Anthropic released Claude Opus 5.5 on September 22, 2026, its first model since CEO Dario Amodei published "We Must Pace the Frontier," an essay from about a week earlier arguing that safety practices need to keep pace with capability jumps. The model was evaluated before launch by outside groups including METR and Frontier Design.
The pitch
Opus 5.5 performs roughly at the level of Anthropic's top public model, Claude Fable 5.1, on most tasks, while costing 40% less to run than the outgoing Opus 5. Anthropic is candid that at this capability tier, benchmark gaps are becoming less reliable as a stand-in for real-world differences.
Benchmarks
Verified against Anthropic's own comparison table.
Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
FrontierCode v1.1 | 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 |
Humanity's Last Exam | 67.7% | 65.6% | 63.6% | 57.2% | - |
AutomationBench | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% | 22.4% |
OSWorld 2.0 (computer use) | 81.8% | 80.7% | 74.0% | - | - |
Chartography (chart reading) | 89.0% | 88.4% | 83.4% | - | - |
These numbers match Anthropic's published figures exactly, though OSWorld 2.0 and Chartography are scored under partial-credit and tool-assisted conditions per Anthropic's own footnotes, and GPT-6 Astra and GPT-5.6 Sol are absent from both rows, so those two comparisons are Claude-only rather than a full field test.
The two bolded losses to GPT-6 Astra have different causes. On Terminal-Bench-Science 0.1, Anthropic's general safeguard footnote applies: when Opus 5.5's safety systems intervened during testing, biology and frontier-AI-research tasks were completed by the older Opus 5 instead, which likely dragged the score down.
On AutomationBench, safeguards are still the cause, just structured differently. Those results were run by Zapier without fallback models, so any safeguard intervention counted as an automatic failure rather than being completed by a substitute model, as on Terminal-Bench-Science, producing a lower score than Opus 5.5 would achieve in practice.
Coding: the headline act
One tester completed a 680,000-line code migration in under a day. Asked to cut load times across a web app, Opus 5.5 succeeded 39 of 40 times without breaking behavior, where Opus 5 made smaller improvements that also altered the app's behavior. A 200,000-line codebase audit took Opus 5.5 under three hours versus Opus 5's 20-plus hours, at roughly 40% of the token cost.
Translating HAProxy from C to Rust, Opus 5.5 passed nearly all regression tests in 9.5 hours against Fable 5.1's 12, at about half the cost. Against GPT-6 Astra, it matches Terminal-Bench 4.0 for around 40% of the price and beats FrontierCode for about 20% of the price. In a separate test where several Claude models built a game from a single prompt, Opus 5.5's build scored highest on graphics and polish.
Customers quoted in the release, including GitHub, Box, and Optiver, independently described similar drops in tokens per task. Stripe is also quoted in the release, though on session efficiency completing a 40-pull-request rebase rather than on token counts.
Security. Opus 5.5 matches or beats Opus 5 against prompt-injection attacks across coding, tool use, computer use, and web browsing, and ties Fable 5.1 for the lowest prompt-injection success rate in third-party testing by the AI security firm Gray Swan. For long, unattended coding sessions, Anthropic also pairs the model with a pre-action classifier, an auditable open-source sandbox, and code review meant to catch vulnerabilities before they merge.
Not just code
In a blind earnings-report test where any invented figure counted as an automatic fail, 16 of 18 Opus 5.5 attempts passed; neither Fable 5.1 nor Opus 5 passed on any attempt. Walleye Capital, an early enterprise tester, said Opus 5.5 caught an error in the firm's own evaluation instructions that no prior model had noticed, a detail confirmed directly in Anthropic's writeup.
In a separate merger-analysis test, Opus 5.5 and Opus 5 each built an Excel financial model and turned it into an executive presentation. Both reached the same conclusion, but Opus 5.5 finished in 63 minutes versus 93 for Opus 5, at half the cost and with a cleaner model.
Two more data points came from named enterprise testers. At its lowest effort setting, Deloitte says Opus 5.5 caught 72% of known bugs in code review, against 56% for Opus 5 at high effort, and Hebbia's finance-workflow rubric put Opus 5.5 at 86.6% coverage versus 60.3% for Opus 5, alongside the model's best-ever citation recall.
Writing style overhaul
Anthropic says Opus 5.5 leads with the important information first and uses less jargon, addressing the most common complaint about Opus 5. Box reports roughly a two-thirds cut in token usage on content-heavy workloads as a result.
Safety
On Anthropic's roughly 2,000-scenario behavioral audit, Opus 5.5 scored better than any recent Claude model on nearly every measure of misaligned behavior and is the company's strongest model to date on most measures of honesty. In a new evaluation built to test how often a model tries to cross containment boundaries, Opus 5.5 attempted this about 85% less often than Opus 5 or Claude Mythos 5.1, and every attempt it did make was low severity and self-reported.
Because Opus 5.5 is now comparable to Mythos 5.1 in biology and cybersecurity, it ships under safeguards similar to Fable 5.1's: most flagged cybersecurity tasks are re-routed to the older Opus 4.8, and flagged biology tasks go to Opus 5, with full biology capability available to vetted organizations through Anthropic's Life Sciences Verification Program.
Anthropic says this reflects real capability rather than caution alone: in biology, Opus 5.5 improved on a long-horizon molecular-design evaluation run with Dyno Therapeutics, and outside red-teamers rated its scientific novelty on par with the best model they had tested.
On the cybersecurity side, Anthropic is expanding its Cyber Verification Program to three tiers of increasingly permissive access for vetted practitioners, up to and including the restricted Mythos models. Anthropic also reports signs that Opus 5.5 often suspects it is being evaluated, which complicates its ability to predict how the model will behave across the real-world settings it gets deployed into. Evaluation awareness is a known open problem across the AI industry, not unique to Anthropic.
Opus 5.5 also launches with preserved thinking, the anti-distillation safeguard introduced with Fable 5.1, which blocks API users from editing Claude's prior context to extract its reasoning. It applies to API accounts created on or after August 31, 2026, and extended thinking can no longer be fully disabled.
Pricing & availability
Opus 5.5 | Opus 5 | |
|---|---|---|
Input (per 1M) | $4 | $5 |
Output (per 1M) | $20 | $25 |
Cache reads (per 1M) | $0.20 | $0.50 |
Cache writes (per 1M) | $5 | $6.25 |
At standard settings, Opus 5.5 also generates output more than 30% faster than Opus 5. A separate fast mode in Claude Code and the Claude Platform trades some of that headroom for up to 2.5x the speed, priced at $8 per million input tokens and $40 per million output tokens.
Live now on the Claude Platform (claude-opus-5-5), AWS, Google Cloud, and Microsoft Azure. Pro/Max/Team subscribers and seat-based Enterprise plans get bumped five-hour limits plus a bankable rate-limit reset. Sonnet 5.5 and Haiku 5.5 are coming in the following weeks.
A note on accuracy
The "smaller improvements that also altered the app's behavior" line about Opus 5 comes straight from Anthropic's own comparison, not an independent claim, since Anthropic is grading its own predecessor here. Worth keeping that context in mind when reading vendor-published benchmarks.
The same goes for every customer statistic in this piece, Deloitte, Hebbia, GitHub, and the rest: those were selected and published by Anthropic for the launch, not gathered independently, so treat them as testimonials rather than a representative sample.
Comments (0)
Join the discussion by logging into your account.