Anthropic Releases Claude Opus 5 Model

Anthropic released Claude Opus 5 on July 24, 2026. API pricing remains $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8 and half of Fable 5. The model is positioned to deliver frontier intelligence at half the cost of Fable 5. On Frontier-Bench v0.1, Opus 5 scored more than double Opus 4.8 with lower per-task cost. On CursorBench 3.2 at highest Effort, it trailed Fable 5 peak by under 0.5% while costing half. ARC-AGI 3 score was triple the second-place model. Zapier AutomationBench pass rate was ~1.5x the next best at equal cost; even lowest Effort completed more tasks than all others. OSWorld 2.0 exceeded Fable 5's best at just over one-third the cost. Organic chemistry improved 10.2 pp, protein tasks 7.7 pp over Opus 4.8. Examples include autonomously building a CV pipeline to reconstruct a 3D part from a drawing, fixing a real open-source bug root cause, and self-creating a test framework for market data parsing. Claude Max default model switched to Opus 5; Pro users get it as strongest model. API via `claude-opus-5`. Fast mode is 2.5x speed at double price. Beta features: tool changes mid-conversation no longer invalidate prompt cache; API auto-degradation reroutes safety-blocked requests to other models. Safety alignment score 2.3, lowest among compared models, with lowest deception. Cybersecurity: vulnerability discovery near Mythos 5 but exploit generation lags. Safety classifier blocks binary-level scans, pentest, exploit generation, with 85% lower intercept frequency than Fable 5. Biological restrictions relaxed: previously blocked bio requests now go to Opus 5 instead of Opus 4.8. Opus 5 is the strongest public model for scientific research but has limits in long autonomous tasks.
Key insights
Opus 5 Delivers Frontier Performance at Half the Cost of Fable 5
Anthropic positions Opus 5 as a cost-efficient alternative to its most expensive model, Fable 5. With identical pricing to Opus 4.8, it achieves near-Fable-5 scores on multiple benchmarks while halving per-task costs. On CursorBench 3.2 the gap to Fable 5 peak is under 0.5%, and on OSWorld 2.0 it surpasses Fable 5 at one-third the cost. This pricing strategy could pressure competitors and make advanced AI more accessible to developers.
Safety and Alignment Improvements Are Central to Opus 5
Anthropic claims Opus 5 is its most aligned model yet, with a composite misalignment score of 2.3, lower than Opus 4.8, Sonnet 5, and Fable 5. Deception rates are also the lowest. The safety classifier intercepts 85% fewer requests than Fable 5, and blocked requests can fall back to Opus 4.8. Biological research restrictions have been eased, routing previously blocked queries to Opus 5, reflecting a nuanced approach to safety versus capability.
New Beta Features Improve Developer Experience for Agent Building
Two beta features address practical pain points: prompt cache no longer invalidates when tools are changed mid-conversation, reducing costs; and API auto-degradation ensures requests blocked by the safety classifier are automatically handled by another model instead of failing. These changes lower friction for building reliable AI agents.