Claude Opus 5 costs half as much as Fable 5 and scores within 0.5% on coding

Claude Opus 5 catches its own mistakes, rewrites its answers when they fall short, and handles multi-step scientific and coding tasks from start to finish — at half the price of Fable 5, Anthropic‘s most powerful model. On coding benchmarks, Opus 5 performs within 0.5% of Fable 5; on OSWorld 2.0 computer-use tests, it actually outperforms the flagship at a third of the cost.

What makes Opus 5 structurally different from its predecessors is an effort toggle: a low, medium, or high setting attached to each API request. A legal team doing routine document review can run on low; a software engineering workflow debugging a complex distributed system can shift to high. The model adjusts how much compute it spends on the task, and therefore how much the user pays, without changing the model itself.

Pricing is $5 per million input tokens and $25 per million output tokens, the same rate as Opus 4.8, Opus 5’s direct predecessor. Fable 5, Anthropic’s flagship model, costs roughly twice as much. Opus 5 is now the default model on Claude Max subscriptions and available through the API endpoint claude-opus-5. A fast mode runs at 2.5 times the inference speed at twice the base token price, aimed at real-time pipelines where latency matters more than cost.

The limitations are real and Anthropic is explicit about them. Opus 5 still trails Mythos 5, the company’s specialized security model, on cybersecurity tasks, and shows no progress on dual-use biology, which Anthropic has defined as a hard ceiling for safety reasons. For autonomous, months-long biological research projects, Fable 5 remains the recommendation. The 0.5% coding gap may widen in practice for complex, open-ended problems that fall outside the benchmarks.

The release arrives against a backdrop of rapid-fire launches across the industry. OpenAI shipped GPT-5.6 in three tiers on July 9; Google released Gemini 3.6 Flash on July 21, cutting output token costs by 17%. Each wave raises the baseline while compressing the time enterprises have to evaluate and deploy what came before.

Opus 5 is available today through the Claude API, Claude.ai, Claude Code, and Claude Cowork, priced at $5 per million input tokens and $25 per million output tokens. Anthropic says the effort toggle is available immediately to all API users and Claude Max subscribers.

Opus 5 is Anthropic’s fourth model release in sixty days. The speed at which the cost-performance frontier is moving now creates a different problem for enterprise buyers: knowing when to stop evaluating and actually ship something.

Tags: Anthropic, Claude, AI models, Fable 5, Claude Max, Claude Opus 5