feat(models): Opus 5.5 and GPT-6 Sol/Luna pricing and current-generation defaults

[pacphi]

Summary

Anthropic and OpenAI released new models on 2026-09-22. This PR adds their pricing, fixes two matcher bugs that were mispricing real usage, and moves the routing defaults to the current generation. Every rate was verified on 2026-09-23 against first-party pages.

New models (USD per million tokens)

Model	Input	Cache read	Cache write	Output	Source
claude-opus-5-5	$4	$0.20 (0.05×)	$5 (5-minute) / $8 (1-hour)	$20	Anthropic pricing
gpt-6-sol	$2	$0.20	$2.50	$10	model page
gpt-6-luna	$0.10	$0.01	$0.125	$0.50	model page

Pricing fixes

• Opus 5.5 was costed as Opus 5. claude-opus-5-5 had no entry, so it matched the claude-opus-5 entry. Re-costing local Opus 5.5 usage (273 messages) gives $19.51, not $37.89: the old cost was 94% too high.
• Provider-prefixed IDs such as anthropic/claude-opus-4.7 fell through to the $3/$15 fallback. They now match on the model name after the last /.
• codex-auto-review stays on the flagged fallback rate. OpenAI publishes no price for it. Codex's backend maps it to some model; the last public mapping was gpt-5.4, per openai/codex#18169.
• All other rates re-verified and unchanged. Third-party summaries give GPT-5.6 Sol as $5/$30; OpenAI's own page says $4/$20, which we keep.

Default model changes

Routes	Before	After	Why
Claude reasoning routes and the implementation/testing escalation step	claude-opus-5	claude-opus-5-5	Anthropic's recommended starting model (overview)
Codex implementation, testing, security-scan	gpt-5.6-terra	gpt-6-sol	Codex's catalog ranks it first ("Workhorse model for coding"); OpenAI's Codex docs make Sol the default preset
Codex documentation, packaging	gpt-5.6-luna	gpt-6-luna	Codex's catalog labels every GPT-5.6 model "Older"

• Tier pairing when mirroring routes between hosts follows role and per-token price: Sol ↔ Sonnet 5 ($2/$10 each), Astra ↔ Fable 5.1 ($10/$50 each), Luna ↔ Haiku.
• Earlier models stay listed so existing pins keep working. Routes seeded before this change show as diverged and keep their model until ak x host refresh.
• Not added as retirements: OpenAI retires GPT-5.4 and 5.4 Mini (2026-08-31) and GPT-5.5 (2026-10-14) only for ChatGPT sign-in; the API still serves them. A retirement entry would override the pins of API-key users.

Other changes

• Opus 5.5 lifecycle facts in the Anthropic catalog, and OpenAI doc links for Sol and Luna.
• The divergence and dashboard tests now derive their "prior catalog" fixtures from DEFAULT_ROUTES, so the next default change doesn't break them.
• Docs: MODEL-PRICING-AUDIT.md, USAGE-SCORECARD-METRICS.md (line-number citations replaced with names), PROVIDERS.md, MODELS.md, and ADR-0002 (Updated note).

Verification

• pnpm run check: exit 0. Tests: 4324 pass, 0 fail, 6 skipped. Lint: 0 errors, with the same 71 warnings as main.
• The new pricing tests fail on the old code and pass with this change. They cover Opus 5.5 rows and multipliers, GPT-6 Sol and Luna, provider-prefixed IDs, and the token-boundary guard.

🤖 Generated with Claude Code