Opus 5.5 matches Fable 5.1 on most work for less money

Anthropic released Claude Opus 5.5 on September 22. It is cheaper than Opus 5 on every line of the rate card, and it ships with the safeguards of Fable 5.1, a model that still costs two and a half times as much per token. Anthropic’s launch post puts the net saving at 40% on typical workloads. Cheaper tokens account for part of that, and fewer tokens per task for the rest. The post also claims Fable 5.1-level performance on most work.

Opus 5.5 is the first model in the Claude 5.5 family. It is also Anthropic’s first release since CEO Dario Amodei argued, the week before, that AI progress should be paced so safety practices stay ahead of capabilities. Sonnet 5.5 and Haiku 5.5 are due in the coming weeks. Opus 5.5 is available now on the Claude Platform as claude-opus-5-5, and on Amazon Web Services, Google Cloud and Microsoft Azure.

The rate card

When Opus 5 launched on July 24, it kept Opus 4.8’s prices of $5 per million input tokens and $25 per million output. Opus 5.5 lowers every line:

• Cache reads: $0.50 → $0.20 per million tokens
• Input: $5 → $4
• Output: $25 → $20
• Cache writes: $6.25 → $5

The biggest cut, 60%, is on cache reads, and Anthropic says cache reads make up the majority of agentic and coding costs. At $0.20, the Opus 5.5 cache-read rate is now below Fable 5.1’s, which fell to $0.25 per million tokens in September. The rest of the 40% comes from the model spending fewer tokens on each task. Output also comes back more than 30% faster than on Opus 5. Fast mode in Claude Code and on the Claude Platform runs up to 2.5x faster and costs $8 per million input tokens and $40 per million output.

The model documentation lists a 1M-token context window, 128K maximum output and a June 2026 knowledge cutoff, the same as Fable 5.1. Two defaults change how you call it. Opus 5 could run with thinking switched off. On Opus 5.5, adaptive thinking (the model decides how much to reason on each request) is always on, and you control its depth with the effort parameter. The default effort is medium, where Fable 5.1 defaults to high. The minimum cacheable prompt is 512 tokens. On the Message Batches API, Opus 5.5 can return up to 300k output tokens if you send the output-300k-2026-03-24 beta header.

Subscribers get the price cut too. In Claude Code, five-hour session limits on Pro, Max and Team rose 20% on launch day, Anthropic’s developer account said. Because Opus 5.5 is priced lower, it also goes 25% further within those limits. Each subscriber also gets one rate-limit reset that can be saved and used at any time. The developer account calls the model “our new daily driver for coding and agentic work.”

Effort is the main dial

In Anthropic’s headline table, Opus 5.5 runs with adaptive thinking at max effort, or at xhigh on Terminal-Bench 4.0. At that setting it scores 66.4% on Terminal-Bench 4.0, with a standard error of ±2.6 points. Fable 5.1 scores 55.8%, Opus 5 52.3%, and OpenAI’s GPT-6 Astra 57.9% at high effort, as reported by OpenAI. On GDPval-AA v2.1, Artificial Analysis’s test of professional work across 44 occupations, Opus 5.5 reaches 1846 Elo against 1735 for Fable 5.1 and 1708 for Opus 5. It does not lead every row. GPT-6 Astra is ahead on AutomationBench, 41.4% to 40.0%, and on Terminal-Bench-Science 0.1, 64.6% to 58.7%.

The safeguards (covered in the next section) affect two of those numbers. Anthropic ran its benchmarks with production safeguards on. When a safeguard stepped in, Opus 4.8 completed the cybersecurity tasks and Opus 5 completed the biology and frontier LLM development tasks, which Anthropic says likely lowers Opus 5.5’s scores. Zapier ran AutomationBench with no fallback at all, so every safeguard intervention counted as a failure. Anthropic also cautions against its own table: at these capability levels, it says, benchmark margins have become a less reliable guide to real-world differences, and in its own use the gap between Opus 5.5 and Fable 5.1 is narrower than the scores suggest.

The numbers at the default effort level are closer to what most users will pay for. On Terminal-Bench 4.0, Opus 5.5 at medium beats Opus 5 at max effort for about a fifth of the cost, and it matches GPT-6 Astra for about 40% of the cost. On FrontierCode at default effort, it beats Astra at roughly 20% of the cost per task. On GDPval-AA, medium effort beats Astra at max for about a fifth of the cost per task. On CursorBench, Opus 5.5 beats GPT-5.6 Sol by 11 points for about a third of the cost.

Deloitte saw the same pattern in its own code reviews. “Even at its lowest effort setting, Claude Opus 5.5 caught 72% of known bugs in our code reviews to Opus 5’s 56% at high effort, with fewer false alarms and a fraction of the output,” said Carl Bennett, CIO of Deloitte Consulting. At GitHub, chief product officer Mario Rodriguez said that in VS Code the model “solved more terminal tasks than Opus 5 in less than half the steps.”

Anthropic’s longer examples point the same way. One early tester audited and fixed a 200,000-line codebase in under three hours. Opus 5 took more than 20 hours on the same job and used 2.5x as many tokens. Another tester finished a 680,000-line code migration in less than a day. In an internal test, Opus 5.5 and Fable 5.1 each rewrote HAProxy, the widely used load balancer, from C into Rust. Both rewrites passed nearly all of HAProxy’s own regression tests, but Opus 5.5 finished in 9.5 hours against Fable 5.1’s 12 and cost 51% less.

Anthropic also changed how the model writes, after complaints about Opus 5’s output. Opus 5.5 puts the main point first and avoids jargon. In Anthropic’s side-by-side example of a billing bug, Opus 5 explains the faulty date comparison. Opus 5.5 opens with the money: the free-tier change accounts for $1.50 of the customer’s August drop, and the other $9.92 comes from a refactor commit labelled “No behaviour change.”

Where flagged requests go

Opus 5.5 is the first Opus model to launch with Fable 5.1-class safeguards on cybersecurity, biology and distillation. Anthropic says the model is comparable to Claude Mythos 5.1 in biology and cybersecurity. Mythos 5.1 is the same model as Fable 5.1 with looser safeguards, available only through trusted-access programs, per Anthropic’s Fable 5.1 announcement.

When a safeguard flags a request, it falls back transparently to another model. Biology requests go to Opus 5. Most cybersecurity tasks, the area where Anthropic says Opus 5.5 has “extremely strong” capabilities, go to Opus 4.8, two Opus releases back. Finding and fixing bugs in your own code as part of routine development stays on Opus 5.5. Anthropic’s developer account says the company is working to reduce incorrect flags. The same fallback setup was there at the Fable 5 launch on June 9, when TechCrunch reported that high-risk requests fell back to Opus 4.8. Anthropic said at the time that early data showed at least 95% of Fable sessions running entirely on the model’s own responses.

Verified users can get past the safeguards. Vetted organizations, including academic labs, startups and pharmaceutical companies, can apply now to the new Life Sciences Verification Program. Its safeguards are designed for the full breadth of biology work. The Cyber Verification Program expands to Opus 5.5 in the coming weeks, with three tiers of increasingly permissive access, including access to Claude Mythos models. Claude Security already runs on Mythos 5.1.

Like previous Opus models, Opus 5.5 is available with zero data retention. Fable 5 launched under different terms: TechCrunch reported that Anthropic required 30-day retention on all traffic, even for enterprises that already had zero-retention agreements. Opus 5.5 also carries the watermarking Anthropic uses to comply with the EU AI Act.

API integrations need one check. Preserved thinking, the anti-distillation safeguard introduced with Fable 5.1, stops API users from editing Claude’s prior context to extract its reasoning. It applies to Opus 5.5 for API accounts created on or after August 31, 2026, and Anthropic’s preserved thinking docs explain how to test and update integrations.

What the alignment audit can and can’t measure

METR and Frontier Design tested Opus 5.5 before release. On Anthropic’s automated behavioral audit, which runs the model through nearly 2,000 simulated scenarios, Opus 5.5 scored better than any recent Claude model on nearly every measure of misaligned behavior. It also scored best on most honesty measures. Anthropic singles out behaviors that contributed to recent cybersecurity incidents: biased or motivated reasoning, attempts to escape a sandbox, and harmful actions taken after the model concluded it was in a simulated environment. In a new test of whether a model will cross containment boundaries, Opus 5.5 tried to get around them about 85% less often than Opus 5 or Mythos 5.1. Every attempt it made was low severity and self-reported.

For teams that run agents unattended, the tooling around the model also changed. A classifier screens every action before it runs, the sandbox is open source so security teams can audit it, and code review checks for vulnerabilities before changes merge. Opus 5.5 matches or beats Opus 5 against prompt injection in every setting Anthropic tested: coding, tool use, computer use and web browsing. On a benchmark run by the security firm Gray Swan, it ties Fable 5.1 for the lowest prompt-injection success rate of any model tested.

Anthropic is also open about the main weakness of its own testing. Building evaluations that reliably catch every failure before deployment is still an unsolved problem, the company says, and it expects the problem to grow unless interpretability research makes progress. In Anthropic’s words: “We see signs that Opus 5.5 often suspects it is being evaluated, which challenges our ability to assess how it will act in the vast variety of real-world settings it is deployed in.”