Claude Opus 5: Near-Fable Performance at Half the Price
Claude Opus 5 launched July 24, 2026 at the same price as Opus 4.8, $5 per million input tokens, $25 per million output tokens, while scoring within 0.5% of Fable 5 on CursorBench 3.2 and tripling the next-best model on ARC-AGI 3. It is the new default on Claude Max and the strongest model on Claude Pro.
Claude Opus 5 is available today on every platform. The model ID is claude-opus-5. Pricing is unchanged from Opus 4.8: $5 per million input tokens and $25 per million output tokens. It is the new default model on Claude Max and the strongest model on Claude Pro.
On CursorBench 3.2, Opus 5 lands within 0.5% of Fable 5 at half the per-task cost. On ARC-AGI 3, a benchmark for solving novel problems, it scores three times higher than the next-best model. On Zapier AutomationBench, which measures end-to-end business task completion, its pass rate is roughly 1.5 times higher than any competitor at the same cost per task.
Opus 5 adds an effort setting you control. Effort is a dial from low to max. High effort means deeper reasoning; low effort means fewer tokens and faster results. This is not a quality compromise: even at its lowest setting, Opus 5 passes more AutomationBench tasks than any other model at any effort level.
Use high effort (the default) for complex, long-horizon, or high-stakes work. Drop to low for drafts, summaries, or speed-sensitive tasks. Fast mode is also available, running approximately 2.5 times the default speed at twice the base price.
Opus 5's most important upgrade is self-verification. It checks its own work before returning results. In benchmark testing, it caught a hidden mobile-layout bug and a product hidden below the mobile fold without being asked to look for either, and fixed both before handing the work back. On code review tasks, it verifies branches, checks templates, and considers test implications before marking work complete.
You can activate this behavior explicitly. Add a verification instruction to your prompt, for example, "cross-check every figure before returning" or "flag anything you're uncertain about rather than guessing." Opus 5 will catch mistakes that other models report as resolved.
Opus 5 also holds its positions under pushback. In testing, it declined to fold when challenged on a design decision, narrowed its objection to the single flaw, and proposed a compromise. Treat that as a feature: ask it to explain its objection and propose a fix, rather than restating your position more firmly.
Two new beta features ship alongside Opus 5. First, mid-conversation tool changes: within an active conversation, developers can now add or remove tools without invalidating the prompt cache. Second, automatic fallbacks: when a request is flagged by a safety classifier on Opus 5 or Fable 5, it routes automatically to the best available model instead of returning an error.
Opus 5 is also Anthropic's most aligned model to date. It scores 2.3 on the automated behavioral misalignment audit, the lowest of any recent model, and is the least susceptible to being tricked into misuse. Like Opus 4.8, it carries no data retention requirements for general access.
Opus 5 is Anthropic's most capable everyday model. At the same price as Opus 4.8, it delivers near-Fable 5 performance on coding, knowledge work, and science. Use the effort setting to trade reasoning depth for speed. Explicitly ask it to verify before returning, it will catch what other models miss. Enable automatic fallbacks so flagged API requests route to the next-best model rather than failing. The model ID is claude-opus-5.