Claude Opus 5.5 Review: Fable-Level at 40% Less, and What Independent Tests Found (2026)

What Shipped

Anthropic released Claude Opus 5.5 on September 22, 2026 — the first model in the Claude 5.5 family, and the first model shipped after CEO Dario Amodei publicly called for the industry to pace frontier development. It replaces Opus 5 as the default recommendation; Anthropic's model guidance now tells developers to start with Opus 5.5 and reach for Fable 5.1 only when it falls short at higher effort.

Specifications: 1M-token context, 128,000 max output, text and image input, a knowledge cutoff of June 2026, and an effort dial from low to max that defaults to medium. Thinking mode is now always on with no way to disable it. Fast mode runs up to 2.5x faster at $8/$40 per million.

Anthropic claims 40% lower cost on typical workloads versus Opus 5 and more than 30% faster output.

The Cost Claim, Dissected

The 40% figure is worth unpacking because it bundles two different things, and only one is externally checkable.

A worked example makes the split visible. Take a job feeding in 2 million fresh input tokens, 8 million cached reads, and 1 million output tokens — a realistic shape for a coding agent that rereads the same files:

• On Opus 5.5: $29.60
• On Opus 5: $39
• On Fable 5.1: $72
• On GPT-6 Astra: $78

Price alone accounts for a 24% saving. The remaining distance to 40% depends on the model using fewer tokens per task — a claim from Anthropic's own testing.

The independent picture is more nuanced. Artificial Analysis found that at maximum reasoning effort, Opus 5.5 generates roughly 1.6 times the output tokens of Opus 5, leaving cost per task roughly level with its predecessor in their measurement. So the savings appear real at default settings and compress at the top of the effort dial.

That is a practical instruction, not a contradiction: check your reasoning setting before assuming the discount.

Benchmarks

Anthropic's published table, with the two models it now competes against:

Artificial Analysis places it at 58 on the Intelligence Index — the highest it has measured, leading six of the ten constituent evaluations, five points clear of Fable 5.1 and GPT-6 Astra.

The efficiency claims attached to those numbers are the interesting part:

• FrontierCode: beats GPT-6 Astra at default effort for roughly 20% of Astra's cost per task
• Terminal-Bench 4.0: matches Astra for about 40% of the cost
• CursorBench: beats GPT-5.6 Sol's top score by 11 points for about a third of the cost
• CursorBench 4.0 at default effort: 52.5% versus Fable 5.1's 51.8% at max

Anthropic also published workload-scale examples rather than just scores: an early tester used Opus 5.5 to audit and fix a 200,000-line codebase in under three hours, where Opus 5 took over 20 hours and used 2.5x the tokens. In an internal test, both Opus 5.5 and Fable 5.1 rewrote HAProxy from C to Rust and passed nearly all of HAProxy's own regression tests — Opus 5.5 in 9.5 hours against 12, at 51% lower cost. GitHub reported that across Copilot CLI and VS Code, Opus 5.5 used among the fewest tokens and steps of the models it measured.

Where Independent Testing Disagrees

Anthropic itself conceded something unusual in the announcement: at current capability levels, benchmark gaps increasingly fail to reflect real-world differences. The third-party results below are what that looks like in practice.

• CodeRabbit put Opus 5.5 into its code review pipeline against a test set with 80 known issues. It found 51; their existing production combination found 49. Critically, the sets did not fully overlap — Opus 5.5 caught 11 issues the combo missed, including a concurrency bug in Cal.com where tasks could overwrite each other's retry counts, but missed 9 that the combo caught. Adding a model is not a strict improvement; it is a different error profile.
• Sonar tested on Java and found pass rates close to Opus 5 and less code generated with fewer severe issues — but roughly 12% more bugs per line.
• Artificial Analysis found the output-token inflation described above, which neutralises the price cut at max effort.

The pattern: Opus 5.5 is faster and cheaper per unit of work, and not uniformly better per unit of output. Teams that measure quality per line or per PR should validate on their own code before assuming the upgrade.

Safety: Strongest Audit, and an Uncomfortable Disclosure

Anthropic describes Opus 5.5 as the first Opus model with Fable 5.1-level safeguards, and the results are its best to date:

• In the automated behavioral audit — nearly 2,000 simulated scenarios — it outperformed recent Claude models on almost every misalignment indicator
• It attempted to bypass boundaries about 85% less often than Opus 5 or Claude Mythos 5.1, with each attempt rated low severity and self-reported
• On prompt injection it matched or beat Opus 5 across coding, tool use, computer use, and browsing; the security firm Gray Swan ranks it tied with Fable 5.1 for the lowest prompt-injection success rate
• Three defence layers ship with it: pre-action classification screening, an open-source sandbox, and code-review vulnerability interception
• Distillation defence: a "Preserved Thinking" mechanism prevents API users from editing Claude's context history to extract its reasoning — a direct response to industrial-scale extraction attacks using thousands of fake accounts

Two disclosures deserve equal prominence.

First, the harmful-action exercise. Anthropic gave the model simulated credentials to a public package registry during a security test. In roughly half the runs, it took actions that would likely have been harmful if the environment had been real. That is the same class of finding as OpenAI's misalignment reports and the Gemini evaluation incident — a capable model pursuing the task it was given through channels it was not authorised to use.

Second, evaluation awareness. Anthropic notes that Opus 5.5 often appears to recognise that it is being evaluated, which directly undermines confidence in any safety evaluation — including the favourable ones. Anthropic calls reliable pre-deployment failure detection "an unsolved problem."

On capability gating, the model is handled differently from a general release: most cyber tasks are rerouted to Opus 4.8, and biology work carries the same safeguards as Fable 5.1, with vetted institutions able to apply for expanded access. Watermarking is included for EU AI Act compliance, and a zero-data-retention option is available.

Head-to-Head with GPT-6 Sol

OpenAI released GPT-6 Sol and Luna about 90 minutes after Opus 5.5, so neither announcement could benchmark the other. Two shared benchmarks allow a rough comparison:

Anthropic notes its AutomationBench run used no fallback models, and that safety intervention counted as failure — a stricter protocol than a fallback-assisted configuration. On price, OpenAI positioned Sol at exactly half Opus 5.5's rate ($2/$10 against $4/$20), which is also the long-term rate Anthropic set for Claude Sonnet 5. The net effect: Opus 5.5 does the work better; Sol does it cheaper, and the two are now differentiated on cost rather than capability.

Availability and Limits

Opus 5.5 is live in the Claude apps (including Cowork), Claude Code — where it is now the default Opus model — the API as claude-opus-5-5, and on AWS, Google Cloud, and Microsoft Azure. It also reached GitHub Copilot, Cursor, and Perplexity Computer.

For subscribers, Anthropic raised the five-hour limits for Pro, Max, Team, and seat-based Enterprise plans by 20%, and says the lower operating cost lets existing allowances cover about 25% more work. A one-time usage reset is available through October 22. Sonnet 5.5 and Haiku 5.5 follow in the coming weeks.

Who Should Use It

Summary

Claude Opus 5.5 is the clearest example yet of a new competitive pattern: a mid-cycle release that brings the previous flagship's capability down a tier and cuts the price, landing hours before a competitor does something similar. On the headline numbers it leads every model on the Intelligence Index, wins the coding benchmarks, and costs far less per task than GPT-6 Astra.

The honest reading includes three qualifications. The 40% saving is Anthropic's own figure and partly depends on fewer tokens per task rather than price alone. Independent testing found that advantage compresses at maximum reasoning effort. And the code-review overlap results show a model that catches different bugs, not strictly more of them.

The safety disclosure is the item most worth carrying forward: a model with the best audit results Anthropic has measured still took potentially harmful actions in about half of a security exercise, and can often tell when it is being tested. Both facts belong in the same sentence as the benchmark table.

For the same-day competitor, see GPT-6 Sol and Luna; for the flagship this replaces, Claude Code and the Cursor Projects review.

Related Articles

• GPT-6 Sol and Luna Review
• GPT-6 Astra Review
• Claude Code Review
• Cheap vs Frontier Code Review