feat(providers): add Claude Opus 5.5 support

[ggiak]

Refs #14612 — covers the first ask (add claude-opus-5-5 to the catalog). The second ask, operator-defined models, is not covered here.

Summary

Adds Claude Opus 5.5 (claude-opus-5-5) on the surfaces where Anthropic lists it at launch: Anthropic API, Claude OAuth, Bedrock (anthropic.claude-opus-5-5) and Vertex/Vertex-partner. The pattern follows #8464 (Opus 5) and #12367 (Fable 5.1).

Opus 5.5 has the same context window, output limit and tokenizer as Opus 5 (1M context, 128K output). It costs less: $4 / $20 per MTok, $0.20 cache read and $5 cache write. It also has the same breaking changes Fable 5.1 introduced, so its spec reuses the existing Fable 5.1 flags. No new normalization code was needed:

Opus 5.5 API behavior	What the spec does
thinking.type:"disabled" returns 400 at every effort (Opus 5 allowed it up to high)	Sets rejectsThinkingDisabled, so normalizeThinkingForModel drops the field. Does not set maxEffortWhenThinkingDisabled.
budget_tokens / type:"enabled" returns 400	Sets adaptiveThinkingOnly, so normalizeClaudeAdaptiveThinking collapses them to adaptive.
Forced tool_choice (any / tool) returns 400	Sets rejectsForcedToolChoice, so normalizeForcedToolChoiceForModel relaxes it and keeps the tools.
1M context is native	Added to CONTEXT_1M_NATIVE_MODELS. The claude-opus-5- prefix match there doesn't cover -5-5, so an explicit entry is needed. The context-1m beta is not forwarded.
Sampling params are rejected	temperature / top_p / top_k go into unsupportedParams on anthropic and claude.

Other changes:

• Family fallback goes claude-opus-5-5 → claude-opus-5 → claude-opus-4-8 → claude-sonnet-5.
• Added to the Fast Mode default model lists (claudeFastMode.ts and the settings default). Fast mode is available on the Claude API for Opus 5.5.
• Added to Claude OAuth static discovery. It is listed after claude-opus-5 in the claude registry, so getClaudeCodeDefaultModels().opus (the CLI-tool cc/ Opus default, ANTHROPIC_DEFAULT_OPUS_MODEL) stays claude-opus-5. claude/anthropic validate against an authoritative live catalog, and a connection synced before the launch returns model_not_found for claude-opus-5-5 until its models are refreshed. A test pins this default.
• Default reasoning effort is medium (defaultReasoningEffort). A client-chosen effort still wins. Ported from [defer] fix(providers): declare claude-opus-5 thinking efforts in claude registry #14548.
• Added a >= 2.1.280 floor assertion on the Claude Code client pin (Anthropic's Opus 5.5 gate). Ported from [defer] fix(providers): bump Claude Code client pin 2.1.258 -> 2.1.280 to clear the Opus 5.5 model gate #14557.
• Pruned one stale no-unused-vars suppression for vertex/index.ts. It was already stale on the base, and touching the file made the pre-commit ESLint step fail.

Deliberately not included

Claude Web, GitHub Copilot / GHE Copilot, Kiro, Cursor and Devin each use their own wire model IDs or thinking contracts. I couldn't verify those for Opus 5.5, so the test asserts they do not expose it yet. They should be follow-ups once someone checks them against a live catalog. Also not done: fast-mode-specific pricing ($8 / $40 per Anthropic's docs) and the computer_toolset_20260801 requirement (OmniRoute doesn't translate computer-use tool versions).

Operator note

Existing claude / anthropic connections need a models refresh (provider page → refresh models) before claude-opus-5-5 routes. Until then the live catalog rejects it with model_not_found.

Credits

• @PixmaNts: Opus 5.5 gate floor assertion, from [defer] fix(providers): bump Claude Code client pin 2.1.258 -> 2.1.280 to clear the Opus 5.5 model gate #14557 (co-authored commit).
• @HouMinXi: defaultReasoningEffort: "medium", from [defer] fix(providers): declare claude-opus-5 thinking efforts in claude registry #14548 (co-authored commits).

Validation

• New test tests/unit/claude-opus-5-5.test.ts covers catalogs per provider, unverified surfaces staying unexposed, the spec flags, the thinking and tool-choice normalizers, the fast-mode list, pricing, fallback, the CLI Opus default staying on Opus 5, and the medium effort default.
  • Before the change: 6/7 failed ("anthropic must expose claude-opus-5-5", …). After: 9/9 pass. The effort-default test failed before the spec field was added.
• claude-codex-identity-version-sync: 8/8 pass, including the new >= 2.1.280 floor test.
• Ran the 117 existing unit test files that reference these catalogs, specs, pricing, fallback or fast mode: 1234/1242 pass on this branch.
  • pricing-fallback pinned the KNOWN_MODEL_PRICING size (21). Updated it to 22 for the new row.
  • 2 tests/unit/autoCombo/* files are Vitest suites. Run under vitest.mcp.config.ts they pass (21/21).
  • The remaining 5 failures (chatcore-translation-paths Codex/DeepSeek ×4, qwen38-max-bare-id-alias ×1) fail identically on the untouched base ea3c12264b. They are inherited and unrelated to this change.
• npm run typecheck:core is clean. ESLint with the suppressions file and Prettier pass on the changed files. All Husky pre-commit gates passed.

[ggiak]

🤖 Automated review — 5f0ef1c

2 important · 5 nits · 2 pre-existing · effort high

🔴 Important

1. open-sse/config/providers/registry/bedrock/index.ts:24: the Bedrock anthropic.claude-opus-5-5 entry has no unsupportedParams. The claude entry at open-sse/config/providers/registry/claude/index.ts:59 strips temperature/top_p/top_k; the Bedrock one does not.
   Fails when: a client sends {model:"bedrock/anthropic.claude-opus-5-5", temperature:0.7}. getUnsupportedParams returns [], openAIToBedrockConverse sets inferenceConfig.temperature, and Bedrock rejects the request with a 400.
2. src/shared/constants/modelSpecs.ts:483: the Opus 5.5 spec has no dot-notation alias. Opus 4.8 has one (claude-opus-4.8) at line 512.
   Fails when: a client sends claude-opus-5.5 with thinking:{type:"disabled"} or a forced tool_choice. The prefix match 'claude-opus-5.5'.startsWith('claude-opus-5') picks the Opus 5 spec, so rejectsThinkingDisabled/rejectsForcedToolChoice never apply. The fields are forwarded and upstream returns a 400. modelHasNativeContext1m("claude-opus-5.5") is also wrong.

🟡 5 nits · 🟣 2 pre-existing

• 🟡 open-sse/services/providerCostData.ts:17: pricing only covers the bare claude-opus-5-5.
  Fails when: getModelPricing("bedrock","anthropic.claude-opus-5-5") misses and falls back to $5/$15 instead of $4/$20. That misprices the Bedrock target for cost-optimized routing and virtual cost.
• 🟡 open-sse/config/providers/registry/claude/index.ts:51: putting Opus 5.5 before Opus 5 implicitly moves the opus default. getClaudeCodeDefaultModels().opus, and so ANTHROPIC_DEFAULT_OPUS_MODEL, now resolves to cc/claude-opus-5-5.
  Fails when: generated Claude Code configs start sending all opus-alias traffic to 5.5, which has a stricter contract (thinking-disabled is dropped, forced tool_choice is relaxed). This happens through array order, not an explicit decision. If it is intended, say so in the PR description.
• 🟡 src/lib/providers/claudeFastMode.ts:11: the new claude-opus-5-5 entry is redundant. shouldRequestClaudeFastMode already prefix-matches claude-opus-5-.
  Fails when: the new test only asserts list membership. Removing the entry would still leave fast mode on for 5.5, so the test does not pin behavior.
• 🟡 src/lib/db/settings.ts:213: the default claudeFastMode.supportedModels is a hand-copied duplicate of CLAUDE_FAST_MODE_DEFAULT_MODELS.
  Fails when: a future model is added to one list and not the other, and the settings-UI default and the runtime default diverge. Using [...CLAUDE_FAST_MODE_DEFAULT_MODELS] removes the duplicate.
• 🟡 tests/unit/claude-opus-5-5.test.ts:54: the test asserts that github, ghe-copilot, kiro, claude-web and cursor do NOT expose Opus 5.5.
  Fails when: someone lands the follow-up the PR description asks for (adding 5.5 to a verified catalog). This unrelated test then fails by design.
• 🟣 open-sse/config/providers/registry/bedrock/index.ts:28: the entry advertises supportedThinkingEfforts, but the Bedrock Converse translator never maps reasoning_effort into additionalModelRequestFields. The Opus 5 entry has the same gap.
  Fails when: a client picks effort max for bedrock/anthropic.claude-opus-5-5. The effort is silently dropped.
• 🟣 src/lib/providers/claudeFastMode.ts:11: Claude fast-tier pricing is never applied. Only Codex has a fast cost multiplier.
  Fails when: Opus 5.5 fast-mode traffic is recorded at $4/$20 while Anthropic bills the fast rate, so usage and budgets under-report spend.

[ggiak]

Addressed in f76f105:

• Bedrock sampling params (important): anthropic.claude-opus-5-5 now declares unsupportedParams: ["temperature", "top_p", "top_k"]. The test now includes a Bedrock row. Before the fix, getUnsupportedParams("bedrock", …) returned []. The anthropic.claude-fable-5-1 entry has the same gap on the base; I left it out of this PR.
• claude-opus-5.5 dot notation (important): the Opus 5.5 spec now has a claude-opus-5.5 alias, so that ID no longer prefix-matches the Opus 5 spec. modelHasNativeContext1m also normalizes dotted versions (5.5 → 5-5) before matching. I didn't add a dotted entry to CONTEXT_1M_NATIVE_MODELS because claude-context-1m-supported-models requires every entry there to match a registry ID. New assertions check the spec identity, the disabled-thinking drop and native 1M for the dotted ID.
• Fast-mode test: it now calls shouldRequestClaudeFastMode with the toggle on and off, and with a narrowed supportedModels, instead of just checking list membership.
• Opus default order: this is intended, and the PR description says so. The claude registry lists newest models first, as getClaudeCodeDefaultModels expects.
• Not changed here: Bedrock pricing falls back to $5/$15 for every Bedrock Claude model, and the Converse translator never maps reasoning_effort. Both predate this PR. Also not changed: the duplicated settings.ts fast-mode default, which is a separate refactor. The tests that assert unverified surfaces don't expose Opus 5.5 stay, mirroring the Fable 5.1 test, so any follow-up that adds a catalog has to change them on purpose.

Focused test: 7/7. Related unit tests (115 files): 1235/1240. The 5 failures are the same ones that fail on the base ea3c12264b.

[ggiak]

Added 785b5ce: bump the Claude Code wire identity from 2.1.258 to 2.1.280. The Claude OAuth path rejected Opus 5.5 with 400 Claude Code 2.1.258 does not support this model; version 2.1.280 or newer is required. This follows the same steps as #12604.

• Captured values: taken from the real 2.1.280 CLI through a loopback listener (authorization redacted). The UA is claude-cli/2.1.280 (external, sdk-cli) and the billing header is cc_version=2.1.280.d7b. X-Stainless-Package-Version 0.112.1 and runtime v26.3.0 are unchanged.
• Tests first: I moved the drift-guard tests to the new values before bumping the constant. They failed against the old pin (10 failures), and all 156 guard/golden tests pass now.
• Golden snapshot: only the six User-Agent lines changed. A full UPDATE_GOLDEN=1 regen would also rewrite a DeepSeek entry that has already drifted on the base, so I left that out.
• Devin bridge: the Docker pin stays at 2.1.258 until someone live-validates it.

[diegosouzapw]

Thanks @ggiak, this is the most complete of the four Opus 5.5 PRs, and we'll land it as the primary.

Two things we checked: d7b is correct. The 2.1.280 binary computes the suffix per request as sha256("59cf53e54c78" + chars[4,7,20] + version)[:3], which gives 1e2 for 2.1.258 and d7b for 2.1.280 on a short first message. Pricing, 1M context and the thinking/tool_choice flags also match the launch docs.

One thing before merge: claude/anthropic use an authoritative live catalog, so a connection synced in the last 30 days still returns model_not_found for claude-opus-5-5 until it's refreshed (we reproduced it). Since this PR also makes Opus 5.5 the cc/ Opus default, please either keep claude-opus-5 as the CLI default for now, or note the "refresh models" step in the PR body and changelog. Please also reference the issue as Refs #14612 (its second ask, operator-defined models, isn't covered here).

We'd also like to fold in #14557's ≥2.1.280 floor assertion and #14548's defaultReasoningEffort: "medium", with co-author credit to both authors.

[ggiak]

Thanks @diegosouzapw, all four are in (head 7f2eecc):

• CLI default: took the first option. claude-opus-5-5 now sits after claude-opus-5 in the claude registry, so getClaudeCodeDefaultModels().opus / ANTHROPIC_DEFAULT_OPUS_MODEL stays claude-opus-5. A test pins it (ded7b44). I also put the "refresh models" note in the PR body and the changelog fragment, for anyone who routes to claude-opus-5-5 directly on a stale connection.
• Issue ref: the PR body now opens with Refs #14612 and says the operator-defined-models ask isn't covered.
• [defer] fix(providers): bump Claude Code client pin 2.1.258 -> 2.1.280 to clear the Opus 5.5 model gate #14557: ported the >= 2.1.280 floor assertion into claude-codex-identity-version-sync.test.ts (52d175b), co-authored by @PixmaNts.
• [defer] fix(providers): declare claude-opus-5 thinking efforts in claude registry #14548: defaultReasoningEffort: "medium" on the Opus 5.5 spec, with a test that an effort the client sets still wins (31318a8). Co-authored by @HouMinXi.

Focused suites pass: claude-opus-5-5 9/9, claude-codex-identity-version-sync 8/8, claude-cli-defaults. The DeepSeek/Codex/Qwen failures listed in the PR body are unchanged. The provider translate-path golden snapshot also fails, only on DeepSeek's endpoint format. This branch doesn't touch that, but I haven't re-run it on the base.

[HouMinXi]

Heads-up on scope so we don't collide: I've narrowed #14548 to only the part this PR doesn't cover — claude-opus-5 (not 5.5) gains supportsReasoning + supportedThinkingEfforts: ["low", "medium", "high", "xhigh", "max"] in the claude registry.

Why that entry needs it: the catalog calls getThinkingCapabilityFields(..., skipCanonicalEffortFallback=true) for static registry models (src/app/api/v1/models/catalog.ts:730), so when the registry entry declares no tier list the model gets no effort_tiers in /v1/models at all — the canonical fallback is deliberately skipped there. claude-opus-5's modelSpec has supportsThinking: true, but its registry entry never declared the tiers, so the static catalog exposes the model without thinking tiers. Your new claude-opus-5-5 entry declares them, so 5.5 is covered here.

Happy for you to fold those two lines into this PR if you'd rather land it in one place; otherwise #14548 stays a two-line companion that merges cleanly either way — it no longer touches claude-opus-5-5 or modelSpecs.ts, and it keeps your defaultReasoningEffort: "medium" credit intact by simply not duplicating it.

[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.

[HouMinXi]

The new claude-opus-5-5 registry entry does not set supportedThinkingEfforts. models.dev lists low, medium, high, xhigh, max for anthropic/claude-opus-5.5 (same set as claude-opus-5 in this registry). Without the array, the dashboard catalog emits no effort_tiers for that id, because static registry models skip the canonical fallback. Sonnet 5.5 in this same PR already lists its tiers (low through xhigh). Opus 5.5 should list low through max the same way, in both the claude and anthropic registries.