Claude Opus 5.5 is live: pin the model id, then check the bill

Anthropic published Claude Opus 5.5 on September 22, 2026 — the first model in the Claude 5.5 family. On the Claude Platform the model id is claude-opus-5-5. It is also available on AWS, Google Cloud, and Microsoft Azure. The release page says Opus 5.5 performs at the level of Claude Fable 5.1 on most work, costs 40% less to run than Opus 5 on typical workloads, and returns output more than 30% faster than Opus 5.

That is a product you can pin, not a routing rumor. If you read the Opus 5.2 routing speculation note, keep the stories apart: that post was an unconfirmed routing idea; this one is the official page at https://www.anthropic.com/claude-opus-5-5. Decide whether the id your app sends should change, price it from your token mix, and do not let four benchmark rows replace the tasks you already run.

What the page commits to, and what it does not

The page commits to a family position, platform id, cloud availability, rate card vs Opus 5, output-speed claim, fast-mode price, four benchmark rows, and a Fable 5.1 real-use caveat. Sonnet 5.5 and Haiku 5.5 are weeks ahead — not on this card.

It does not give retirement dates, a Fable 5.1 price, a prompt-injection success rate, millisecond latency, a promise every request is 40% cheaper, fast-mode cache prices or model id, or provider id strings for AWS, Google Cloud, or Microsoft Azure (only availability plus the Claude Platform id). Copy the id from the provider you call. A console label that reads Opus 5.5 is not the pin (Gemini arena label versus API pin).

The rate card against Opus 5

Prices are US dollars per 1M tokens, from the release page.

Three lines share a 20% discount; cache reads drop from $0.50 to $0.20. Anthropic also says Opus 5.5 costs 40% less to run than Opus 5 on typical workloads — that is a workload claim, not the list-price column (fresh input/output are 20% on the card). Your blended percent is a property of the mix.

When the bill falls 20%, 40%, or somewhere else

Call Opus 5 dollars from input, output, and cache writes U, and cache-read dollars C. Those three lines become 0.8U on Opus 5.5; cache reads become 0.4C. A 40% cut means 0.8U + 0.4C = 0.6(U + C), or U = C — cache-read spend at Opus 5 prices equals the other three lines. Larger cache-read share → cut above 40% (up to 60% pure cache-read). Smaller share → closer to the 20% floor when C is zero. Arithmetic on the published card, not a claim about Anthropic's unpublished typical mix.

Uncached: 5M input and 1M output, no cache. Opus 5 is $50; Opus 5.5 is $40 — 20%. Output can still be more than 30% faster; the invoice does not move by 40%.

The 40% point: 1M input, 1M output, 60M cache-read tokens. Opus 5 is $5 + $25 + $30 = $60 (U = C). Opus 5.5 is $4 + $20 + $12 = $36 — 40%.

Replay a representative window of production counts through the new card before you change the pin. Split counters into input, output, cache write, and cache read.

Output speed, and what fast mode costs

Output is more than 30% faster than Opus 5 — output tokens, not tool round-trips, retrieval, or the editor. If output is only part of wall clock, end-to-end gain is smaller; measure before repeating the figure.

Fast mode, in Claude Code and on the Claude Platform, is listed at up to 2.5× speed, priced at $8 input and $40 output per 1M tokens. The page does not name the 2.5× baseline, publish fast-mode cache prices, or give a distinct model id — enable only through the documented control, and do not bill those requests at the $4 / $20 card by accident.

$8 / $40 is double standard Opus 5.5 (1.6× Opus 5). The same uncached 5M + 1M mix costs $80 in fast mode versus $40 on standard Opus 5.5 — a premium for interactive wait time, not a migration discount. Prefer the $0.20 cache-read line for backfills, overnight jobs, and cache-heavy traffic.

Four benchmark rows, and the caveat beside them

On every row Opus 5.5 sits above both Fable 5.1 and Opus 5. Smallest gap versus Fable 5.1 here: FrontierCode at 4.1 points. GDPval-AA is Elo, not percent — do not average the four gaps.

Bound: Opus 5.5 performs at Fable 5.1's level on most work, and the gap versus Fable 5.1 is narrower in real use than the scores suggest. Scores can sit higher while day-to-day work stays close.

A higher public score does not mean your repository's failure modes moved (the honest codebase problem on Opus 4.8). Use these rows to decide a trial is worth running; use your issues, tests, and production traces to decide the pin stays.

Safety lines on the page

Opus 5.5 is stronger on automated behavioral audit than prior models, more resistant to prompt injection than Opus 5, and similar to Fable 5.1 on biology and cyber safeguards (Life Sciences Verification Program; Cyber Verification Program expanding). No attack success rate is published — do not remove allowlists, tool sandboxes, or human confirmation on destructive actions because a comparison improved.

When to change the id you send

Pinned to Opus 5, with cost or latency as the complaint, and you can re-run evals: move the Claude Platform pin to claude-opus-5-5 in one environment. Replay token counts before promising a finance number. Expect ~20% off fresh input, output, and cache writes, 60% off cache reads, and a whole-bill cut near 40% only if your mix is near U = C or matches the workload Anthropic called typical. Expect output more than 30% faster than Opus 5, then measure end to end. Keep the old pin for rollback.

Pinned to Fable 5.1 as the quality pin: Opus 5.5 is at that level on most work; scores sit higher; real use narrows the gap. No Fable price here, so cost is not the argument — run your tasks; switch one environment.

Person waiting in Claude Code or the Claude Platform: fast mode at $8 / $40 is a deliberate premium, not the new default. Wanted a smaller default: Sonnet 5.5 and Haiku 5.5 are weeks ahead with no prices or ids — do not block an Opus 5 cutover on a model you cannot call. Multi-cloud availability does not mean one id string is correct everywhere; confirm each provider string and pin ids, not marketing names.

If the rumor post changed your router, undo that

The routing-speculation post was not a changelog. If a branch sends traffic down an Opus 5.2 guess, a hidden alias, or a "latest Opus" label you do not control, retire it. Put claude-opus-5-5 only where you have decided this model is the one; leave Fable 5.1 on its own explicit id until your tasks agree.

Order of work: (1) one config value for the Claude Platform model id; (2) four token counters before cutover; (3) paper replay onto the Opus 5.5 card vs the 20% floor, 40% U = C case, and 60% cache-read ceiling; (4) canary scored on your tasks — Terminal-Bench 4.0 at 66.4% is not your task success; (5) previous id for rollback; fast mode if enabled meters at $8 / $40 as its own bucket; (6) provider ids from AWS, Google Cloud, and Microsoft Azure docs, then pinned.

What this release is not

Not a Sonnet 5.5 or Haiku 5.5 launch. Not evidence that prompt-injection controls are finished. Not a reason to treat biology or cyber safeguards as looser or tighter than Fable 5.1. Not permission to quote 40% savings unless your replay lands near 40% — if uncached, say 20% on input and output, and output more than 30% faster than Opus 5.

Primary page: https://www.anthropic.com/claude-opus-5-5. Quantitative claims above are stated there or arithmetic on its prices and scores. If a later doc revises a price or score, the page wins.

Sources

• Anthropic, Claude Opus 5.5 release (model id, availability, prices, speed, fast mode, benchmarks, safety): https://www.anthropic.com/claude-opus-5-5
• Related, not a source for this model card — routing speculation: https://otf-kit.dev/blog/claude-opus-5-2-routing-speculation
• Related, not a source for these scores — benchmark ≠ repository: https://otf-kit.dev/blog/opus-4-8-honest-codebase-problem
• Related, not a source for Claude pricing — public label ≠ API pin: https://otf-kit.dev/blog/gemini-arena-label-vs-api-pin