Claude Opus 5 pricing undercuts Fable 5 as Anthropic claims benchmark gains

Anthropic announced Claude Opus 5 pricing at $5 per million input tokens and $25 per million output tokens, half the listed token rates for Claude Fable 5. The company is pitching the new flagship model as a lower-cost way to get near-Fable 5 performance, with stronger results on several coding and knowledge-work benchmarks, according to Anthropic’s own testing.

Opus 5 is becoming the default model for Claude Max users and the most capable model available to Claude Pro subscribers. Anthropic says the model keeps the 1 million-token context window and base rates used by Opus 4.8. A Fast Mode is also available, which Anthropic says runs 2.5 times faster and costs twice as much.

How much does Claude Opus 5 cost?

Claude Opus 5 costs $5 per million input tokens, $6.25 per million five-minute cache writes, $10 per million one-hour cache writes, $0.50 per million cache hits and updates, and $25 per million output tokens, according to Anthropic’s pricing. Claude Fable 5 and the limited-availability Claude Mythos 5 are listed at $10 per million input tokens and $50 per million output tokens, with proportionally higher cache rates.

Base token rates are only part of the cost picture. Anthropic’s recent model releases have drawn scrutiny because task-level costs can rise even when posted token prices stay flat. The company says Opus 5 improves value at each of its five effort settings: low, medium, high, xhigh, and max. Anthropic’s prompting guidance recommends low and medium for many use cases, while coding and agentic workloads should start at xhigh.

The max setting is not uniformly better. Anthropic’s charts show Opus 5 scoring slightly lower at max than at the next-highest setting on Frontier-Bench v0.1 and the Artificial Analysis Coding Agent Index, despite the higher cost.

Where Anthropic says Opus 5 leads

On Frontier-Bench v0.1 agentic terminal coding, Anthropic reports Opus 5 at 43.3%, ahead of Fable 5 at 33.7%, GPT-5.6 Sol at 34.4%, and Opus 4.8 at 21.1%. On GDPval-AA v2, a knowledge-work benchmark, Opus 5 posted an Elo score of 1,861, compared with 1,747 for Fable 5 and 1,736 for GPT-5.6 Sol.

Opus 5 does not top every test Anthropic disclosed. On DeepSWE v1.1 agentic coding, GPT-5.6 Sol led with 72.7%, followed by Fable 5 at 69.7% and Opus 5 at 68.8%. Anthropic also says Fable 5 performs better on health tasks, while Mythos 5 leads on legal benchmarks.

The most unusual result is ARC-AGI-3, a benchmark intended to test novel problem-solving rather than pattern recall. Anthropic reports Opus 5 at 30.2%, compared with 7.8% for GPT-5.6 Sol and 1.5% for Opus 4.8. Anthropic did not provide a Fable 5 score for that test, and the company’s materials do not establish whether the gap carries over to production workloads.

Tool use, cyber limits and beta features

Anthropic says Opus 5 is better at revising its own work and creating tools when needed. In one Frontier-Bench task, the model was given a drawing of a machine part and asked to build a 3D model in FreeCAD without direct access to view the drawing. Anthropic says Opus 5 wrote a computer vision pipeline to extract geometry from pixels and solved the task, while no other model completed it within five attempts.

The company also says Opus 5 can support source-code vulnerability research but blocks binary vulnerability scanning, penetration testing and exploit generation. Anthropic says its cyber classifiers trigger about 85% less often than Fable 5’s. Blocked requests in Claude.ai, Claude Code and Claude Cowork fall back to Opus 4.8, matching the approach used for Fable 5.

Alongside the model, Anthropic is releasing beta features for developers: Mid-Conversation Tool Changes, which allow tool swaps during a conversation without invalidating the prompt cache, and Automatic Fallbacks, which route blocked API requests to another model.

This story draws on original reporting from The Decoder.