Claude Opus 5 Launches With Fable 5 Performance at Half the Price
Claude Opus 5 collapses the price-performance gap between Anthropic's tiered model lineup, offering near-flagship capability at a cost that makes it viable for routine coding and knowledge work, while the safety relaxation compared to Fable 5 suggests Anthropic is calibrating restrictions per model tier rather than applying a uniform policy.
Key Facts
- Claude Opus 5 launched on July 24, 2026, priced at $5 per million input tokens and $25 per million output tokens, half the cost of Claude Fable 5.
- On Frontier-Bench v0.1, Opus 5 scored higher than Fable 5 at a lower per-task cost.
- On ARC-AGI-3, Opus 5 scored 30.2% compared to about 20% for Fable-series models.
- Safety intervention frequency is reduced by approximately 85% compared to Claude Fable 5.
- Fast mode responds about 2.5 times faster than normal mode at double the cost.
Reporting from 3 sources: ASCII.jp, GameBusiness.jp, GIGAZINE.
Anthropic released Claude Opus 5 on July 24, 2026, a new AI model that delivers performance comparable to the top-tier Claude Fable 5 while priced the same as the previous generation Claude Opus 4.8. API pricing is $5 per million input tokens and $25 per million output tokens, half the cost of Fable 5. In benchmarks, Opus 5 outperformed Fable 5 on Frontier-Bench v0.1 for coding agent performance and scored higher on the ARC-AGI-3 test for adapting to unknown environments, recording 30.2% compared to about 20% for Fable-series models. On CursorBench 3.2 under maximum reasoning settings, the gap with Fable 5 was within 0.5 points at half the per-task cost. Anthropic emphasized improvements in autonomous judgment and task completion for agent use, including the ability to argue against instructions when proposing better alternatives. Safety features include reduced intervention rates by about 85% compared to Fable 5, with fallback to Opus 4.8 for blocked cybersecurity and biology requests. A new Fast mode responds about 2.5 times faster at double the cost. Opus 5 is available through Claude's paid plans and API, and becomes the default model for the Max plan.
Anthropic released Claude Opus 5 on July 24, 2026, positioning it as a cost-effective alternative to the flagship Claude Fable 5. The model is priced at $5 per million input tokens and $25 per million output tokens, identical to the previous Claude Opus 4.8 and half the cost of Fable 5. Third-party evaluator Artificial Analysis rated Opus 5 one point higher than Fable 5 in intelligence performance and ranked it first in both coding and agent performance.
Anthropic highlighted Opus 5's improvements in autonomous judgment and task completion for agent use cases. The model can self-verify and iterate on its own work, reducing the pattern of reporting tasks as done when issues remain unresolved. It can also argue against instructions based on task purpose, with examples including pushing back on proposed designs and writing its own image processing pipeline to work around limitations. On OSWorld 2.0, which evaluates PC operations, Opus 5 surpassed Fable 5's highest scores at all cost levels while achieving costs just over one-third.
Safety features include a misalignment behavior score of 2.3, the lowest in the Opus series. The frequency of intervention by cybersecurity classifiers is reduced by about 85% compared to Fable 5. Requests for binary scanning, penetration testing, and exploit generation are blocked and redirected to Opus 4.8. For biology-related requests that Fable 5 blocks, they are now redirected to Opus 5 instead of Opus 4.8. A new Fast mode operates at about 2.5 times normal speed at double the base price. Two beta features include mid-conversation tool changes without invalidating the prompt cache, and automatic redirection of safety-flagged requests to another model.
Synthesized by Yomimono from the 3 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.