From c389d452e9325c86476f211b5aa2070b97eb0bb4 Mon Sep 17 00:00:00 2001
From: Cursor Agent <cursoragent@cursor.com>
Date: Wed, 5 Aug 2026 17:30:12 +0000
Subject: [PATCH 1/2] Add Kimi K2.6 serverless rate limiting test report

Characterize adaptive TPM headers and 429 behavior for
accounts/fireworks/models/kimi-k2p6 across phased, stress,
and burst test scenarios.

Co-authored-by: tarunipaleru <tarunipaleru@users.noreply.github.com>
---
 kimi-k2p6-rate-limit-report.md       |  150 +++
 kimi-k2p6-rate-limit-results.json    | 1344 ++++++++++++++++++++++++++
 kimi-k2p6-rpm-stress.json            |  270 ++++++
 kimi-k2p6-stress-results.json        |   20 +
 scripts/kimi_k2p6_rate_limit_test.py |  256 +++++
 scripts/kimi_k2p6_stress_test.py     |   94 ++
 6 files changed, 2134 insertions(+)
 create mode 100644 kimi-k2p6-rate-limit-report.md
 create mode 100644 kimi-k2p6-rate-limit-results.json
 create mode 100644 kimi-k2p6-rpm-stress.json
 create mode 100644 kimi-k2p6-stress-results.json
 create mode 100644 scripts/kimi_k2p6_rate_limit_test.py
 create mode 100644 scripts/kimi_k2p6_stress_test.py

diff --git a/kimi-k2p6-rate-limit-report.md b/kimi-k2p6-rate-limit-report.md
new file mode 100644
index 0000000..b4524d4
--- /dev/null
+++ b/kimi-k2p6-rate-limit-report.md
@@ -0,0 +1,150 @@
+# Kimi K2.6 Serverless Rate Limiting Test Report
+
+**Date:** August 5, 2026  
+**Provider:** Fireworks AI (serverless)  
+**Model:** `accounts/fireworks/models/kimi-k2p6`  
+**Endpoint:** `POST https://api.fireworks.ai/inference/v1/chat/completions`
+
+---
+
+## Executive Summary
+
+Kimi K2.6 on Fireworks serverless exposes **adaptive per-model token rate limits** via response headers. Under moderate parallel load (≤15 concurrent, ≤512 output tokens), all **224/224 requests succeeded** with no throttling. Under an aggressive burst of **200 tiny requests at 100-way concurrency**, **12/200 (6%) returned HTTP 429**, triggered by **generated-token TPM exhaustion** (`x-ratelimit-remaining-tokens-generated: 0`), not prompt-token or account RPM limits.
+
+The observed adaptive ceiling for generated tokens **dropped from 216,000 TPM to 36,000 TPM** during the burst, demonstrating the adaptive shrink behavior described in Fireworks docs.
+
+---
+
+## Observed Rate Limit Headers
+
+After sustained load, all three TPM dimensions are reported:
+
+| Header | Observed Limit (TPM) | Docs Default Ceiling |
+|--------|---------------------:|---------------------:|
+| `x-ratelimit-limit-tokens-prompt` | 11,250,000 | 21,600,000 |
+| `x-ratelimit-limit-tokens-cache-adjusted-prompt` | 5,400,000 | 5,400,000 |
+| `x-ratelimit-limit-tokens-generated` | 216,000 → **36,000** (after burst) | 216,000 |
+
+Equivalent TPS (limit ÷ 60):
+
+| Dimension | Initial TPS | Post-burst TPS |
+|-----------|------------:|---------------:|
+| Total prompt | 187,500 | 187,500 |
+| Uncached prompt | 90,000 | 90,000 |
+| Generated | 3,600 | **600** |
+
+The account's effective prompt limit (11.25M TPM) is below the documented default ceiling, consistent with adaptive limits that grow with usage history.
+
+---
+
+## Test Phases
+
+### Phase 1 — Characterization (74 requests, moderate load)
+
+| Phase | Requests | Concurrency | Success | 429 | 503 | Avg Latency |
+|-------|----------|-------------|---------|-----|-----|-------------|
+| Baseline single | 1 | 1 | 1/1 | 0 | 0 | 0.45s |
+| Steady low parallel | 10 | 2 | 10/10 | 0 | 0 | 1.25s |
+| Moderate parallel | 20 | 5 | 20/20 | 0 | 0 | 4.13s |
+| Burst parallel | 30 | 15 | 30/30 | 0 | 0 | 1.73s |
+| Heavy output (256 tok) | 8 | 4 | 8/8 | 0 | 0 | 5.44s |
+| Recovery (2s spacing) | 5 | 1 | 5/5 | 0 | 0 | 4.64s |
+
+**Finding:** No throttling at ≤15 concurrency with output up to 256 tokens. Remaining-token counters decrement predictably but stay well above zero.
+
+### Phase 2 — Token Stress (150 requests, high concurrency)
+
+| Scenario | Requests | Concurrency | Success | 429 | 503 |
+|----------|----------|-------------|---------|-----|-----|
+| stress_50x20 | 50 | 20 | 50/50 | 0 | 0 |
+| stress_100x50 | 100 | 50 | 100/100 | 0 | 0 |
+
+**Finding:** Even 50-way concurrent requests with 512 max output tokens did not trigger 429. Generated-token remaining stayed above zero (lowest observed: ~208,898 of 216,000).
+
+### Phase 3 — RPM / Generated-TPM Burst (200 requests, 100-way concurrency)
+
+| Metric | Value |
+|--------|------:|
+| Total requests | 200 |
+| Concurrency | 100 |
+| `max_tokens` | 1 |
+| Wall time | 3.13s |
+| Effective RPM | ~3,840 |
+| Success (200) | 188 (94%) |
+| Throttled (429) | 12 (6%) |
+| Load shed (503) | 0 |
+
+**429 trigger condition:**
+
+```
+x-ratelimit-remaining-tokens-generated: 0
+x-ratelimit-limit-tokens-generated: 36000   (shrunk from 216000)
+x-ratelimit-over-limit: no                  (still reports "no")
+```
+
+**429 error body:**
+
+```json
+{
+  "object": "error",
+  "type": "invalid_request_error",
+  "code": "invalid_request_error",
+  "message": "rate limit exceeded, please try again later"
+}
+```
+
+First 429 appeared at request index 18 (~0.23s into the burst). All 429 responses returned in <260ms (fail-fast, no queueing).
+
+---
+
+## Key Findings
+
+### 1. Generated TPM Is the First Bottleneck for Burst Traffic
+
+When firing many concurrent requests with even minimal output (`max_tokens=1`), the **generated-token bucket exhausts first**. Prompt-token remaining stayed above 11.2M (of 11.25M limit) at the time of 429s.
+
+### 2. Adaptive Limits Shrink Under Burst
+
+The generated-token limit dropped from **216,000 → 36,000 TPM** (6× reduction) during the burst test. This matches Fireworks' documented adaptive behavior: limits grow and shrink based on usage patterns; ramping too quickly causes 429s.
+
+### 3. `x-ratelimit-over-limit: no` on 429 Responses
+
+Even when returning HTTP 429, the `x-ratelimit-over-limit` header remained `"no"`. Clients should treat **HTTP status code 429** and **`remaining-tokens-*: 0`** as the authoritative throttle signals, not the over-limit flag alone.
+
+### 4. No 503 Load Shedding Observed
+
+Across 374 total requests, zero `503 Service Overloaded` responses were observed. Throttling manifested exclusively as 429 rate-limit errors.
+
+### 5. Account RPM Limit Not Hit
+
+Peak effective throughput was ~3,840 RPM, well below the documented 6,000 RPM account-wide ceiling. The 429s in Phase 3 were token-bucket driven, not request-count driven.
+
+---
+
+## Recommendations for Kimi K2.6 Clients
+
+1. **Always set `max_tokens` explicitly** — Kimi K2.6 can produce long reasoning traces; unbounded output accelerates generated-TPM consumption.
+2. **Use exponential backoff on 429** — Fail-fast responses (<300ms) indicate bucket exhaustion; retry after a brief delay.
+3. **Monitor all three header dimensions** — Track `remaining-tokens-prompt`, `remaining-tokens-cache-adjusted-prompt`, and `remaining-tokens-generated`.
+4. **Ramp concurrency gradually** — Jumping to 50–100 concurrent requests causes adaptive limit shrink and 429s even when per-request token usage is tiny.
+5. **Separate rate-limit from load-shed** — 429 = token/RPM bucket; 503 = deployment saturation. Different remediation strategies apply.
+
+---
+
+## Artifacts
+
+| File | Description |
+|------|-------------|
+| `scripts/kimi_k2p6_rate_limit_test.py` | Phased characterization test harness |
+| `scripts/kimi_k2p6_stress_test.py` | High-concurrency token stress test |
+| `kimi-k2p6-rate-limit-results.json` | Raw results from Phase 1 |
+| `kimi-k2p6-stress-results.json` | Raw results from Phase 2 |
+| `kimi-k2p6-rpm-stress.json` | Raw results from Phase 3 burst |
+
+---
+
+## References
+
+- [Fireworks Serverless Rate Limits](https://docs.fireworks.ai/serverless/rate-limits)
+- [Account Quotas & RPM Limits](https://docs.fireworks.ai/guides/quotas_usage/account-quotas)
+- [Kimi K2 Family on Fireworks](https://docs.fireworks.ai/models/kimi-k2)
diff --git a/kimi-k2p6-rate-limit-results.json b/kimi-k2p6-rate-limit-results.json
new file mode 100644
index 0000000..51db3df
--- /dev/null
+++ b/kimi-k2p6-rate-limit-results.json
@@ -0,0 +1,1344 @@
+{
+  "model": "accounts/fireworks/models/kimi-k2p6",
+  "timestamp_utc": "2026-08-05T17:26:41Z",
+  "phases": {
+    "baseline_single": {
+      "total": 1,
+      "status_counts": {
+        "200": 1
+      },
+      "success_rate": 1.0,
+      "latency_s": {
+        "min": 0.4495302739999829,
+        "max": 0.4495302739999829,
+        "mean": 0.4495302739999829,
+        "p50": 0.4495302739999829
+      },
+      "latest_rate_limit_headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249984",
+        "x-ratelimit-over-limit": "no"
+      },
+      "first_429_index": null,
+      "first_503_index": null,
+      "sample_errors": []
+    },
+    "steady_low_parallel": {
+      "total": 10,
+      "status_counts": {
+        "200": 10
+      },
+      "success_rate": 1.0,
+      "latency_s": {
+        "min": 0.8768106880000062,
+        "max": 1.7742476210000007,
+        "mean": 1.246523089100006,
+        "p50": 1.2186630235000138
+      },
+      "latest_rate_limit_headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249981",
+        "x-ratelimit-over-limit": "no"
+      },
+      "first_429_index": null,
+      "first_503_index": null,
+      "sample_errors": []
+    },
+    "moderate_parallel": {
+      "total": 20,
+      "status_counts": {
+        "200": 20
+      },
+      "success_rate": 1.0,
+      "latency_s": {
+        "min": 1.969659390000004,
+        "max": 13.489280185000013,
+        "mean": 4.133620091849998,
+        "p50": 2.4672788430000026
+      },
+      "latest_rate_limit_headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246712",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397325",
+        "x-ratelimit-remaining-tokens-generated": "213145",
+        "x-ratelimit-over-limit": "no"
+      },
+      "first_429_index": null,
+      "first_503_index": null,
+      "sample_errors": []
+    },
+    "burst_parallel": {
+      "total": 30,
+      "status_counts": {
+        "200": 30
+      },
+      "success_rate": 1.0,
+      "latency_s": {
+        "min": 0.9237531930000102,
+        "max": 5.309711937000003,
+        "mean": 1.7330934568333343,
+        "p50": 1.2951424485000018
+      },
+      "latest_rate_limit_headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11246421",
+        "x-ratelimit-over-limit": "no"
+      },
+      "first_429_index": null,
+      "first_503_index": null,
+      "sample_errors": []
+    },
+    "heavy_output": {
+      "total": 8,
+      "status_counts": {
+        "200": 8
+      },
+      "success_rate": 1.0,
+      "latency_s": {
+        "min": 3.6082314209999993,
+        "max": 8.83252007499999,
+        "mean": 5.444448210249995,
+        "p50": 4.831071655499997
+      },
+      "latest_rate_limit_headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246402",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397037",
+        "x-ratelimit-remaining-tokens-generated": "208898",
+        "x-ratelimit-over-limit": "no"
+      },
+      "first_429_index": null,
+      "first_503_index": null,
+      "sample_errors": []
+    },
+    "recovery_wait": {
+      "total": 5,
+      "status_counts": {
+        "200": 5
+      },
+      "success_rate": 1.0,
+      "latency_s": {
+        "min": 0.7583341669999868,
+        "max": 6.704704643000014,
+        "mean": 4.639633552999999,
+        "p50": 4.8563677970000185
+      },
+      "latest_rate_limit_headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11245059",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5396694",
+        "x-ratelimit-remaining-tokens-generated": "209111",
+        "x-ratelimit-over-limit": "no"
+      },
+      "first_429_index": null,
+      "first_503_index": null,
+      "sample_errors": []
+    }
+  },
+  "raw_results": [
+    {
+      "phase": "baseline_single",
+      "index": 0,
+      "status_code": 200,
+      "latency_s": 0.4495302739999829,
+      "prompt_tokens": 16,
+      "completion_tokens": 16,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249984",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "steady_low_parallel",
+      "index": 0,
+      "status_code": 200,
+      "latency_s": 1.7430589570000166,
+      "prompt_tokens": 19,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249981",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "steady_low_parallel",
+      "index": 1,
+      "status_code": 200,
+      "latency_s": 1.7742476210000007,
+      "prompt_tokens": 19,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249962",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "steady_low_parallel",
+      "index": 2,
+      "status_code": 200,
+      "latency_s": 0.9262403500000005,
+      "prompt_tokens": 19,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-remaining-tokens-prompt": "11249891",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399910",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "steady_low_parallel",
+      "index": 3,
+      "status_code": 200,
+      "latency_s": 0.8847273050000126,
+      "prompt_tokens": 19,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249981",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "steady_low_parallel",
+      "index": 4,
+      "status_code": 200,
+      "latency_s": 0.8768106880000062,
+      "prompt_tokens": 19,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-remaining-tokens-prompt": "11249872",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399891",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "steady_low_parallel",
+      "index": 5,
+      "status_code": 200,
+      "latency_s": 0.9169327669999916,
+      "prompt_tokens": 19,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-remaining-tokens-prompt": "11249853",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399891",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "steady_low_parallel",
+      "index": 6,
+      "status_code": 200,
+      "latency_s": 1.2291536480000218,
+      "prompt_tokens": 19,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249929",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399948",
+        "x-ratelimit-remaining-tokens-generated": "215706",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "steady_low_parallel",
+      "index": 7,
+      "status_code": 200,
+      "latency_s": 1.2081723990000057,
+      "prompt_tokens": 19,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249796",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399948",
+        "x-ratelimit-remaining-tokens-generated": "215706",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "steady_low_parallel",
+      "index": 8,
+      "status_code": 200,
+      "latency_s": 1.3116970739999942,
+      "prompt_tokens": 19,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249777",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399869",
+        "x-ratelimit-remaining-tokens-generated": "215706",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "steady_low_parallel",
+      "index": 9,
+      "status_code": 200,
+      "latency_s": 1.5941900820000114,
+      "prompt_tokens": 19,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249981",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 0,
+      "status_code": 200,
+      "latency_s": 2.112493157999978,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249756",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399860",
+        "x-ratelimit-remaining-tokens-generated": "215130",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 1,
+      "status_code": 200,
+      "latency_s": 2.0788152279999963,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249735",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399860",
+        "x-ratelimit-remaining-tokens-generated": "215130",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 2,
+      "status_code": 200,
+      "latency_s": 13.457547730999977,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249958",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 3,
+      "status_code": 200,
+      "latency_s": 13.489280185000013,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249979",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 4,
+      "status_code": 200,
+      "latency_s": 1.969659390000004,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249714",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399860",
+        "x-ratelimit-remaining-tokens-generated": "215130",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 5,
+      "status_code": 200,
+      "latency_s": 11.484700843000013,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249632",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 6,
+      "status_code": 200,
+      "latency_s": 2.724387605000004,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249958",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 7,
+      "status_code": 200,
+      "latency_s": 2.5724797000000024,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249979",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 8,
+      "status_code": 200,
+      "latency_s": 7.027807866999979,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249569",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399745",
+        "x-ratelimit-remaining-tokens-generated": "214482",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 9,
+      "status_code": 200,
+      "latency_s": 2.3967286900000033,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249979",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 10,
+      "status_code": 200,
+      "latency_s": 2.073006573999976,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249653",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399787",
+        "x-ratelimit-remaining-tokens-generated": "214682",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 11,
+      "status_code": 200,
+      "latency_s": 2.157902584999988,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249156",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399370",
+        "x-ratelimit-remaining-tokens-generated": "214194",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 12,
+      "status_code": 200,
+      "latency_s": 2.537828996000002,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249166",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399387",
+        "x-ratelimit-remaining-tokens-generated": "214281",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 13,
+      "status_code": 200,
+      "latency_s": 2.110844061999984,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249171",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399387",
+        "x-ratelimit-remaining-tokens-generated": "214006",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 14,
+      "status_code": 200,
+      "latency_s": 2.8061243049999973,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249154",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399410",
+        "x-ratelimit-remaining-tokens-generated": "214006",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 15,
+      "status_code": 200,
+      "latency_s": 2.8343422460000056,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249173",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399382",
+        "x-ratelimit-remaining-tokens-generated": "214114",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 16,
+      "status_code": 200,
+      "latency_s": 2.540057544000007,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249152",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399382",
+        "x-ratelimit-remaining-tokens-generated": "214114",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 17,
+      "status_code": 200,
+      "latency_s": 2.0378839930000083,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249593",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399357",
+        "x-ratelimit-remaining-tokens-generated": "214251",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 18,
+      "status_code": 200,
+      "latency_s": 2.1159235450000153,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246730",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399357",
+        "x-ratelimit-remaining-tokens-generated": "214251",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "moderate_parallel",
+      "index": 19,
+      "status_code": 200,
+      "latency_s": 2.1445875900000146,
+      "prompt_tokens": 21,
+      "completion_tokens": 128,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246712",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397325",
+        "x-ratelimit-remaining-tokens-generated": "213145",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 0,
+      "status_code": 200,
+      "latency_s": 0.9237531930000102,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249970",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 1,
+      "status_code": 200,
+      "latency_s": 1.3765895720000003,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246784",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397349",
+        "x-ratelimit-remaining-tokens-generated": "212498",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 2,
+      "status_code": 200,
+      "latency_s": 1.5193182350000143,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249163",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399360",
+        "x-ratelimit-remaining-tokens-generated": "212907",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 3,
+      "status_code": 200,
+      "latency_s": 0.9528534089999994,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249955",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 4,
+      "status_code": 200,
+      "latency_s": 0.9517804159999912,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249985",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 5,
+      "status_code": 200,
+      "latency_s": 1.4148320680000097,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246754",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397349",
+        "x-ratelimit-remaining-tokens-generated": "212498",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 6,
+      "status_code": 200,
+      "latency_s": 1.1147605699999872,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246785",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397326",
+        "x-ratelimit-remaining-tokens-generated": "212341",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 7,
+      "status_code": 200,
+      "latency_s": 1.4158787460000042,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246769",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397349",
+        "x-ratelimit-remaining-tokens-generated": "212498",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 8,
+      "status_code": 200,
+      "latency_s": 1.1494519590000039,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246770",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397326",
+        "x-ratelimit-remaining-tokens-generated": "212341",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 9,
+      "status_code": 200,
+      "latency_s": 1.2462186180000003,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246800",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397326",
+        "x-ratelimit-remaining-tokens-generated": "212341",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 10,
+      "status_code": 200,
+      "latency_s": 1.1293881349999992,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249970",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 11,
+      "status_code": 200,
+      "latency_s": 1.6545148459999837,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249148",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399360",
+        "x-ratelimit-remaining-tokens-generated": "212907",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 12,
+      "status_code": 200,
+      "latency_s": 1.0952261200000066,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249955",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 13,
+      "status_code": 200,
+      "latency_s": 1.0968196240000054,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249985",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 14,
+      "status_code": 200,
+      "latency_s": 1.6555409970000028,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249133",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399360",
+        "x-ratelimit-remaining-tokens-generated": "212907",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 15,
+      "status_code": 200,
+      "latency_s": 2.139889748999991,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249985",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 16,
+      "status_code": 200,
+      "latency_s": 2.092999118999984,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249955",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 17,
+      "status_code": 200,
+      "latency_s": 2.086777927000014,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249970",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 18,
+      "status_code": 200,
+      "latency_s": 1.144329980000009,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249970",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 19,
+      "status_code": 200,
+      "latency_s": 1.0560014700000124,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249985",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 20,
+      "status_code": 200,
+      "latency_s": 1.2520901720000097,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249955",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 21,
+      "status_code": 200,
+      "latency_s": 1.525174015999994,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11246656",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 22,
+      "status_code": 200,
+      "latency_s": 1.6753569660000096,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249985",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 23,
+      "status_code": 200,
+      "latency_s": 1.338194724999994,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11246641",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 24,
+      "status_code": 200,
+      "latency_s": 5.309711937000003,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249205",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397319",
+        "x-ratelimit-remaining-tokens-generated": "212828",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 25,
+      "status_code": 200,
+      "latency_s": 5.130342642999977,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246465",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397319",
+        "x-ratelimit-remaining-tokens-generated": "212828",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 26,
+      "status_code": 200,
+      "latency_s": 5.186301154000006,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249190",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397319",
+        "x-ratelimit-remaining-tokens-generated": "212828",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 27,
+      "status_code": 200,
+      "latency_s": 1.235240860999994,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249985",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 28,
+      "status_code": 200,
+      "latency_s": 1.0229154160000178,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11246436",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "burst_parallel",
+      "index": 29,
+      "status_code": 200,
+      "latency_s": 1.1005510619999939,
+      "prompt_tokens": 15,
+      "completion_tokens": 64,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11246421",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "heavy_output",
+      "index": 0,
+      "status_code": 200,
+      "latency_s": 5.2359516649999875,
+      "prompt_tokens": 21,
+      "completion_tokens": 256,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249937",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "heavy_output",
+      "index": 1,
+      "status_code": 200,
+      "latency_s": 5.66713300699999,
+      "prompt_tokens": 21,
+      "completion_tokens": 256,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249958",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "heavy_output",
+      "index": 2,
+      "status_code": 200,
+      "latency_s": 7.799037986999991,
+      "prompt_tokens": 21,
+      "completion_tokens": 256,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246505",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397059",
+        "x-ratelimit-remaining-tokens-generated": "209741",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "heavy_output",
+      "index": 3,
+      "status_code": 200,
+      "latency_s": 4.426191646000007,
+      "prompt_tokens": 21,
+      "completion_tokens": 256,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-remaining-tokens-prompt": "11249979",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "heavy_output",
+      "index": 4,
+      "status_code": 200,
+      "latency_s": 4.187831545999984,
+      "prompt_tokens": 21,
+      "completion_tokens": 256,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246382",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397013",
+        "x-ratelimit-remaining-tokens-generated": "209741",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "heavy_output",
+      "index": 5,
+      "status_code": 200,
+      "latency_s": 3.6082314209999993,
+      "prompt_tokens": 21,
+      "completion_tokens": 256,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246419",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397013",
+        "x-ratelimit-remaining-tokens-generated": "209741",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "heavy_output",
+      "index": 6,
+      "status_code": 200,
+      "latency_s": 8.83252007499999,
+      "prompt_tokens": 21,
+      "completion_tokens": 256,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11249979",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5399335",
+        "x-ratelimit-remaining-tokens-generated": "213747",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "heavy_output",
+      "index": 7,
+      "status_code": 200,
+      "latency_s": 3.7986883350000085,
+      "prompt_tokens": 21,
+      "completion_tokens": 256,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246402",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397037",
+        "x-ratelimit-remaining-tokens-generated": "208898",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "recovery_wait",
+      "index": 0,
+      "status_code": 200,
+      "latency_s": 4.8563677970000185,
+      "prompt_tokens": 11,
+      "completion_tokens": 8,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11246029",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5397041",
+        "x-ratelimit-remaining-tokens-generated": "207563",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "recovery_wait",
+      "index": 1,
+      "status_code": 200,
+      "latency_s": 4.493718208999979,
+      "prompt_tokens": 11,
+      "completion_tokens": 8,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11243768",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5395120",
+        "x-ratelimit-remaining-tokens-generated": "206663",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "recovery_wait",
+      "index": 2,
+      "status_code": 200,
+      "latency_s": 6.704704643000014,
+      "prompt_tokens": 11,
+      "completion_tokens": 8,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11243851",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5395186",
+        "x-ratelimit-remaining-tokens-generated": "206226",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "recovery_wait",
+      "index": 3,
+      "status_code": 200,
+      "latency_s": 6.385042948999995,
+      "prompt_tokens": 11,
+      "completion_tokens": 8,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11243952",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5395642",
+        "x-ratelimit-remaining-tokens-generated": "206921",
+        "x-ratelimit-over-limit": "no"
+      }
+    },
+    {
+      "phase": "recovery_wait",
+      "index": 4,
+      "status_code": 200,
+      "latency_s": 0.7583341669999868,
+      "prompt_tokens": 11,
+      "completion_tokens": 8,
+      "error": null,
+      "headers": {
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "216000",
+        "x-ratelimit-remaining-tokens-prompt": "11245059",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5396694",
+        "x-ratelimit-remaining-tokens-generated": "209111",
+        "x-ratelimit-over-limit": "no"
+      }
+    }
+  ]
+}
\ No newline at end of file
diff --git a/kimi-k2p6-rpm-stress.json b/kimi-k2p6-rpm-stress.json
new file mode 100644
index 0000000..221a6a5
--- /dev/null
+++ b/kimi-k2p6-rpm-stress.json
@@ -0,0 +1,270 @@
+{
+  "elapsed_s": 3.13,
+  "rpm_equivalent": 3839.8,
+  "status_counts": {
+    "200": 188,
+    "429": 12
+  },
+  "first_429": {
+    "i": 18,
+    "status": 429,
+    "lat": 0.233,
+    "err": {
+      "object": "error",
+      "type": "invalid_request_error",
+      "code": "invalid_request_error",
+      "message": "rate limit exceeded, please try again later"
+    },
+    "h": {
+      "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+      "x-ratelimit-limit-tokens-generated": "36000",
+      "x-ratelimit-limit-tokens-prompt": "11250000",
+      "x-ratelimit-over-limit": "no",
+      "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5398911",
+      "x-ratelimit-remaining-tokens-generated": "0",
+      "x-ratelimit-remaining-tokens-prompt": "11247531"
+    }
+  },
+  "sample_errors": [
+    {
+      "i": 18,
+      "status": 429,
+      "lat": 0.233,
+      "err": {
+        "object": "error",
+        "type": "invalid_request_error",
+        "code": "invalid_request_error",
+        "message": "rate limit exceeded, please try again later"
+      },
+      "h": {
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "36000",
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-over-limit": "no",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5398911",
+        "x-ratelimit-remaining-tokens-generated": "0",
+        "x-ratelimit-remaining-tokens-prompt": "11247531"
+      }
+    },
+    {
+      "i": 53,
+      "status": 429,
+      "lat": 0.225,
+      "err": {
+        "object": "error",
+        "type": "invalid_request_error",
+        "code": "invalid_request_error",
+        "message": "rate limit exceeded, please try again later"
+      },
+      "h": {
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "36000",
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-over-limit": "no",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5398911",
+        "x-ratelimit-remaining-tokens-generated": "0",
+        "x-ratelimit-remaining-tokens-prompt": "11247531"
+      }
+    },
+    {
+      "i": 90,
+      "status": 429,
+      "lat": 0.218,
+      "err": {
+        "object": "error",
+        "type": "invalid_request_error",
+        "code": "invalid_request_error",
+        "message": "rate limit exceeded, please try again later"
+      },
+      "h": {
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "36000",
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-over-limit": "no",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5398911",
+        "x-ratelimit-remaining-tokens-generated": "0",
+        "x-ratelimit-remaining-tokens-prompt": "11247531"
+      }
+    },
+    {
+      "i": 103,
+      "status": 429,
+      "lat": 0.15,
+      "err": {
+        "object": "error",
+        "type": "invalid_request_error",
+        "code": "invalid_request_error",
+        "message": "rate limit exceeded, please try again later"
+      },
+      "h": {
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "36000",
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-over-limit": "no",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5398911",
+        "x-ratelimit-remaining-tokens-generated": "0",
+        "x-ratelimit-remaining-tokens-prompt": "11250000"
+      }
+    },
+    {
+      "i": 104,
+      "status": 429,
+      "lat": 0.101,
+      "err": {
+        "object": "error",
+        "type": "invalid_request_error",
+        "code": "invalid_request_error",
+        "message": "rate limit exceeded, please try again later"
+      },
+      "h": {
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "36000",
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-over-limit": "no",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5398911",
+        "x-ratelimit-remaining-tokens-generated": "0",
+        "x-ratelimit-remaining-tokens-prompt": "11247531"
+      }
+    },
+    {
+      "i": 110,
+      "status": 429,
+      "lat": 0.15,
+      "err": {
+        "object": "error",
+        "type": "invalid_request_error",
+        "code": "invalid_request_error",
+        "message": "rate limit exceeded, please try again later"
+      },
+      "h": {
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "36000",
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-over-limit": "no",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5398911",
+        "x-ratelimit-remaining-tokens-generated": "0",
+        "x-ratelimit-remaining-tokens-prompt": "11250000"
+      }
+    },
+    {
+      "i": 111,
+      "status": 429,
+      "lat": 0.149,
+      "err": {
+        "object": "error",
+        "type": "invalid_request_error",
+        "code": "invalid_request_error",
+        "message": "rate limit exceeded, please try again later"
+      },
+      "h": {
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "36000",
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-over-limit": "no",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5398911",
+        "x-ratelimit-remaining-tokens-generated": "0",
+        "x-ratelimit-remaining-tokens-prompt": "11250000"
+      }
+    },
+    {
+      "i": 112,
+      "status": 429,
+      "lat": 0.149,
+      "err": {
+        "object": "error",
+        "type": "invalid_request_error",
+        "code": "invalid_request_error",
+        "message": "rate limit exceeded, please try again later"
+      },
+      "h": {
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "36000",
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-over-limit": "no",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5398911",
+        "x-ratelimit-remaining-tokens-generated": "0",
+        "x-ratelimit-remaining-tokens-prompt": "11250000"
+      }
+    },
+    {
+      "i": 134,
+      "status": 429,
+      "lat": 0.144,
+      "err": {
+        "object": "error",
+        "type": "invalid_request_error",
+        "code": "invalid_request_error",
+        "message": "rate limit exceeded, please try again later"
+      },
+      "h": {
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "36000",
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-over-limit": "no",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5398911",
+        "x-ratelimit-remaining-tokens-generated": "0",
+        "x-ratelimit-remaining-tokens-prompt": "11250000"
+      }
+    },
+    {
+      "i": 144,
+      "status": 429,
+      "lat": 0.255,
+      "err": {
+        "object": "error",
+        "type": "invalid_request_error",
+        "code": "invalid_request_error",
+        "message": "rate limit exceeded, please try again later"
+      },
+      "h": {
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "36000",
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-over-limit": "no",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5398911",
+        "x-ratelimit-remaining-tokens-generated": "0",
+        "x-ratelimit-remaining-tokens-prompt": "11250000"
+      }
+    },
+    {
+      "i": 153,
+      "status": 429,
+      "lat": 0.152,
+      "err": {
+        "object": "error",
+        "type": "invalid_request_error",
+        "code": "invalid_request_error",
+        "message": "rate limit exceeded, please try again later"
+      },
+      "h": {
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "36000",
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-over-limit": "no",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5398911",
+        "x-ratelimit-remaining-tokens-generated": "0",
+        "x-ratelimit-remaining-tokens-prompt": "11250000"
+      }
+    },
+    {
+      "i": 178,
+      "status": 429,
+      "lat": 0.104,
+      "err": {
+        "object": "error",
+        "type": "invalid_request_error",
+        "code": "invalid_request_error",
+        "message": "rate limit exceeded, please try again later"
+      },
+      "h": {
+        "x-ratelimit-limit-tokens-cache-adjusted-prompt": "5400000",
+        "x-ratelimit-limit-tokens-generated": "36000",
+        "x-ratelimit-limit-tokens-prompt": "11250000",
+        "x-ratelimit-over-limit": "no",
+        "x-ratelimit-remaining-tokens-cache-adjusted-prompt": "5398911",
+        "x-ratelimit-remaining-tokens-generated": "0",
+        "x-ratelimit-remaining-tokens-prompt": "11250000"
+      }
+    }
+  ]
+}
\ No newline at end of file
diff --git a/kimi-k2p6-stress-results.json b/kimi-k2p6-stress-results.json
new file mode 100644
index 0000000..18f125a
--- /dev/null
+++ b/kimi-k2p6-stress-results.json
@@ -0,0 +1,20 @@
+{
+  "scenarios": {
+    "stress_50x20": {
+      "status_counts": {
+        "200": 50
+      },
+      "first_429": null,
+      "sample_429s": [],
+      "sample_503s": []
+    },
+    "stress_100x50": {
+      "status_counts": {
+        "200": 100
+      },
+      "first_429": null,
+      "sample_429s": [],
+      "sample_503s": []
+    }
+  }
+}
\ No newline at end of file
diff --git a/scripts/kimi_k2p6_rate_limit_test.py b/scripts/kimi_k2p6_rate_limit_test.py
new file mode 100644
index 0000000..58ada01
--- /dev/null
+++ b/scripts/kimi_k2p6_rate_limit_test.py
@@ -0,0 +1,256 @@
+#!/usr/bin/env python3
+"""Rate limit characterization for Kimi K2.6 on Fireworks serverless."""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import json
+import os
+import statistics
+import time
+from dataclasses import asdict, dataclass, field
+from typing import Any
+
+import httpx
+
+MODEL = "accounts/fireworks/models/kimi-k2p6"
+API_URL = "https://api.fireworks.ai/inference/v1/chat/completions"
+RATE_LIMIT_HEADERS = [
+    "x-ratelimit-limit-tokens-prompt",
+    "x-ratelimit-limit-tokens-cache-adjusted-prompt",
+    "x-ratelimit-limit-tokens-generated",
+    "x-ratelimit-remaining-tokens-prompt",
+    "x-ratelimit-remaining-tokens-cache-adjusted-prompt",
+    "x-ratelimit-remaining-tokens-generated",
+    "x-ratelimit-over-limit",
+    "retry-after",
+]
+
+
+@dataclass
+class RequestResult:
+    phase: str
+    index: int
+    status_code: int
+    latency_s: float
+    prompt_tokens: int | None = None
+    completion_tokens: int | None = None
+    error: str | None = None
+    headers: dict[str, str] = field(default_factory=dict)
+
+
+def extract_rate_headers(headers: httpx.Headers) -> dict[str, str]:
+    out: dict[str, str] = {}
+    for key in RATE_LIMIT_HEADERS:
+        if key in headers:
+            out[key] = headers[key]
+    return out
+
+
+async def send_request(
+    client: httpx.AsyncClient,
+    *,
+    phase: str,
+    index: int,
+    prompt: str,
+    max_tokens: int,
+) -> RequestResult:
+    payload = {
+        "model": MODEL,
+        "messages": [{"role": "user", "content": prompt}],
+        "max_tokens": max_tokens,
+    }
+    started = time.perf_counter()
+    try:
+        response = await client.post(API_URL, json=payload)
+        latency = time.perf_counter() - started
+        body: dict[str, Any] = {}
+        try:
+            body = response.json()
+        except json.JSONDecodeError:
+            body = {}
+
+        usage = body.get("usage") or {}
+        error = None
+        if response.status_code >= 400:
+            error = body.get("error", body)
+
+        return RequestResult(
+            phase=phase,
+            index=index,
+            status_code=response.status_code,
+            latency_s=latency,
+            prompt_tokens=usage.get("prompt_tokens"),
+            completion_tokens=usage.get("completion_tokens"),
+            error=str(error) if error else None,
+            headers=extract_rate_headers(response.headers),
+        )
+    except Exception as exc:  # noqa: BLE001
+        return RequestResult(
+            phase=phase,
+            index=index,
+            status_code=0,
+            latency_s=time.perf_counter() - started,
+            error=str(exc),
+        )
+
+
+async def run_phase(
+    client: httpx.AsyncClient,
+    *,
+    phase: str,
+    count: int,
+    concurrency: int,
+    prompt: str,
+    max_tokens: int,
+    delay_s: float = 0.0,
+) -> list[RequestResult]:
+    semaphore = asyncio.Semaphore(concurrency)
+    results: list[RequestResult] = []
+
+    async def one(index: int) -> None:
+        async with semaphore:
+            if delay_s > 0:
+                await asyncio.sleep(index * delay_s)
+            results.append(
+                await send_request(
+                    client,
+                    phase=phase,
+                    index=index,
+                    prompt=prompt,
+                    max_tokens=max_tokens,
+                )
+            )
+
+    await asyncio.gather(*(one(i) for i in range(count)))
+    return sorted(results, key=lambda item: item.index)
+
+
+def summarize(results: list[RequestResult]) -> dict[str, Any]:
+    by_status: dict[int, int] = {}
+    latencies = [r.latency_s for r in results if r.status_code == 200]
+    header_samples: list[dict[str, str]] = []
+    for result in results:
+        by_status[result.status_code] = by_status.get(result.status_code, 0) + 1
+        if result.headers:
+            header_samples.append(result.headers)
+
+    latest_headers = header_samples[-1] if header_samples else {}
+    return {
+        "total": len(results),
+        "status_counts": by_status,
+        "success_rate": by_status.get(200, 0) / len(results) if results else 0.0,
+        "latency_s": {
+            "min": min(latencies) if latencies else None,
+            "max": max(latencies) if latencies else None,
+            "mean": statistics.mean(latencies) if latencies else None,
+            "p50": statistics.median(latencies) if latencies else None,
+        },
+        "latest_rate_limit_headers": latest_headers,
+        "first_429_index": next(
+            (r.index for r in results if r.status_code == 429),
+            None,
+        ),
+        "first_503_index": next(
+            (r.index for r in results if r.status_code == 503),
+            None,
+        ),
+        "sample_errors": [
+            {"index": r.index, "status": r.status_code, "error": r.error}
+            for r in results
+            if r.status_code != 200
+        ][:5],
+    }
+
+
+async def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--output", required=True)
+    args = parser.parse_args()
+
+    api_key = os.environ.get("FIREWORKS_API_KEY")
+    if not api_key:
+        raise SystemExit("FIREWORKS_API_KEY is required")
+
+    timeout = httpx.Timeout(120.0, connect=30.0)
+    headers = {"Authorization": f"Bearer {api_key}"}
+
+    report: dict[str, Any] = {
+        "model": MODEL,
+        "timestamp_utc": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()),
+        "phases": {},
+        "raw_results": [],
+    }
+
+    async with httpx.AsyncClient(headers=headers, timeout=timeout) as client:
+        phases = [
+            {
+                "name": "baseline_single",
+                "count": 1,
+                "concurrency": 1,
+                "prompt": "Reply with exactly one word: hello.",
+                "max_tokens": 16,
+            },
+            {
+                "name": "steady_low_parallel",
+                "count": 10,
+                "concurrency": 2,
+                "prompt": "In one short sentence, explain what rate limiting means.",
+                "max_tokens": 64,
+            },
+            {
+                "name": "moderate_parallel",
+                "count": 20,
+                "concurrency": 5,
+                "prompt": "Write a 2-sentence summary of adaptive token rate limits.",
+                "max_tokens": 128,
+            },
+            {
+                "name": "burst_parallel",
+                "count": 30,
+                "concurrency": 15,
+                "prompt": "List five synonyms for 'fast'.",
+                "max_tokens": 64,
+            },
+            {
+                "name": "heavy_output",
+                "count": 8,
+                "concurrency": 4,
+                "prompt": "Write a detailed 150-word explanation of HTTP 429 errors.",
+                "max_tokens": 256,
+            },
+            {
+                "name": "recovery_wait",
+                "count": 5,
+                "concurrency": 1,
+                "prompt": "Say OK.",
+                "max_tokens": 8,
+                "delay_s": 2.0,
+            },
+        ]
+
+        for spec in phases:
+            phase_name = spec["name"]
+            print(f"Running phase: {phase_name}", flush=True)
+            results = await run_phase(
+                client,
+                phase=phase_name,
+                count=spec["count"],
+                concurrency=spec["concurrency"],
+                prompt=spec["prompt"],
+                max_tokens=spec["max_tokens"],
+                delay_s=spec.get("delay_s", 0.0),
+            )
+            report["phases"][phase_name] = summarize(results)
+            report["raw_results"].extend(asdict(r) for r in results)
+            await asyncio.sleep(3)
+
+    with open(args.output, "w", encoding="utf-8") as handle:
+        json.dump(report, handle, indent=2)
+
+    print(f"Wrote {args.output}")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/scripts/kimi_k2p6_stress_test.py b/scripts/kimi_k2p6_stress_test.py
new file mode 100644
index 0000000..9ce213a
--- /dev/null
+++ b/scripts/kimi_k2p6_stress_test.py
@@ -0,0 +1,94 @@
+#!/usr/bin/env python3
+"""Aggressive Kimi K2.6 rate-limit stress test."""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import os
+import time
+
+import httpx
+
+MODEL = "accounts/fireworks/models/kimi-k2p6"
+URL = "https://api.fireworks.ai/inference/v1/chat/completions"
+
+
+async def one(client: httpx.AsyncClient, i: int) -> dict:
+    payload = {
+        "model": MODEL,
+        "messages": [
+            {
+                "role": "user",
+                "content": (
+                    "Write a detailed 300-word essay about distributed systems, "
+                    "covering consistency, availability, and partition tolerance."
+                ),
+            }
+        ],
+        "max_tokens": 512,
+    }
+    started = time.perf_counter()
+    try:
+        resp = await client.post(URL, json=payload)
+        body = resp.json() if resp.headers.get("content-type", "").startswith("application/json") else {}
+        return {
+            "index": i,
+            "status": resp.status_code,
+            "latency_s": round(time.perf_counter() - started, 3),
+            "error": body.get("error"),
+            "headers": {
+                k: v
+                for k, v in resp.headers.items()
+                if k.startswith("x-ratelimit") or k == "retry-after"
+            },
+        }
+    except Exception as exc:  # noqa: BLE001
+        return {
+            "index": i,
+            "status": 0,
+            "latency_s": round(time.perf_counter() - started, 3),
+            "error": str(exc),
+            "headers": {},
+        }
+
+
+async def main() -> None:
+    api_key = os.environ["FIREWORKS_API_KEY"]
+    timeout = httpx.Timeout(180.0, connect=30.0)
+    headers = {"Authorization": f"Bearer {api_key}"}
+
+    scenarios = [
+        {"name": "stress_50x20", "count": 50, "concurrency": 20},
+        {"name": "stress_100x50", "count": 100, "concurrency": 50},
+    ]
+
+    report: dict = {"scenarios": {}}
+    async with httpx.AsyncClient(headers=headers, timeout=timeout) as client:
+        for spec in scenarios:
+            print(f"Running {spec['name']}", flush=True)
+            sem = asyncio.Semaphore(spec["concurrency"])
+
+            async def run(i: int) -> dict:
+                async with sem:
+                    return await one(client, i)
+
+            results = await asyncio.gather(*(run(i) for i in range(spec["count"])))
+            status_counts: dict[int, int] = {}
+            for r in results:
+                status_counts[r["status"]] = status_counts.get(r["status"], 0) + 1
+            report["scenarios"][spec["name"]] = {
+                "status_counts": status_counts,
+                "first_429": next((r for r in results if r["status"] == 429), None),
+                "sample_429s": [r for r in results if r["status"] == 429][:3],
+                "sample_503s": [r for r in results if r["status"] == 503][:3],
+            }
+            await asyncio.sleep(5)
+
+    with open("/workspace/kimi-k2p6-stress-results.json", "w", encoding="utf-8") as f:
+        json.dump(report, f, indent=2)
+    print(json.dumps(report, indent=2))
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

From 02b86a50e1ac2a03c8d74151dab1065963f359f8 Mon Sep 17 00:00:00 2001
From: Cursor Agent <cursoragent@cursor.com>
Date: Wed, 5 Aug 2026 17:46:43 +0000
Subject: [PATCH 2/2] Add Clay (ryan-chang-33dcc0) rate-limit investigation

Prometheus shows 62 kimi-k2p6 HTTP 429s with rejection_reason=tokens_prompt
during a sharp ramp at 2026-08-05 01:21-01:23 UTC, despite smoothed TPM
utilization staying well below adaptive ceilings.

Co-authored-by: tarunipaleru <tarunipaleru@users.noreply.github.com>
---
 ryan-chang-33dcc0-rate-limit-investigation.md | 144 ++++++++++++++++++
 1 file changed, 144 insertions(+)
 create mode 100644 ryan-chang-33dcc0-rate-limit-investigation.md

diff --git a/ryan-chang-33dcc0-rate-limit-investigation.md b/ryan-chang-33dcc0-rate-limit-investigation.md
new file mode 100644
index 0000000..bf31a50
--- /dev/null
+++ b/ryan-chang-33dcc0-rate-limit-investigation.md
@@ -0,0 +1,144 @@
+# Investigation: Rate Limiting for `ryan-chang-33dcc0` (Clay)
+
+**Account:** `accounts/ryan-chang-33dcc0` (`Clay [Production]`, ryan.chang@clay.com)  
+**Window:** last 24h from 2026-08-05 ~17:40 UTC (covering 2026-08-04 17:40 → 2026-08-05 17:40)  
+**Primary model involved:** `accounts/fireworks/models/kimi-k2p6` / deployment `accounts/fireworks/deployments/kimi-k2p6`  
+**Source:** GCP Managed Prometheus (`fw-ai-cp-prod`) + billing usage API
+
+---
+
+## Verdict
+
+Clay was rate-limited on **Kimi K2.6**, but **not because sustained TPM sat at their adaptive ceilings**.
+
+Prometheus records **62 HTTP 429s** with:
+
+| Field | Value |
+|-------|-------|
+| `rejection_reason` | `tokens_prompt` |
+| `http_code_allowlisted` | `429` |
+| Deployment | `kimi-k2p6` |
+| When | **2026-08-05 01:21–01:23 UTC** (8 + 39 + 15) |
+
+That lines up with a sharp traffic ramp on kimi-k2p6 (≈25 → **349 req/min** within ~6 minutes). Adaptive limits were still raising during the burst. This matches Fireworks’ documented behavior: ramping too quickly causes 429s even when average utilization looks “fine.”
+
+---
+
+## Why dashboards can look under-limit
+
+Against **global RLM** limits for kimi-k2p6 (5‑minute averages), peak utilization in the last 24h was:
+
+| Dimension | Peak rate | Adaptive limit | Peak util |
+|-----------|----------:|---------------:|----------:|
+| `tokens_generated` | ~7,040 | ~17,334 | **~41%** |
+| `tokens_prompt` | ~68,264 | ~533,334 | **~13%** |
+| `tokens_cache_adjusted_prompt` | ~27,860 | 200k→250k | **~11–14%** |
+
+So if you only compare smoothed rate mirrors to current limits, Clay does **not** look like they hit TPM caps.
+
+Two important caveats:
+
+1. **`tokens_prompt` rate_mirror was missing/0 during the exact 429 window**, so that dimension’s utilization is not observable at the moment of rejection.
+2. Limits were **still adapting upward** during the incident (see below), so a short admission burst can exceed the *then-current* bucket even if later averages look low.
+
+---
+
+## Incident timeline (kimi-k2p6, 2026-08-05 UTC)
+
+Successful request volume (per-minute deltas from `requests_total_per_account_total`):
+
+| Time (UTC) | Successful req/min | `tokens_prompt` 429s | Other errors |
+|------------|-------------------:|---------------------:|-------------:|
+| 01:16 | 25 | 0 | 0 |
+| 01:17 | 118 | 0 | 0 |
+| 01:18 | 206 | 0 | 0 |
+| 01:19 | 307 | 0 | 0 |
+| 01:20 | 345 | 0 | 0 |
+| **01:21** | **349** | **8** | 0 |
+| **01:22** | **277** | **39** | 2 |
+| **01:23** | **194** | **15** | 5 |
+| 01:24 | 154 | 0 | 5 |
+| 01:25 | 110 | 0 | 3 |
+
+Adaptive limits on the same deployment during that hour:
+
+| Metric | Before burst | After raise |
+|--------|-------------:|------------:|
+| `global_rlm_tokens_prompt_limit_per_account` | 426,667 | **533,334** |
+| `global_rlm_tokens_cache_adjusted_prompt_limit_per_account` | 200,000 | **250,000** |
+| `global_rlm_tokens_generated_limit_per_account` | 13,867 | **17,334** |
+
+Additional signal: **`global_rlm_tokens_cache_adjusted_prompt_limit_at_upper_bound_per_account = 1`** for kimi-k2p6 throughout — uncached/cache-adjusted prompt limit is pinned at its ceiling and cannot auto-raise further. No 429s were labeled `tokens_cache_adjusted_prompt` in this window (count ≈ 0), but the ceiling being maxed is relevant for headroom.
+
+No meaningful volume of:
+
+- `concurrency_limit` (583)
+- `pid_shed_load` (583)
+
+in this window.
+
+---
+
+## Usage context (billing API, serverless)
+
+Rough 24h serverless token volume for this account:
+
+| Model | Prompt tokens | Cached prompt | Completion |
+|-------|--------------:|--------------:|-----------:|
+| **kimi-k2p6** | **~335M** (across day buckets) | **~266M** | **~11.6M** |
+| deepseek-v4-flash | ~46M | ~0 | ~9.0M |
+| glm-5p2 | ~15.4M | ~9.6M | ~1.1M |
+| kimi-k3 | ~0.5M | ~0.2M | ~19k |
+
+Kimi K2.6 dominates. Cache hit rate on kimi-k2p6 is high (~79% of prompt tokens cached in the latest day bucket), which is why cache-adjusted headroom matters.
+
+Account has **no on-demand deployments** (`deployments` list empty) — this is serverless-only for inference.
+
+---
+
+## What did *not* cause it
+
+- **Account-wide 6,000 RPM cap:** peak observed request throughput on kimi-k2p6 was ~5–6 RPS (~300–350 RPM), far below 6,000.
+- **Sustained generated-TPM exhaustion:** peak ~41% of adaptive generated limit.
+- **Sustained total-prompt TPM exhaustion (as averaged):** peak ~13% of adaptive prompt limit in rate_mirror (with the caveat that mirror data was absent during the 429 minute).
+- **Load shed / concurrency 583s:** essentially none in the incident window.
+
+Note on `requests_limit_per_account`: the series is stuck at **1.0** for kimi-k2p6 while successful traffic clearly runs at several RPS. Treat that series as a **stale/floor signal**, not the effective request limiter. The authoritative rejection label is `tokens_prompt`.
+
+---
+
+## Root-cause summary
+
+1. Clay ramped **kimi-k2p6** hard around **01:16–01:21 UTC**.
+2. The rate limiter rejected **62 requests** as **`tokens_prompt` / HTTP 429** at **01:21–01:23**.
+3. Adaptive prompt / cache-adjusted / generated limits were mid-raise; cache-adjusted prompt was already **at upper bound**.
+4. Smoothed TPM utilization metrics therefore **understate** why 429s happened — this is a **burst / adaptive-ramp** problem on the prompt-token dimension, not “they sat at 100% of the published ceiling for hours.”
+
+---
+
+## Suggested follow-ups
+
+1. Confirm with Clay the client concurrency / retry behavior around 01:20 UTC (likely a batch or agent fan-out).
+2. If they need higher prompt/uncached headroom on kimi-k2p6, raise the **cache-adjusted prompt ceiling** (already at upper bound) and/or grant a reservation — adaptive alone cannot lift it further.
+3. Advise gradual ramp + exponential backoff on 429; document that `rejection_reason=tokens_prompt` can fire during bursts even when Grafana averages look &lt;50%.
+4. Investigate why `tokens_prompt_rate_mirror_per_account` went to 0 during the incident (observability gap on the exact rejecting dimension).
+
+---
+
+## Prometheus queries used
+
+```promql
+# Adaptive limits
+global_rlm_tokens_prompt_limit_per_account{account="ryan-chang-33dcc0", deployment="accounts/fireworks/deployments/kimi-k2p6"}
+global_rlm_tokens_cache_adjusted_prompt_limit_per_account{account="ryan-chang-33dcc0", deployment="accounts/fireworks/deployments/kimi-k2p6"}
+global_rlm_tokens_generated_limit_per_account{account="ryan-chang-33dcc0", deployment="accounts/fireworks/deployments/kimi-k2p6"}
+global_rlm_tokens_cache_adjusted_prompt_limit_at_upper_bound_per_account{account="ryan-chang-33dcc0", deployment="accounts/fireworks/deployments/kimi-k2p6"}
+
+# Rate mirrors
+tokens_prompt_rate_mirror_per_account{account="ryan-chang-33dcc0", deployment="accounts/fireworks/deployments/kimi-k2p6"}
+tokens_cache_adjusted_prompt_rate_mirror_per_account{account="ryan-chang-33dcc0", deployment="accounts/fireworks/deployments/kimi-k2p6"}
+tokens_generated_rate_mirror_per_account{account="ryan-chang-33dcc0", deployment="accounts/fireworks/deployments/kimi-k2p6"}
+
+# Rejections (authoritative)
+requests_total_per_account_total{account="ryan-chang-33dcc0", deployment="accounts/fireworks/deployments/kimi-k2p6", rejection_reason="tokens_prompt"}
+```