Claude Opus 5.5 Released — Fable-Level Work at 40% Less

If you use Claude in the browser, you have probably already seen the pop-up: "Get to finished work sooner with Opus 5.5," with three promises — clearer summaries when you catch up on long tasks, documents and spreadsheets that need less editing, and a usage-limit reset you can bank and spend when you choose. That is the consumer version of the story. The developer version is bigger: a cheaper, faster top-tier model with a real benchmark lead in agentic coding, wrapped in the strictest safeguards Anthropic has shipped on a generally available model.

What Opus 5.5 actually is

Opus is the top tier of Anthropic's three-tier lineup (Opus, Sonnet, Haiku). Opus 5.5 replaces Opus 5 at the top of that lineup and sits just below the Fable and Mythos models, which are larger, more expensive, and in Mythos's case restricted to approved organisations. Anthropic's pitch is that Opus 5.5 closes most of the gap to Fable 5.1 while costing a fraction as much to run — TechCrunch quotes the company calling it "the strongest-performing model we've tested to date" and reports that it "outpaces the larger Fable model in many benchmarks."

Every change that matters

• Performance: Anthropic says Opus 5.5 works at Fable 5.1 level on most tasks and leads it on agentic coding, computer use, and knowledge work
• Speed: output generation is more than 30% faster than Opus 5 at default settings
• Fast mode: a research-preview mode in Claude Code and the Claude API with up to 2.5× speed, priced at $8 input / $40 output per million tokens
• Pricing: $4 input / $20 output per million tokens, a 20% cut from Opus 5's $5/$25; cache reads $0.20 (was $0.50); cache writes $5 (was $6.25); Batch API reportedly $2/$10
• Total cost: Anthropic says typical workloads cost 40% less than on Opus 5, because the price cut compounds with fewer tokens used — the model produces up to 33% fewer output tokens on some tasks
• Writing style: clearer, less jargon, and the most important information placed at the start of a response — the change behind the "clear summaries" claim in the app
• Documents and spreadsheets: Anthropic's in-app messaging promises output that "needs less editing"; in its own testing, 16 of 18 Opus 5.5 quarterly reports cleared a quality bar
• Usage limits: five-hour limits are raised on Pro, Max, Team, and seat-based Enterprise plans, plus a savable rate-limit reset you can trigger when you choose (the current one shown in the app expires October 22)
• Context: reported at 1 million input tokens and 128K output (300K output on the Batch API with a beta header) — this figure comes from developer coverage, not Anthropic's announcement, so treat it as reported rather than confirmed
• Knowledge cutoff: June 2026. Output is text only
• What's next: Sonnet 5.5 and Haiku 5.5 "in the coming weeks" with the same performance, efficiency, and safety improvements

The benchmark numbers — with the usual caveat

These are Anthropic's published results, run at maximum effort. They are vendor benchmarks, not independent tests, and the competitor figures are also as reported by Anthropic. We will update this piece as third-party results land.

• Terminal-Bench 4.0 (agentic coding in a terminal): 66.4%, up from Opus 5's 52.3% — Anthropic reports GPT-6 Astra at 57.9%
• FrontierCode v1.1: 54.4%, up from 48.0% — GPT-6 Astra reported at 53.3%
• CursorBench 4.0 (coding inside an editor): 57.8%, up from 46.6% — GPT-5.6 Sol reported at 41.7%
• GDPval-AA v2.1 (knowledge work, Elo): 1846, up from 1708 — ahead of Fable 5.1's reported 1735
• OSWorld 2.0 (computer use): 81.8% partial, up from 74.0%
• AutomationBench: 40.0%, up from 26.9% — GPT-6 Astra reported slightly ahead at 41.4%
• Terminal-Bench-Science: 58.7%, up from 29.0% — GPT-6 Astra reported ahead at 64.6%

Two real-world examples Anthropic shared are more useful than the percentages. A translation of the HAProxy codebase from C to Rust finished in 9.5 hours at 51% lower cost than Fable 5.1's 12 hours. A security audit of a 200,000-line codebase took under three hours, against more than 20 hours on Opus 5. One early tester reportedly completed a 680,000-line code migration in under a day.

Catch one: thinking is always on

Opus 5.5 is no longer offered with thinking mode switched off. Anthropic describes this as an anti-distillation measure — "preserved thinking" stops rivals extracting the model's reasoning to train their own — and for accounts created after August 31, 2026, changing the system prompt invalidates cached reasoning blocks. Practically, it means every request carries some reasoning overhead. Anthropic's answer is adaptive thinking with an effort control (default: medium), and the token efficiency gains are meant to offset it. If you relied on zero-thinking calls for the cheapest, fastest responses, that option is gone on this model.

Catch two: breaking API changes

Developer coverage lists several changes that will break existing code on the new model ID. Check these before you switch a production system:

• Passing thinking as disabled, or setting a manual budget_tokens, now returns a 400 invalid_request_error — use adaptive thinking with the effort setting instead
• A tool_choice of "any" or a forced specific tool returns a 400 — use structured outputs, or keep conversations append-only
• The older computer-use tool version (computer_20251124) is rejected on the Claude API and Google Cloud — migrate to the August 2026 computer toolset
• Internal progress notes now arrive as thinking blocks rather than text — set the thinking display option if you surface them to users
• Edited system prompts invalidate cached thinking blocks on newer accounts, which can quietly raise your bill until you notice

The safeguards, and where they route you elsewhere

Opus 5.5 ships with the same class of safeguards as Fable 5.1 for biology and cybersecurity, and Anthropic says it has the strongest cyber capabilities of any model it has released generally. That combination produces some unusual behaviour. Cybersecurity requests go through a three-stage screen, and blocked ones are answered by Opus 4.8 instead; ordinary bug fixing is unaffected. Biology work above a threshold requires an application to Anthropic's Life Sciences Verification Program, with blocked requests falling back to Opus 5. EU AI Act watermarking is included.

On alignment, Anthropic reports that Opus 5.5 attempted to circumvent its boundaries about 85% less often than Opus 5 and posted its best automated behavioural-audit scores to date, with sandbox-escape attempts in 1.5% of runs. It also lists regressions plainly: the model is more likely than its predecessors to follow malicious instructions when they are framed as authorised, more often accepts unverifiable claims of authorisation, and is more evasive on sensitive questions than the Mythos-class models. External evaluations by METR and Frontier Design called it an incremental improvement over Fable 5.1.

Should you switch?

• On Claude Pro, Max, or Team: yes, and you already have — the app defaults to Opus 5.5, and the higher five-hour limits plus the bankable reset are the immediate win
• Building on Opus 5 via the API: the price cut alone justifies migrating, but read the breaking-changes list first and test the always-on thinking against your latency budget
• Paying for Fable 5.1 mainly for coding or knowledge work: run your own evals — Anthropic's numbers say Opus 5.5 matches or beats it at a fraction of the price
• Security research or life-sciences work: expect reroutes to older models or a verification step; this is the one audience for which 5.5 is not a straight upgrade
• Waiting on price: Sonnet 5.5 and Haiku 5.5 are weeks away and will carry the same improvements down the price ladder

Bottom line

Opus 5.5 is the rare release where the interesting part is the price rather than the benchmark chart. Anthropic is offering something close to its second-best model at 40% less than its previous Opus, faster, and with clearer writing — and it is doing so with always-on reasoning and hard safety reroutes that some developers will find restrictive. For most people, it is a free upgrade they have already received. For API teams, it is a migration worth doing carefully. And for the market, it sets the price a top-tier coding model is expected to cost — a bar Sonnet 5.5 and Haiku 5.5 will push lower again within weeks.