Anthropic released Claude Opus 5 on July 24, its fourth model in two months, but this one is different from the others. Where Fable 5 and Mythos 5 push the absolute frontier of capability at premium prices, Opus 5 is designed to be the model teams actually use every day, and the benchmarks suggest Anthropic may have pulled it off.
At $5 per million input tokens and $25 per million output tokens, the same price as the older Opus 4.8, the new model delivers performance that rivals Fable 5 on several key evaluations. On Frontier-Bench v0.1, a test of agentic coding, Opus 5 scored 43.3%, beating Fable 5’s 33.7% and GPT-5.6 Sol’s 34.4% outright. On ARC-AGI-3, which measures novel problem-solving, it hit 30.2%, roughly four times the score of the next-best competitor. On OSWorld 2.0, a test of computer-use capabilities, it surpassed Fable 5 at just over one-third the cost per task.
The headline numbers mask a more interesting story about how Anthropic is thinking about deployment. Opus 5 ships with an effort control dial, low, medium, or high per request, that lets users trade reasoning depth for speed and cost. A routine data-entry task can run on low effort to conserve tokens; a complex architecture problem can dial up to high. It is a flexible approach that acknowledges most production workloads do not need maximum reasoning on every call.
Safety architecture: a fallback, not a hard block
The safety system introduces a novel architectural choice: when Opus 5’s guardrails flag a request as potentially problematic, such as a prompt involving binary-based vulnerability scanning or exploit generation, the system can automatically fall back to Opus 4.8 rather than rejecting the request outright. This preserves the user’s workflow while routing high-risk operations to a model with less advanced offensive capability.
Anthropic describes Opus 5 as its “most aligned model to date” based on automated behavioral testing, with a lower rate of deceptive behavior than recent releases. On offensive cybersecurity tasks, it trails the specialized Mythos 5 but matches it on vulnerability identification.
Self-correction and the effort dial
Anthropic highlights a capability that may prove the most consequential for enterprise users: self-correction. In internal demonstrations, Opus 5 detected execution errors mid-task and built its own workarounds, in one case constructing a computer vision pipeline from scratch to extract geometry from raw pixels when the model could not directly view an object. Competing models failed after five attempts under identical conditions.
The model supports Anthropic’s full suite of developer features, including tool-switching mid-conversation without invalidating prompt caches, and the ability to route flagged API requests to another model automatically. It is the default model for Claude Max subscribers and the strongest available option for Claude Pro users.
Sources: “Claude Opus 5 delivers top coding scores at half the per-task cost” (Interesting Engineering, Jul 24, 2026); “Anthropic Introduces Claude Opus 5” (TheWinCentral, Jul 24, 2026); Anthropic official announcement