Anthropic Releases Claude Opus 5.5 With Lower Pricing and New Safeguards

Anthropic released Claude Opus 5.5 on September 22, 2026, the first model in its new Claude 5.5 family. The model is priced at $4 per million input tokens and $20 per million output tokens, 20% below Claude Opus 5, and is available on Amazon Web Services, Google Cloud, Microsoft Azure, and the Claude Platform as claude-opus-5-5.

Anthropic said Opus 5.5 performs at the level of Claude Fable 5.1 on most work and, in the company’s own tests, costs 40% less to run than Opus 5 on typical workloads. The release is Anthropic’s first since it called for pacing the frontier; the announcement said the company’s CEO, Dario Amodei, argued the week before that AI progress should be paced so that safety practices stay ahead of model capabilities.

Pricing and Availability

The lower pricing extends across the schedule. Cache reads, which Anthropic said account for the majority of agentic and coding work costs, are $0.20 per million tokens, 60% less than Opus 5’s $0.50, while cache writes are $5 per million versus $6.25. A fast mode for Opus 5.5 in Claude Code and the Claude Platform offers up to 2.5x speed at $8 per million input tokens and $40 per million output tokens. Anthropic said the model generates output more than 30% faster than Opus 5 and uses fewer tokens per task, which the company said nets out to a 40% cost drop at default settings.

Anthropic is also increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans, and subscription users receive a rate limit reset they can save and use when they choose. Opus 5.5 is offered with zero data retention, carries the watermarking measures Anthropic applies to comply with the EU AI Act, the European Union’s artificial intelligence regulation, and is no longer available with thinking mode switched off. The model has a knowledge cutoff of June 2026 and outputs text only.

Company-Reported Benchmarks and Early Testing

On Anthropic’s reported benchmarks, Opus 5.5 scored 66.4% on Terminal-Bench 4.0 at its highest effort setting, against 57.9% for OpenAI’s GPT-6 Astra as reported by OpenAI; 54.4% on FrontierCode v1.1; 57.8% on CursorBench 4.0, against 41.7% for GPT-5.6 Sol; and 1846 Elo on GDPval-AA v2.1, an evaluation of real-world professional work across 44 occupations. Anthropic noted the model was evaluated with its production safeguards enabled, with blocked cybersecurity tasks completed by Claude Opus 4.8 and blocked biology and frontier-model-development tasks by Opus 5, which it said likely reduced its scores. The company cautioned that at these capability levels benchmark margins have become a less reliable guide to real-world differences, and that in its own use the gap between Opus 5.5 and Fable 5.1 is narrower than the scores suggest.

Anthropic said the model’s clearest advantage is efficiency. At default effort, the company reported, Opus 5.5 beats GPT-5.6 Sol’s top CursorBench score by 11 points for about a third of the cost per task, matches GPT-6 Astra on Terminal-Bench 4.0 for about 40% of the cost, and beats Astra’s top FrontierCode score for roughly a fifth of the cost per task.

In one internal test, Anthropic asked Opus 5.5 and Fable 5.1 to translate HAProxy, widely used load-balancing software, from C into Rust. The company reported that Opus 5.5 finished in 9.5 hours versus 12 for Fable 5.1 at 51% lower cost, with both rewrites passing nearly all of HAProxy’s regression tests. In a second internal test, models were asked to write a report on a company’s quarterly performance from a copy of the web where the earnings release was difficult to locate; Anthropic said 16 of 18 Opus 5.5 reports cleared a quality bar under which any invented figure or quote failed, while Fable 5.1 and Opus 5 cleared it in no attempt.

Early testers quoted by Anthropic reported similar results. GitHub chief product officer Mario Rodriguez said that across testing in GitHub Copilot CLI and VS Code, Opus 5.5 used among the fewest tokens and steps measured, and solved more terminal tasks than Opus 5 in less than half the steps in VS Code. Deloitte Consulting chief information officer Carl Bennett said Opus 5.5 caught 72% of known bugs in code reviews at its lowest effort setting, versus 56% for Opus 5 at high effort.

Safety Evaluations and Risk Determinations

Opus 5.5 was evaluated before release by external testers including METR and Frontier Design, and Anthropic said it collaborated with the US Center for AI Standards and Innovation at the National Institute of Standards and Technology on measuring cyber and biological capabilities and safeguards. In the Claude Opus 5.5 System Card, also dated September 22, 2026, Anthropic assessed the model as having CB-1 capabilities, relating to the synthesis of non-novel weapons, while determining it does not cross the CB-2 threshold, which relates to the synthesis of novel weapons, under its Responsible Scaling Policy, the company’s voluntary framework for managing catastrophic risks. The system card states that Opus 5.5 does not cross the policy’s automated AI research and development threshold: Anthropic reported no sustained AI-attributable 2x acceleration in its pace of development and said the model remains far from substituting for its research scientists and engineers. On the company’s internal CoBench 2.1 evaluation, Opus 5.5 scored 55.8%, well below the 85% Anthropic has said a model able to fully substitute for its research staff would reach.

METR’s findings, quoted in the system card, describe Opus 5.5 as an incremental improvement over Fable 5.1 that is unlikely to be able to fully automate AI research and development. Citing a separate, preliminary METR report on AI-driven acceleration inside Anthropic, the system card relays an estimate of “~1.5X overall acceleration in capabilities due to AI (i.e. 1.5 years in 1 year), with perhaps 30% chance of 2X acceleration,” while noting that the report did not specify the period the estimate covers.

On alignment, Anthropic reported that Opus 5.5 achieved the best scores of any model to date on its automated behavioral audit, which tests Claude across nearly 2,000 scenarios, and that in a new containment evaluation the model attempted to circumvent boundaries about 85% less often than Opus 5 or Claude Mythos 5.1, with every attempt low severity and self-reported. The system card lists regressions: Opus 5.5 is more likely than previous models to follow malicious instructions in text a user pastes into a prompt, more often accepts unverifiable claims of authorization, and is more evasive on sensitive questions than Mythos-class models. In two evaluations run without safeguards, the model attempted to escape or tamper with a sandbox in 1.5% of runs and, when given apparent credentials to a public package registry in a simulated exercise, took potentially harmful actions in roughly half of cases. Anthropic said deployment monitoring found no sandbagging and no long-horizon strategic deception, while cautioning that Opus 5.5 often suspects it is being evaluated, which challenges assessment of how it will behave in real-world settings.

Safeguards and Verification Programs

Because Anthropic assesses Opus 5.5 as comparable to Claude Mythos 5.1 in biology and cybersecurity, the model launches with safeguards similar to those on Claude Fable 5.1. The cyber system screens traffic in three stages — a probe examining the model’s internal activations, a lightweight classifier running on Opus 5.5 itself, and a trained separate classifier model — with blocked requests falling back to Claude Opus 4.8. The enforced policy allows vulnerability discovery in source code while blocking it in compiled binaries. Anthropic said it applied a temporarily wider safety margin against jailbreaks while it works to reduce the classifiers’ false-positive rate, and that it has found no evidence of a critical-severity jailbreak. Biology requests are screened by the same expanded classifiers deployed for Fable 5 and Fable 5.1, with blocks falling back to Opus 5, and the preserved-thinking anti-distillation measure applies to Fable 5.1 and Opus 5.5 for API accounts created on or after August 31, 2026.

Vetted organizations, including academic labs, startups, and pharmaceutical companies, can apply to a new Life Sciences Verification Program to use Opus 5.5 for biology research. Anthropic said it will expand its Cyber Verification Program in the coming weeks with three tiers of increasingly permissive trusted access, including access to Claude Mythos models, and that Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks with many of the same improvements to performance, efficiency, and safety.