Anthropic has released Claude Opus 5, a model positioned between the frontier-level Claude Fable 5 and the previous Opus 4.8. It aims to provide near-frontier intelligence at roughly half the price of Fable 5, making it suitable for everyday use across coding, knowledge work, and scientific research.

On several benchmarks, Opus 5 sets new state-of-the-art results. It achieves the highest scores on Frontier-Bench v0.1 and GDPval-AA, and on CursorBench 3.2 at max effort its peak score is within 0.5% of Fable 5’s, but at half the cost per task. On ARC-AGI 3 (novel problem solving) its score is three times the next-best model, and on Zapier AutomationBench its pass rate is about 1.5 times the next-best model at the same cost per task. On OSWorld 2.0 (computer use) it outperforms every other model at any given cost and surpasses Fable 5’s best result at just over a third of the cost.

Compared to Opus 4.8, the model shows meaningful gains in scientific domains. It improves on every life sciences evaluation, with notable gains on organic chemistry tasks (10.2 percentage points higher on spectroscopy inference) and protein prediction tasks (7.7 percentage points higher).

Early-access testers reported that Opus 5 exhibits stronger agency, self-verification, and iterative problem-solving. Engineering firms noted it can build complete workflows from scratch, push back on flawed designs, check its own work for edge cases, and maintain high performance with fewer reasoning tokens (up to 26% fewer tokens compared to Opus 4.8 at comparable quality). Several testers described its judgment and consistency as a clear generational step up.

On alignment and safety, Opus 5 scored 2.3 on an automated behavioral audit of misaligned behavior, the lowest (best) among recent models. It shows lower rates of deceptive behavior and is less susceptible to being tricked into misuse. However, the model does not advance the frontier in risky dual-use capabilities: it remains behind Mythos 5 on both biology research and offensive cybersecurity. While Opus 5 is close to Mythos 5 at identifying software vulnerabilities, it is considerably less capable at developing exploits for them. Safeguards have been calibrated to be proportionally less restrictive than those on Fable 5, and flagged requests fall back to Opus 4.8 by default.

Pricing is $5 per million input tokens and $25 per million output tokens, matching Opus 4.8. Fast mode runs at around 2.5 times default speed for twice the base price. Opus 5 is available on the Claude API, Claude Max, and Claude Pro.