feat(claude): add Claude Sonnet 5.5 and correct Sonnet 5 pricing

[HouMinXi]

Summary

Claude Sonnet 5.5 (claude-sonnet-5-5) shipped on 2026-09-28. A request routed to it currently fails upstream in three ways this change closes:

• thinking: {"type":"disabled"} is rejected. The new model spec sets rejectsThinkingDisabled, so normalizeThinkingForModel drops the field and the model uses its adaptive default.
• Forced tool_choice (any, tool, function) returns 400. rejectsForcedToolChoice makes normalizeForcedToolChoiceForModel drop it back to the default auto behavior.
• Reasoning effort tops out at xhigh; max returns 400. The model is excluded from supportsClaudeMaxEffort, so a max request is lowered to high.

The published rate for Sonnet 5 and Sonnet 5.5 is $2 input, $10 output, $0.20 cache reads, and $2.50 for a five-minute cache write. The catalog still carried the retired $3/$15 rate. reasoning is not published as its own rate; $10 follows the output price.

Test plan

• node --import tsx/esm --test tests/unit/claude-sonnet-5-5.test.ts — 4 passed
• tests/unit/catalog-updates-v3x.test.ts passed
• Injected defect: removing the two reject flags fails the thinking/tool-choice test; removing the max-effort exclusion fails the effort test. Restoring each passes.

[ggiak]

🤖 Automated review — a8d8a24

5 important · 5 nits · 1 pre-existing · effort high

External facts not verifiable from the repo: the $2/$10 Sonnet 5 / 5.5 list price, and whether Sonnet 5.5 rejects temperature/top_p/top_k and needs the 1M-context beta the way Sonnet 5 does. Please confirm these against Anthropic's docs. The findings below judge whether the repo is internally consistent with what this PR claims.

🔴 Important

1. open-sse/config/providers/registry/claude/index.ts:100 — claude-sonnet-5-5 is not added to any provider registry (claude, anthropic, vertex, vertex/partner). The sibling Opus 5.5 PR (#14546) added it to all of these.
   Fails when: cc/claude-sonnet-5-5 or anthropic/claude-sonnet-5-5 is requested. The model is missing from /v1/models and the provider catalogs. getUnsupportedParams() (open-sse/config/providerRegistry.ts:261) finds no unsupportedParams, so a client-sent temperature/top_p is forwarded. Sonnet 5 lists these because the adaptive-only model returns 400 on them, and the new spec marks 5.5 adaptiveThinkingOnly: true too. There are also no supportedThinkingEfforts, which causes finding 2.
2. open-sse/executors/base/reasoningEffort.ts:624 — the executor sanitizer still forwards effort: "max" for Sonnet 5.5. The model has no registry supportedThinkingEfforts, and supportsXHigh defaults to true, so the body is returned unchanged. supportsClaudeMaxEffort is not read on this path.
   Fails when: a native Anthropic or Claude Code request with output_config.effort: "max" is sent to claude-sonnet-5-5. Upstream receives max and returns the 400 this PR is meant to prevent. A registry entry with supportedThinkingEfforts: ["low","medium","high","xhigh"] would lower max to xhigh here.
3. src/shared/constants/modelSpecs.ts:408 — only the exact id claude-sonnet-5-5 gets the new spec. Dated, -thinking and dotted ids fall through to prefix matching, where the earlier claude-sonnet-5 key wins. There is also no claude-sonnet-5.5 dot alias, although 4.6 has one.
   Fails when: getCanonicalModelSpecId resolves claude-sonnet-5-5-20260915, claude-sonnet-5-5-thinking and claude-sonnet-5.5 to claude-sonnet-5 (checked at the PR head). rejectsThinkingDisabled and rejectsForcedToolChoice then never apply, so thinking: {type:"disabled"} or tool_choice: {type:"any"} passes through and upstream returns 400.
4. src/shared/constants/pricing/frontier-labs.ts:222 — no claude-sonnet-5-5 key in the anthropic DEFAULT_PRICING table, and none in the cc table in oauth-subscriptions.ts. #14546 added claude-opus-5-5 to both. getPricingForModel (src/shared/constants/pricing.ts:18) matches exact keys only.
   Fails when: usage for anthropic/claude-sonnet-5-5 or cc/claude-sonnet-5-5 resolves to null pricing, so it is costed at $0 unless the synced LiteLLM/models.dev layer has the model. The new test does not catch this: getModelPricing("claude", …) passes only through the generic KNOWN_MODEL_PRICING fallback in providerCostData.ts.
5. src/shared/constants/pricing/oauth-subscriptions.ts:81 — the Sonnet 5 price change was applied to CLAUDE_SONNET_5_PRICING (anthropic) and KNOWN_MODEL_PRICING, but the cc table still prices claude-sonnet-5 at $3 / $15 / $0.30 cached / $3.75 cache-write.
   Fails when: the same model is costed 1.5× higher through cc/ than through anthropic/, so the dashboard shows two different Sonnet 5 prices. Other tables look consistent: devin.ts is already $2/$10, inference-hosts.ts ($1.40/$7) is a host discount, and the kiro table ($3/$15, line 464) is kept on purpose per the updated test comment.

🟡 5 nits (+2 more) · 🟣 1 pre-existing

• 🟡 open-sse/translator/request/openai-to-claude.ts:224 — when a model doesn't support max, it is lowered to high. Sonnet 5.5 accepts xhigh, so max should become xhigh. Fails when: reasoning_effort: "max" on claude-sonnet-5-5 goes upstream as high, two tiers below what was asked. claudeCodeCompatible.ts:462 does the same.
• 🟡 open-sse/config/providerModels.ts:330 — the new pattern only matches the hyphen form, and supportsClaudeMaxEffort does not turn dots into hyphens. Fails when: supportsClaudeMaxEffort("claude-sonnet-5.5") returns true and max is forwarded.
• 🟡 open-sse/config/context1m.ts:17 — claude-sonnet-5-5 is not eligible for the 1M beta. The digit guard in modelSupportsContext1mBeta rejects it as a child of claude-sonnet-5, while the spec says contextWindow: 1000000. Fails when: a claude-sonnet-5-5[1m] request has its context-1m beta header stripped, if 5.5 needs the beta (needs confirmation).
• 🟡 open-sse/services/modelFamilyFallback.ts:94 — no claude-sonnet-5-5 fallback chain, and no entry in src/lib/providers/staticModels.ts:44. #14546 added both for Opus 5.5. Fails when: a family fallback from claude-sonnet-5-5 finds no siblings, and the static catalog does not list the model.
• 🟡 src/shared/constants/modelSpecs.ts:409 — the comment says "(use between_tools)", but normalizeThinkingForModel just deletes thinking. Fails when: a reader expects a between_tools rewrite that the code never does.
• (+2 more nits) Adding a single model version to CLAUDE_MAX_EFFORT_UNSUPPORTED_FAMILY_PATTERNS (meant for whole families) is the wrong level; a per-model supportedThinkingEfforts would cover every effort path. The four-line pricing comment at shared-tiers.ts:121 repeats the values set just below it.
• 🟣 open-sse/config/providers/registry/bedrock/index.ts:24 — the Bedrock registry lists no Sonnet 5 at all (only 4.6 / 4.5), so Bedrock Sonnet-5-family ids never get registry unsupportedParams. This predates this PR, but #14546 had to add Bedrock sampling stripping for Opus 5.5, and Sonnet 5.5 probably needs the same.

[diegosouzapw]

Re-homed to release/v3.8.52: v3.8.51 entered its release freeze, so the branch now belongs to the release captain and development continues on the next cycle. Nothing is wrong with this PR — it just needed a live base. No action needed from you; CI will re-run against the new base.