🤖 feat: add support for Claude Opus 5 - #3750
Merged
Merged
Conversation
Promotes Claude Opus 5 (released July 24, 2026) to the default opus/OPUS model in place of Opus 4.8. Same pricing and envelope as 4.8: $5/$25 per MTok, native 1M context, 128K max output. The existing native-xhigh regex already routes Opus 5 through adaptive thinking, summarized display, and the full effort ladder including native xhigh and max; the Opus 5 restriction that disabled thinking is only legal at effort high or below is satisfied because "off" maps to effort low. --- _Generated with `mux` • Model: `anthropic:claude-fable-5` • Thinking: `xhigh` • Cost: `$13.54`_ <!-- mux-attribution: model=anthropic:claude-fable-5 thinking=xhigh costs=13.54 -->
Contributor
Author
|
@codex review |
|
Preview deployment for your docs. Learn more about Mintlify Previews.
💡 Tip: Enable Workflows to automatically generate PRs for you. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 873a944914
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Contributor
Author
|
@codex review |
|
Codex Review: Didn't find any major issues. Swish! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
This was referenced Aug 27, 2026
yermakoffivan
pushed a commit
to yermakoffivan/mux
that referenced
this pull request
Sep 6, 2026
## Summary Prepares Xum for the Claude Fable 5.1 drop: promotes Fable 5.1 (`anthropic:claude-fable-5-1`) to the `FABLE` known model, so the `fable` alias, the `/fable` command, tokenizer warming, and the default refusal-fallback chain all route to the new model. Fable 5 stays usable as the custom model string `anthropic:claude-fable-5` and keeps its metadata entry and tokenizer approximation. > [!NOTE] > **Launch confirmed.** Anthropic released Claude Fable 5.1 on September 1, 2026. All assumptions below were verified against the official model overview; the API id and $10/$50 pricing were correct as assumed. The one correction: cache reads are $0.25/M (0.025x input), a quarter of the assumed 0.1x ratio, fixed in `models-extra.ts`. ## Background This PR was prepared ahead of the drop (same play as the Opus 5 promotion in coder#3750), with assumptions confirmed on release day against the official docs: - **API id**: `claude-fable-5-1` (dash form, dateless, matching the 5-generation convention of `claude-sonnet-5` / `claude-opus-5` and the dash style of `claude-haiku-4-5`). - **Pricing**: unchanged from Fable 5 for input/output, $10/M input, $50/M output, cache write 1.25x input. Cache read turned out cheaper at launch: $0.25/M (0.025x input), corrected in this PR. - **Envelope**: unchanged, native 1M context, 128K max output, full effort ladder with native xhigh and max, adaptive thinking only (disabled thinking rejected), safeguard classifiers retained so the refusal fallback stays relevant. ## Implementation - `knownModels.ts`: `FABLE` -> `claude-fable-5-1` (keeps the `fable` alias and tokenizer warming; tokenizer override stays on the Opus 4.5 approximation). - `models-extra.ts`: new `claude-fable-5-1` pricing/limits entry; the Fable 5 entry is retained. - `models.ts`: `claude-fable-5-1` added to the native-1M patterns (the existing `claude-fable-5` pattern only tolerates date suffixes, not `-1`). - No wire-format changes needed for thinking/effort: the Mythos-class wildcard matchers (`anthropicSupportsNativeXhigh`, `anthropicRejectsDisabledThinking`) and native web-fetch id parsing already match the new id; tests now pin that. - **Fallback seed migration**: `DEFAULT_MODEL_FALLBACKS` is keyed by the FABLE id and the one-time seed is guarded by `migrations.defaultModelFallbacksSeeded`, so already-seeded configs would never get a chain for the new source key. A second one-shot flag (`defaultModelFallbacksSeededFable51`) re-runs the gap-check once: pre-5.1 configs get the Fable 5.1 -> Opus chain, an existing 5.1 chain is never overwritten, and the legacy Fable 5 chain is left byte-identical on disk. - **Downgrade safety** (Codex round 1): completing the original seed pass claims `defaultModelFallbacksSeeded`, which pre-5.1 builds trust for their own `claude-fable-5` key. The original seed pass (and the fresh-install default config) now also carries a `LEGACY_DEFAULT_MODEL_FALLBACKS` chain for `anthropic:claude-fable-5`, so a downgrade still finds its expected default instead of silently losing refusal fallback. - **Legacy tokenizer overrides** (Codex round 1): `TOKENIZER_MODEL_OVERRIDES` is derived from current KNOWN_MODELS ids, so retiring an id dropped its approximate-tokenizer mapping. A `LEGACY_TOKENIZER_MODEL_OVERRIDES` map retains `anthropic:claude-fable-5` and the identical pre-existing case `anthropic:claude-opus-4-8`; exact-id lookup resolves both to the intended Opus 4.5 approximation instead of warning and falling back to the generic per-provider tokenizer. - Docs model table and built-in skill content regenerated. - **Launch-day verification**: rebased onto main and corrected the `models-extra.ts` cache-read cost to the official $0.25/M; everything else matched the official model page. ## Validation - Behavioral coverage added for the new id: native-1M classification, display formatting (`Fable 5.1`), 5-level thinking policy with off-clamp, provider options (adaptive + summarized display, no disabled thinking), native web-fetch support, and the config seeding semantics (legacy chain carried by the original seed pass, 5.1 chain seeded once for pre-5.1 configs, existing 5.1 chains and deletions respected). - The WorkflowRunner alias-mapping test caught the alias flip (`fable` -> 5.1) and was updated; remaining `claude-fable-5` fixtures were audited and stay valid as explicit custom model strings. ## Risks Low: additive registry/metadata changes plus config migration flags. The main user-visible effect is that the `fable` alias and `/fable` route to Fable 5.1. Until the model exists at the API, alias sends would fail, which is why the merge gate above matters. Existing Fable 5 selections and user-edited fallback chains are unaffected; note that with current ai-tokenizer data the legacy tokenizer fix changes no token counts (both approximations share the `claude` encoding), it removes warning spam and future drift risk. --- _Generated with `xum` • Model: `anthropic:claude-fable-5` • Thinking: `xhigh` • Cost: `$45.64`_ <!-- mux-attribution: model=anthropic:claude-fable-5 thinking=xhigh costs=45.64 -->
asm
pushed a commit
to asm/mux
that referenced
this pull request
Sep 22, 2026
## Summary Prepares Xum for the Claude Opus 5.5 drop: promotes Opus 5.5 (`anthropic:claude-opus-5-5`) to the `OPUS` known model, so the `opus` alias, the `/opus` command, tokenizer warming, the default model, the mux-gateway first-time model list, and the CI agent workflows all route to the new model. Opus 5 stays usable as the custom model string `anthropic:claude-opus-5` and keeps its metadata entry and tokenizer approximation. > [!IMPORTANT] > **Opus 5.5 launched Sept 22, 2026** ([announcement](https://www.anthropic.com/claude-opus-5-5), [what's new](https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5)). Every provisional value below was verified against the official docs in `c86761dae`; the launch checklist tracks what is still open. Undrafted Sept 22 18:07 UTC on the maintainer's instruction to merge once CI is green; that triggered Codex's final-head code + security reviews on `c86761dae` (both completed clean, 👍). **Scope note (Sept 22):** this PR was originally prepared as an _Opus 5.1_ promotion. Opus 5.1 never shipped — the Sept 1 wave was Fable 5.1 / Mythos 5.1 only and `main` still ships `claude-opus-5` — so rather than promoting a nonexistent id, the same PR was retargeted in place to the anticipated Opus 5.5 launch (mechanical id/label rename + repricing; commit `605816945`), then aligned with the official launch docs (`c86761dae`). The review record below was earned on the 5.1-era diff and is preserved; no new review loops were run for the retarget or the alignment. ## Background This PR was prepared ahead of the drop (same play as the Opus 5 promotion in coder#3750 and the Fable 5.1 promotion in coder#3988) and then verified against the official docs on launch day: - **API id** ✅ `claude-opus-5-5` (Claude API; Bedrock `anthropic.claude-opus-5-5`, Vertex/Foundry `claude-opus-5-5`). - **Pricing** ✅ $4/M input, $20/M output; 5-minute cache write $5/M (1.25×); cache read $0.20/M (0.05×). The 1-hour cache-write tier ($8/M, 2×) and fast mode ($8/$40) have no field/entry in `models-extra.ts`, matching how Opus 5 is modelled (noted in the entry comment). - **Envelope** ✅ native 1M context, 128K max output, effort ladder with native xhigh + max; default effort `medium` (Xum's default is already `medium`). - **Breaking changes vs Opus 5** (docs "What's new"): (1) thinking **cannot be disabled** — `disabled` and `enabled+budget_tokens` both return 400 → handled in `c86761dae`; (2) forced `tool_choice` any/tool returns 400 → Xum only forces tool choice for xai, so no change; (3) thinking blocks are bound to the producing model (Opus 5 → 5.5 keeps reasoning; Fable/Mythos → 5.5 drops it) → informational; (4) `computer_20251124` not accepted → not referenced in the repo; (5) Opus 5.5 ships the Fable-class safeguard classifiers → it is not a proven refusal-fallback _target_, see Implementation. ## Implementation - `knownModels.ts`: `OPUS` -> `claude-opus-5-5` (keeps the `opus` alias, warming, and the Opus 4.5 tokenizer approximation). The retired `anthropic:claude-opus-5` id joins `LEGACY_TOKENIZER_MODEL_OVERRIDES` (alongside main's `claude-fable-5` / `claude-mythos-5` entries) so exact-id lookup keeps its approximation instead of warning and falling back to the generic per-provider tokenizer. - `models-extra.ts`: new `claude-opus-5-5` pricing/limits entry (official numbers); the Opus 5 entry is retained. - `models.ts`: `claude-opus-5-5` added to the native-1M patterns — the existing `claude-opus-5` pattern only tolerates date suffixes, not `-5` (the exact trap coder#3988 documented for Fable). - Thinking/effort wire format: `anthropicSupportsNativeXhigh` already matches any Opus 5+ (`claude-opus-[5-9]`, unanchored) and the native web-fetch parser reads `claude-opus-5-5` as major 5 / minor 5. **`anthropicRejectsDisabledThinking` now also matches the exact `claude-opus-5-5` id** (optional date suffix, Bedrock `anthropic.` prefix tolerated), so the thinking policy offers `low..max` without "off" (a requested "off" clamps to `low`, as for Fable/Mythos) and provider options omit `thinking` on a stray "off" instead of sending the rejected `{ type: "disabled" }`. Opus 5 keeps its 6-level ladder. Tests pin all of this plus display formatting (`Opus 5.5`). - **Refusal-fallback target stays Opus 5** (decision on launch day, replacing the earlier chain-target migration): Opus 5.5 ships the same safeguard classifiers as Fable 5.1, so it is not a proven recovery target for Fable refusals. `DEFAULT_MODEL_FALLBACKS` and `LEGACY_DEFAULT_MODEL_FALLBACKS` pin the literal `anthropic:claude-opus-5` instead of `KNOWN_MODELS.OPUS.id`, so fresh installs and already-seeded configs keep byte-identical `Fable 5.1 → Opus 5` / `Fable 5 → Opus 5` chains with no migration or new flag. The one-shot `defaultModelFallbacksSeededOpus55` migration and `SUPERSEDED_DEFAULT_MODEL_FALLBACKS` from the reviewed diff were removed as no longer needed (pure deletion; `config/index.ts` and `appConfigOnDisk.ts` are back to `main`). **Follow-up:** evaluate Opus 5.5 as a fallback target (or `Opus 5.5 → Opus 5` as a new default chain) once its refusal behavior is known. - Gateway first-time defaults (`providerService.ts`), CI agent workflows (auto-cleanup, terminal-bench), terminal-bench leaderboard `MODEL_METADATA`, docs model table, `docs/guides/github-actions.mdx`, and built-in skill content regenerated — mirroring coder#3750's surface set. ## Validation - Behavioral coverage for the new id: native-1M classification (incl. gateway-prefixed form), 5-level thinking policy with "off" excluded and clamped to `low` (Opus 5 and a Bedrock-style `anthropic.claude-opus-5-5` id covered as boundaries), provider options (adaptive + summarized display, native xhigh/max effort, `thinking` omitted on "off"), native web-fetch support, display formatting, alias/tokenizer-override registry wiring, and the unchanged fallback seeding semantics with the target pinned to Opus 5. - Audited remaining `anthropic:claude-opus-5` fixtures: all use it as an explicit custom model string, which stays valid; default-model-derived fixtures (CLI `--help`, workspace creation, e2e display name) were updated. - 5.5 retarget (`605816945`): zero `5-1`/`5.1`/`Opus51` residue outside the regenerated skill file; the nine model/config/provider suites pass on the pinned Bun (754 tests, 0 failures); `make static-check` green. - Launch alignment (`c86761dae`): eleven model/config/provider/thinking/aiService suites pass on the pinned Bun (929 tests, 0 failures) plus the eight suites that reference the `OPUS` default (207 pass); `make static-check` green. CI on the final head is the authoritative gate. <details> <summary>Earlier 5.1-era validation (rebase + full unit suite classification)</summary> Rebase onto `main` (post-coder#3988-merge) re-verified: conflicts resolved additively (Mythos 5.1 promotion, gemini 3.8 fixture churn, `config.ts` → `config/index.ts` split) and `make static-check` green. The full unit suite was re-run locally on the pinned Bun in fail-closed chunks; every failure was classified as pre-existing: a host git `init.templateDir` quirk (passes with `GIT_TEMPLATE_DIR` set), Storybook snapshot-budget and BackupRepoCache failures that reproduce identically on clean `origin/main`, and suite-load flakes that pass in isolation. </details> ### Review record Six review rounds were used on the 5.1-era diff (round 1: Codex code review with one P2 fixed + security review; rounds 2–3: clean code + security pairs, the last on `e78cf41e7`). The independent advisor pass recommended **ready with tracked release follow-ups** (the launch checklist below), held as draft. The 5.5 retarget (mechanical rename + reprice) and the launch alignment (`anthropicRejectsDisabledThinking` match for Opus 5.5, removal of the chain-target migration) sit on top of that reviewed diff and did not open new review loops, so those six reviews and the advisor recommendation apply to `e78cf41e7`; the final head `c86761dae` is covered by the automatic code + security pair described in the note below. A second advisor consultation on launch day recommended dropping the target migration (scope reduction) rather than migrating users' recovery chains onto a classifier-bearing model; a third advised holding at the six-review cap, which the maintainer's explicit merge instruction superseded. > [!NOTE] > **Final-head review record.** Six Codex reviews (three code + security pairs on `26d3cc49`, `2259d651`, `e78cf41e`) were completed under the six-review budget. Marking the PR ready for review on Sept 22 (required for merge) triggered Codex's automatic final-head pair on `c86761dae` — code review and security review both **Completed** with no findings (👍, no threads) — bringing the total to **eight completed assessments**. The two extra assessments were accepted deliberately to satisfy the "required checks must cover the final commit" rule after the maintainer authorized delivery; no fix rounds resulted from them. ### Launch checklist (before undrafting) 1. ~~Confirm the official API id is `claude-opus-5-5`~~ ✅ verified Sept 22 (Claude API / Bedrock / Vertex / Foundry ids in the docs). 2. ~~Confirm pricing~~ ✅ $4/$20, 5-min cache write $5, cache read $0.20 verified; 1-hour cache write ($8) and fast mode ($8/$40) intentionally not modelled (same as Opus 5). 3. ~~Confirm envelope~~ ✅ 1M/128K, native xhigh + max verified; thinking-off **not** allowed and safeguard classifiers **present** — both handled in `c86761dae` (policy drops "off"; fallback target pinned to Opus 5, migration removed). 4. ~~Obtain clean Codex code **and** security reviews for the final commit~~ ✅ both completed clean on `c86761dae` (Sept 22, trigger: draft marked ready); `Codex Comments` + `Required` re-run on that same commit. ## Risks Low: additive registry/metadata changes, no config migration. The main user-visible effects are that the `opus` alias, `/opus`, and the default model route to Opus 5.5, and that users on the `opus` alias lose the thinking "off" level (persisted "off" clamps to `low`, matching the API's always-on thinking). Existing Opus 5 selections and all fallback chains (default or user-edited) are unaffected. Cost tracking uses the official $4/$20 and $0.20/M cache-read prices; 1-hour cache writes are under-counted at the 5-minute rate, as for every other Anthropic entry. --- _Generated with `xum` • Model: `coder:anthropic/claude-fable-5-1` • Thinking: `xhigh` • Cost: `$106.53`_ <!-- mux-attribution: model=coder:anthropic/claude-fable-5-1 thinking=xhigh costs=106.53 -->
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds first-class support for Claude Opus 5 (released July 24, 2026) and promotes it to Mux's default
opus/OPUSmodel in place of Opus 4.8.Background
Anthropic shipped Opus 5 as the successor to Opus 4.8 with the same pricing and envelope, verified against the official What's new in Claude Opus 5 and pricing pages:
claude-opus-5(dateless, same convention as Sonnet 5 / Fable 5).lowthroughmax, nativexhigh), thinking on by default, no beta header.Per the request-handling review: Opus 5 introduces one breaking API change vs 4.8,
thinking: {type: "disabled"}returns 400 when effort isxhighormax. Mux is unaffected because theoffthinking level maps to effortlow, and all other levels send adaptive thinking. The existinganthropicSupportsNativeXhigh()regex already matchesclaude-opus-5, so adaptive thinking,display: "summarized", and native xhigh/max effort apply automatically with no wire-format changes. New Opus 5 beta features (mid-conversation tool changes, server-side default fallbacks) and fast mode were deliberately not adopted.Implementation
knownModels.ts:OPUS→claude-opus-5(keepsopusalias, app default, and the Fable→Opus refusal fallback via the shared constant). Tokenizer override stays onanthropic/claude-opus-4.5for approximate counting since the newer 4.7+ tokenizer is not published upstream.models-extra.ts: newclaude-opus-5pricing/limits entry.models.ts:claude-opus-5added to the native-1M patterns (not the beta-header toggle).providerService.ts: mux-gateway first-time default model list now seeds Opus 5.claude-opus-5has no minor version to exercise dot-vs-dash normalization.Risks
Low. Opus 4.8 remains usable as a custom model string (
anthropic:claude-opus-4-8); its metadata entry is retained. The main user-visible effect is that theopusalias and default model route to Opus 5. Existing per-user fallback chains are not migrated (seed-once semantics preserved).Generated with
mux• Model:anthropic:claude-fable-5• Thinking:xhigh• Cost:$28.06