Claude Opus 5.5: Pricing, Benchmarks and Breaking Changes

Anthropic released Claude Opus 5.5 on September 22, 2026. It is the first model in the Claude 5.5 family, priced at $4 per million input tokens and $20 per million output tokens (Opus 5: $5 and $25), with cache reads cut from $0.50 to $0.20. Anthropic says it performs “at the level of Claude Fable 5.1 on most work” and costs 40% less to run than Opus 5 on typical workloads.

Both claims hold up on Anthropic’s own published numbers, but they are measured at different settings, and the difference matters for anyone planning a migration. The 40% saving is measured at each model’s default effort, which is medium on Opus 5.5 and high on Opus 5. Most rows in the headline benchmark table are measured at max effort. And four API changes will return errors on code that runs cleanly on Opus 5 today.

Every figure in this post is Anthropic’s own, or a partner’s as quoted by Anthropic. None has been independently replicated at the time of writing.

Two other launch details affect planning. According to the announcement, Claude Sonnet 5.5 and Claude Haiku 5.5 will follow “in the coming weeks.” Anthropic is also raising five-hour usage limits on the Pro, Max and Team plans, and subscribers get a one-off rate-limit reset that they can save and use whenever they choose.

Anthropic’s ClaudeDevs account put numbers on the subscription change the same day. In Claude Code, five-hour session limits rise 20% from September 22, and because Opus 5.5 is priced lower, it goes about 25% further within those limits. The same post describes Opus 5.5 as about 30% faster and about 40% cheaper per task than Opus 5.

The release also has a policy context. It is Anthropic’s first model since its CEO argued the frontier should be paced, and METR and Frontier Design evaluated it before launch.

The cache-read line is the structural change. According to Anthropic’s pricing page, a cache hit costs 10% of the input price on most Claude models. On Opus 5.5 it costs 5%, and on Fable 5.1 it costs 2.5%. The announcement says cache reads “make up the majority of agentic and coding work costs,” so the 60% cut lands on the biggest part of an agent’s bill. The minimum cacheable prompt stays at 512 tokens, the same as Opus 5 and Fable 5.1.

Fast mode (research preview, Claude API only) runs Opus 5.5 at up to 2.5× the output speed for 2× the standard price. It is not available on Bedrock, Claude Platform on AWS, Google Cloud or Microsoft Foundry. US-only inference (inference_geo: "us") adds a 1.1× multiplier to every rate above, fast mode included.

A lower per-token price does not cut every bill by the same percentage, because each token type fell by a different amount. The table below is an illustrative agent task. The token counts are invented for the example; the rates are Anthropic’s published prices.

With identical token counts, the price change alone saves about 29% on this cache-heavy task. A workload with no caching saves exactly 20%. The rest of Anthropic’s 40% has to come from Opus 5.5 using fewer tokens per task, which is its stated second source of savings (“it costs less per token than Opus 5 and uses fewer tokens per task”). That part depends on your workload and your effort setting, so only your own measurements can confirm it.

This is the most useful thing to understand before switching models.

The cost claim compares default settings. Anthropic’s wording is “at default settings it will cost 40% less than Opus 5 on typical workloads.” Opus 5.5 defaults to medium; Opus 5 defaults to high. The prompting guide says that in Anthropic’s testing, “Claude Opus 5.5 at medium matches or exceeds Claude Opus 5 at high on coding and knowledge-work evaluations.” So the comparison is between Opus 5.5 at a lower effort level and Opus 5 at a higher one, delivering the same or better results.

The headline table uses max effort. Its footnote reads: “Unless otherwise noted, all Claude Opus 5.5 results use adaptive thinking at max effort.” The only exception is Terminal-Bench 4.0, reported at xhigh because that was Opus 5.5’s best score on it. So the 1846 GDPval-AA score and the 40% lower bill come from different settings.

Keeping your old effort setting raises the cost. Anthropic’s What’s new page says that “at the same effort setting the model tends to think more per turn than Claude Opus 5, most of all at xhigh and max.” A team that pinned effort: "high" on Opus 5 and only changes the model ID will get a model that thinks more than it used to at that setting. That team gets the per-token cut and may lose some or all of the token savings.

What to do: remove any effort value inherited from Opus 5, start at medium, and compare two or three levels on your own traffic. Anthropic also recommends leaving max_tokens large enough for thinking (it suggests 128,000, the model’s maximum, for long agentic turns) and using xhigh and max only where you have measured a quality gain.

Every figure below comes from Anthropic’s comparison table. The last column shows the highest score among the other four models Anthropic compared: Fable 5.1, Opus 5, GPT-6 Astra and GPT-5.6 Sol. Where GPT-6 Astra is shown, Fable 5.1 scored 55.8% on Terminal-Bench 4.0, 50.3% on FrontierCode, 31.4% on AutomationBench and 52.6% on Terminal-Bench-Science.

Five caveats come from Anthropic’s own footnotes, and each one affects how the table should be read:

1. Margins sit close to the noise. Anthropic gives Terminal-Bench 4.0 a standard error of ±2.6 points for Opus 5.5, and ±3.5–5 points per model on Terminal-Bench-Science. A 1.1-point lead on FrontierCode over GPT-6 Astra is not a meaningful difference.
2. OpenAI’s numbers come from OpenAI. The GPT-6 Astra and GPT-5.6 Sol figures on Terminal-Bench are “as reported by OpenAI,” so they were not run in Anthropic’s harness.
3. Safeguards affected the scores. Opus 5.5 was evaluated with its production safeguards on. When they intervened, cybersecurity tasks were completed by Claude Opus 4.8, and biology and frontier-LLM-development tasks by Opus 5. Anthropic says this “likely reduces” Opus 5.5’s scores.
4. AutomationBench is Zapier’s run. Zapier ran it without fallback models, so every safeguard intervention counted as a failure. The Opus 5.5 result comes from Zapier’s early-access evaluation; the other models’ results come from its public leaderboard.
5. Anthropic discounts its own table. “At these levels of capability we’ve found that benchmark margins have become a less reliable guide to real-world differences.”

In short, Opus 5.5 leads on coding, knowledge work and computer use by margins that range from clear (Terminal-Bench, GDPval-AA) to within noise (FrontierCode, OSWorld, Chartography). It trails GPT-6 Astra on two agentic benchmarks.

Anthropic also published score-against-cost curves for each effort level on four benchmarks. Their shapes show where extra effort helps.

Anthropic’s text states the cost comparisons directly. At default effort, Opus 5.5 beats GPT-6 Astra on FrontierCode at “roughly 20% of the cost per task” and matches Astra on Terminal-Bench 4.0 for “about 40% of the cost.” At default effort on GDPval-AA, it beats Astra at max effort “for about a fifth of the cost per task.” On CursorBench, it beats GPT-5.6 Sol by 11 points “for about a third of the cost.”

The practical reading is that coding agents may be best served at medium or high, with a re-run at higher effort only when a task fails. Research, analysis and document work gain more from higher effort. Anthropic’s cost-optimisation guide reached the same conclusion with earlier models: where results can be checked, the cheapest option is often to run at a low setting and re-run only the failures at a higher one. A partner result points the same way. In the announcement, Deloitte’s CIO says that “even at its lowest effort setting,” Opus 5.5 caught 72% of known bugs in their code reviews, against Opus 5’s 56% at high effort.

The What’s new page lists four changes that break code running on Opus 5. The first three also apply to Fable 5.1, so teams that already moved to Fable have done most of this work.

The model-switching rule matters for any router. Opus 5.5 can read thinking blocks from Opus 5 and earlier Opus, Sonnet and Haiku models, but not from Fable or Mythos models. On the Claude API, only Fable 5.1 and Mythos 5.1 can read Opus 5.5’s blocks. A conversation that switches between models outside those pairs continues without the earlier reasoning. The request still succeeds, and the dropped blocks are not billed.

Three behaviour changes also need attention in agent loops:

• Unattended loops can stop early. On long tasks, Opus 5.5 sometimes ends a turn with a text progress report rather than a tool call. A loop that reads end_turn as “done” will stop there. Anthropic recommends treating a text-only turn as a report, keeping a checklist of the task’s parts, and sending a short continuation message when items are still open, with a limit of two or three automatic continuations.
• Changing effort mid-session invalidates the prompt cache. To change effort for a single turn, use the per-message effort beta instead.
• More refusal categories. A biology classifier now runs alongside the cybersecurity one, and prompts that ask the model to reproduce its internal reasoning in its reply can be declined under a new reasoning_extraction category. Handle stop_reason: "refusal" and set up a fallback; the server-side fallbacks: "default" option is in beta.

Our Fable 5.1 migration post covers the first three changes in more depth, and the deprecation calendar tracks when older models stop accepting traffic.

Opus 5.5 is the first Opus model to launch with safeguards in the same class as Fable 5.1’s, covering cybersecurity, biology and distillation. When a safeguard triggers, the request falls back to another model without any visible sign to the user.

On alignment, Anthropic reports that Opus 5.5 scored better than any recent Claude model “on nearly every measure of misaligned behavior” across an automated audit of nearly 2,000 scenarios. In a new containment test, it tried to get around boundaries about 85% less often than Opus 5 or Claude Mythos 5.1, and Anthropic says every attempt was low severity and reported by the model itself.

The announcement names the limit of this evidence plainly: “We see signs that Opus 5.5 often suspects it is being evaluated.” For teams running agents unattended, the improved scores still don’t replace your own guardrails.

The announcement includes several case studies. All are Anthropic’s or its early testers’ results, and none has been reproduced independently:

• A tester completed a 680,000-line code migration in less than a day.
• Another audited and fixed a 200,000-line codebase in under three hours. Opus 5 took more than 20 hours and used 2.5× as many tokens.
• In an internal test, Opus 5.5 and Fable 5.1 each rewrote the HAProxy load balancer from C into Rust. Both passed nearly all of HAProxy’s regression tests. Opus 5.5 finished in 9.5 hours against 12, at 51% lower cost.
• In a research-report test where every figure and quote was checked against sources, 16 of 18 Opus 5.5 reports passed. Neither Fable 5.1 nor Opus 5 passed in any attempt.
• Asked to cut page load times across a web app, Opus 5.5 succeeded 39 of 40 times. Opus 5’s improvements were smaller and also changed how the app behaved.

The communication change is harder to measure but easy to test. Anthropic says Opus 5.5 “puts the most important information up front, is less likely to use jargon or idiosyncratic phrases, and follows the writing rules you give it.” Its side-by-side example shows the difference. Asked to explain a billing bug, Opus 5 opens with the commit and the code. Opus 5.5 opens with the result: the free-tier change explains $1.50 of the drop, and the other $9.92 comes from a refactor that stopped billing the last day of each month. Teams that give their agents a style guide can check this claim in an afternoon on their own prompts.