Anthropic released Claude Opus 5.5 on September 22, 2026. The model arrives two months after Opus 5, which launched on July 24, 2026, and it targets better performance without preserving the old model’s appetite for compute. Anthropic describes it as “the strongest-performing model we’ve tested to date” on its internal alignment testing.
The headline change is cost. Claude Opus 5.5 costs 40% less to run than Opus 5, while performing at roughly the level of Claude Fable 5.1 on most work. Sonnet 5.5 and Haiku 5.5 are also set to launch soon after, giving Anthropic a broader model lineup rather than one expensive flagship doing all the talking.
Performance That Scales With the Budget
Opus 5.5 scores 58 on the Artificial Analysis Intelligence Index at maximum reasoning effort. Its output speed ranges from 74 to 86 tokens per second, depending on the effort level, while the cost per intelligence-index task ranges from $0.55 at low effort to $5.98 at maximum effort.
That spread matters because the model’s strongest results do not require maximum effort for every task. Walleye Capital reported that Opus 5.5 largely solved its evaluation suite at the lowest effort setting, and an internal test found that it analyzed a fictional company merger in 63 minutes, compared with 93 minutes for Opus 5.
The coding results are more concrete than the usual benchmark parade. An early tester completed a 680,000-line code migration in under a day, while another tester audited and fixed a 200,000-line codebase in under three hours. Opus 5.5 also translated HAProxy from C to Rust in 9.5 hours; Fable 5.1 took 12 hours for the same task.
On FrontierCode, Opus 5.5 beats GPT-6 Astra for roughly a fifth of the cost per task. It matches Astra on Terminal-Bench 4.0 for about 40% of the cost, then beats GPT-5.6 Sol on CursorBench by 11 points for roughly a third of the cost. The pattern is clear: Anthropic is selling throughput and economics alongside raw capability.
Safety and Real-World Testing
Opus 5.5 ships with three coding-specific safeguards: a classifier, an open-source sandbox, and a code review process. In an independent benchmark by AI security firm Gray Swan, Opus 5.5 tied Fable 5.1 for the lowest prompt injection success rate.
Its internal research testing also produced a notable result. Opus 5.5 cleared the quality bar in 16 of 18 attempts when writing a company earnings report, while neither Fable 5.1 nor Opus 5 cleared the quality bar even once. That is a narrow test, but the gap is large enough to resist polite dismissal.
External evaluations point in the same direction across several types of work. Deloitte Consulting caught 72% of known code review bugs with Opus 5.5 at its lowest setting, compared with 56% for Opus 5. Rogo beat Opus 5’s best result using about 60% fewer output tokens with Opus 5.5.
Legal and research users also reported useful gains. LexisNexis consistently identified relevant citations and legal frameworks in early evaluations with Opus 5.5, while Thomson Reuters Labs reported better results on internal benchmarks. Hebbia covered 86.6% of an expert grading with the model.
Claude Opus 5.5 is not presented as a clean victory lap; its task costs still rise from $0.55 to $5.98 as reasoning effort increases. But the combination of lower operating cost, faster completion, coding safeguards, and stronger test results gives Anthropic a practical upgrade—not just another model number wearing a ceremonial hat.
