Update OpenAI and Gemini pricing to current published rates

[mando]

Summary

Checked every price family against the providers' current rate cards.

Claude rates were already correct (Fable and Mythos, Opus 5 through 4.5, the Opus 4.1/4 legacy tier, Sonnet 5's introductory rate, Sonnet 4.x, Haiku 4.5, Haiku 3.5 legacy), so nothing changed there.

Gemini had drifted furthest: a single gemini-3-flash entry was pricing every Gemini 3.x Flash model. Adds 3.6 Flash, 3.5 Flash, 3.5 Flash-Lite and 3.1 Flash-Lite at their own rates, plus long-context tier pairs for 3.1 Pro and 3 Pro. gemini-3-flash stays as the fallback for older transcripts. The Gemini branch now shares one 200k-token check, so the tier step applies to all three Pro generations instead of only 2.5.

OpenAI had several variants quietly billing at their base model's rate, because they matched it by substring:

Model	Was billing	Actual
gpt-5-pro	$1.25 / $10	$15 / $120
gpt-5.4-pro	$2.50 / $15	$30 / $180
gpt-5.5-pro	$5 / $30	$30 / $180
gpt-5.2-pro	$1.75 / $14	$21 / $168
gpt-5-mini	$1.25 / $10	$0.25 / $2
gpt-5-nano	$1.25 / $10	$0.05 / $0.40
gpt-5.1	unpriced (cost 0)	$1.25 / $10

Each gets its own entry, matched ahead of the base model. gpt-5.1 was unpriced entirely because the model-id regex rejected the dot; it now resolves to the gpt-5 rate it shares. The pro tiers get no cached-input discount, so their cacheRead is the full input rate rather than a tenth of it.

Also fills in the price-family table in the settings reference, which was missing fable, mythos, opus-legacy, sonnet-5, haiku-legacy, the whole gpt-5.6 family and gpt-5.4-nano. Names gpt-5.6-sol on the gpt-5.6 row, since OpenAI publishes that tier suffixed and the override key isn't otherwise guessable. Stamps a verification date on the defaults so the next refresh knows the baseline.

Notes for review

• Two entries carry rates that have dropped off the providers' current pages but are kept for historical transcripts: gemini-3-flash and codex-mini.
• Deliberately no separate gpt-5.6-sol key. Two keys with the same rate is a footgun, since overriding one in pricing.json would silently leave the other at the default.

Test plan

• bun test (694 pass)
• bun run typecheck
• New cases cover the mini/nano/pro tiers, gpt-5.1, and the Gemini 3.x rates and Pro tier boundaries

🤖 Generated with Claude Code