Introducing Claude Opus 5.5 (Anthropic)
Anthropic's launch post for Claude Opus 5.5, dated 22 September 2026. It is the first model of a "Claude 5.5 family"; Sonnet 5.5 and Haiku 5.5 are promised "in the coming weeks". Arrived by Telegram on 2026-09-23, a day after opus-and-sol-and-luna-willison had already given this spoke the price.
T1 for what Anthropic decided (price, availability, safeguards, API behaviour). Every benchmark and every customer result here is first-party or vendor-selected, and should be read as a claim. Two of the benchmark rows were not run by Anthropic (AutomationBench by Zapier, the GPT figures as reported by OpenAI), and the post says so in its footnotes.
The headline claims
- Near Fable 5.1 at lower cost. "It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5."
- The 40% is two effects added together: a cheaper token and fewer tokens per task. The per-token cut alone is 20%.
- Output is "more than 30% faster" than Opus 5.
- Consumer plans: five-hour usage limits go up on Pro, Max, Team and seat-based Enterprise, and subscribers get a rate-limit reset they can save and use later.
Price (per million tokens)
| Opus 5.5 | Opus 5 | Change | |
|---|---|---|---|
| Input | $4 | $5 | −20% |
| Output | $20 | $25 | −20% |
| Cache reads | $0.20 | $0.50 | −60% |
| Cache writes | $5 | $6.25 | −20% |
| Fast mode (input / output) | $8 / $40 | — | "up to 2.5x speed" |
The post calls cache reads "the majority of agentic and coding work costs", which is why it puts the biggest cut there. This agrees with the Anthropic pricing page this spoke read the same day (anthropic-model-docs).
Benchmarks (Anthropic's table)
Opus 5.5 at adaptive thinking, max effort, unless noted.
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 (xhigh) | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| FrontierCode v1.1 (Main) | 54.4% | 50.3% | 48.0% | 53.3% | 47.5% |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% | — | 41.7% |
| GDPval-AA v2.1 (Elo) | 1846 | 1735 | 1708 | 1542 | 1588 |
| AutomationBench (Zapier) | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
| Humanity's Last Exam (tools) | 67.7% | 65.6% | 63.6% | 57.2% | — |
| Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% | 22.4% |
| OSWorld 2.0 (partial) | 81.8% | 80.7% | 74.0% | — | — |
| Chartography (tools) | 89.0% | 88.4% | 83.4% | — | — |
What the footnotes add:
- Opus 5.5 does not win everything. GPT-6 Astra leads on AutomationBench and on Terminal-Bench-Science.
- Error bars are wide. Terminal-Bench 4.0 is ±2.6 points for Opus 5.5; Terminal-Bench-Science is ±3.5–5 points per model. Anthropic re-ran Opus 5 on both public leaderboards and landed within noise of the published figure (52.3% vs 51.8%; 29.0% vs 30.0%).
- The safeguards took part in the scoring. Opus 5.5 ran with production safeguards on. When they stepped in, cyber tasks were answered by Opus 4.8 and biology and "frontier LLM development" tasks by Opus 5. Zapier ran AutomationBench without fallback, so every intervention counted as a failure there.
- 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 our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest."
Cost per task, at default effort
The post's charts plot score against cost per task for five effort levels (low, medium, high, xhigh, max). Medium is the default. The claims it draws from them:
- FrontierCode: 54.6% at medium, above every other model's best, including Astra's 53.3%, at "about a fifth of the cost per task". Medium scores slightly higher here than the max-effort 54.4% in the table.
- Terminal-Bench 4.0: medium beats Opus 5 at max "for about a fifth of the cost" and matches Astra at about 40% of the cost.
- CursorBench: 52.5% at medium, 11 points over GPT-5.6 Sol's best at about a third of the cost.
- GDPval-AA: medium beats Astra at max effort for about a fifth of the cost per task.
- WANDR (Perplexity's data-collection benchmark): Opus 5.5 beats Fable 5.1 and Opus 5 at lower cost. Anthropic ran it with its own tool setup and a 980k-token task budget, and says the scores cannot be compared with Perplexity's published ones.
Anthropic's own task tests
- A HAProxy rewrite from C to Rust: both Opus 5.5 and Fable 5.1 passed nearly all of HAProxy's regression tests; Opus 5.5 took 9.5 hours against 12, at 51% lower cost.
- A quarterly-performance report written from a copy of the web, with every figure and quote checked by an automated grader: 16 of 18 Opus 5.5 reports passed; no Fable 5.1 or Opus 5 report did.
- An Excel merger model plus a slide deck: same conclusion from both models; Opus 5.5 took 63 minutes against Opus 5's 93, at 50% lower cost, and Opus 5's version had "minor errors".
- A web-app load-time task: 39 of 40 successes; Opus 5 made smaller gains and changed the app's behaviour.
Customer quotes worth keeping
Twenty-one early-access companies are quoted. The ones that carry a number:
- Deloitte: at low effort, Opus 5.5 caught 72% of known bugs in code review, against Opus 5's 56% at high effort.
- Rogo: at low effort it beat Opus 5 at high effort on their BigFinance Bench with about 60% fewer output tokens.
- Optiver: Opus 5's quality in about half the turns, time and output tokens; 40–50% lower cost.
- Kiro: about 40% fewer calls and half the tokens on a public command-line benchmark.
- Quantium: a task that took 38 prompts over four days took 11 prompts over three hours.
- Hebbia: on finance workflows graded against expert rubrics, 86.6% coverage against Opus 5's 60.3%.
- Factory: the first model where they defaulted to medium effort; it matched Opus 5 at high effort with 20–25% fewer output tokens.
Several quotes (Ramp, Box, Clio) praise shorter, clearer writing. The post says this answers "common feedback we heard about Opus 5".
Safety and API changes
- External testing before release by METR and Frontier Design.
- Alignment audit (about 2,000 scenarios): best of any recent Claude model on nearly every measure of misaligned behaviour. A new containment test found it tried to cross boundaries about 85% less often than Opus 5 or Claude Mythos 5.1, and every attempt was low severity and self-reported. The post also says Opus 5.5 "often suspects it is being evaluated", which limits what the audit can show.
- Fallback safeguards, a first for an Opus model. Like Fable 5.1, it hands some work to another model, which the post calls falling back "transparently": most cybersecurity tasks go to Opus 4.8. Biology safeguards match Fable 5.1's, with a new Life Sciences Verification Program for vetted labs and companies. The Cyber Verification Program will add Opus 5.5 with three access tiers.
- Preserved thinking (the anti-distillation measure from Fable 5.1): API users cannot edit Claude's earlier context to pull out its reasoning. It applies to accounts created on or after 2026-08-31.
- Thinking cannot be switched off on Opus 5.5.
- EU AI Act watermarking, as on Fable 5.1; zero data retention available.
Availability
All platforms at launch, including AWS, Google Cloud and Microsoft Azure. API id claude-opus-5-5.
Fast mode is on the Claude Platform and in Claude Code.
What this source does not settle
- Context window and output limit are not stated. The 128,000-token output limit on claude-opus-5-5 comes from Willison.
- No benchmark here is independent. Anthropic chose the benchmarks, the effort settings and the harness. It is still a vendor document, however careful the footnotes are. (Checked 2026-10-02: outside this page, Artificial Analysis ran Opus 5.5 on launch day and measured Terminal-Bench 59.6% against Anthropic's 66.4%, and Humanity's Last Exam 61.4% against 67.7%, per heise and Vellum's write-ups. Not ingested.)
Related
claude-opus-5-5 · claude-opus-5 · claude-fable-5-1 · astra · gpt-6 · opus-and-sol-and-luna-willison · llm-benchmarks · coderabbit-fable-5-1-review · claude-mythos-5