Claude Fable 5 for Coding: Benchmarks, Pricing, and Where to Use It

On June 9, 2026, Anthropic released Claude Fable 5, the first publicly available version of its closely watched Mythos model and its most capable model to date. The headline for developers: it is built for coding. Anthropic pitches it at large migrations, complex multi-file implementations, and autonomous sessions that run for hours while the model writes its own tests and uses vision to check its output against the goal.

This is not a breaking-news post. It is a practical look at the coding numbers, what the model costs, and where you can actually use it today, so you can decide whether it earns a place in your workflow or stays a tool you reach for only on the hard problems.

The coding benchmarks

Fable 5’s launch numbers are the strongest Anthropic has published. On SWE-bench Pro, the harder successor evaluation, Anthropic’s own chart puts it well ahead of the field:

It also posted roughly 95% on SWE-bench Verified and 72.9% on CursorBench at max effort, and led the FrontierCode benchmark. The short version: on the hardest software-engineering tasks the benchmarks measure, Fable 5 has a clear lead right now.

One honest caveat, the same one we apply in our best AI coding assistant guide: treat the absolute percentages with some skepticism. OpenAI’s auditors found that a large share of the hardest SWE-bench Verified problems had flawed test cases and that frontier models could reproduce solutions from memory. The rankings hold up better than the raw numbers do.

Pricing: powerful, not cheap

Fable 5 costs $10 per million input tokens and $50 per million output tokens, with batch pricing at $5 and $25. That is roughly two to five times the per-token cost of the cheaper frontier models. But one discount changes the math for exactly the work Fable 5 is built for: prompt caching bills repeated input at one tenth the rate, a 90% discount, so a cache hit costs about $1 per million input tokens instead of $10. Agentic and multi-file runs send the same large prefix (the system prompt, the codebase, the spec) on every step, so in practice the running cost for its best use case is far lower than the headline rate suggests. It is still a precision instrument rather than the model you leave on autocomplete all day, but the caching discount is the reason a long agent session is more affordable than the sticker price implies.

Where you can use it for coding

As of launch, Fable 5 is live across most of the places developers already work:

• Claude apps and Claude Code (terminal agent), included in Claude Pro and the Max tiers.
• The Anthropic API and Amazon Bedrock, for builders and teams.
• GitHub Copilot, generally available as of June 9.
• Cursor, selectable as a model in the IDE.

Anthropic is also letting Claude subscribers try Fable 5 free through June 22, 2026, which is the no-risk way to test it before committing token budget. For how these coding tools compare overall, see our best AI coding assistant guide, and for the broader model picture our Claude vs ChatGPT vs Gemini comparison.

Should you switch?

Not wholesale. The smart pattern is to keep a cheaper, faster model for routine coding and reach for Fable 5 on the jobs where its lead pays for the price: large migrations, gnarly multi-file refactors, and long autonomous agent sessions that would otherwise need a senior engineer babysitting them. It is worth knowing that in high-risk domains like biology, chemistry, cybersecurity, and model-distillation prompts the model deliberately blocks and falls back to Claude Opus 4.8, a safety design Anthropic shipped alongside it.

If you spend real money on AI coding tools, the free trial window through June 22 is the obvious move: point it at your hardest open ticket and see whether the benchmark lead shows up in your actual codebase.