Anthropic released Claude Fable 5 last week, their new model specifically designed for software engineering.
The headline number is 95% on SWE-bench. This benchmark measures AI's ability to solve real GitHub issues—the previous best was Opus 4.8's 88%. A 7-point jump is unusual in AI; typically progress comes in 1-2 point increments.
The Senior Engineer evaluation scored 91/100. This test measures more than just writing code—it includes architectural decisions, design pattern recognition, and system-level tradeoffs. In other words, Fable 5 starts thinking like an experienced engineer, not just spitting out code.
The context window expanded to 1 million tokens with a 128K token output limit. This means you can feed it an entire large codebase and get back a complete, multi-file implementation in a single response. No chunking, no manual context management, no "continue from where you left off."
Pricing isn't cheap: $10 per million tokens input, $50 per million tokens output. This is the most expensive major API currently available. Anthropic's positioning is clear—it's worth the price for hard problems; for routine tasks, Haiku or Sonnet are more cost-effective.
At the same time, Anthropic released Claude Mythos 5, their research-oriented model. It's optimized for mathematical proofs, scientific reasoning, and deep analytical work. Releasing both models simultaneously shows Anthropic is pursuing a specialization strategy—one model for each use case rather than one-size-fits-all.
Fable 5 vs Major Models Comparison
| Metric | Claude Fable 5 | Claude Opus 4.8 | GPT-5.5 |
|---|---|---|---|
| SWE-bench | 95% | 88% | 82% |
| Context Window | 1M | 200K | 1M |
| Output Limit | 128K | 32K | 64K |
| Input Price | $10/M | $15/M | $2.5/M |
| Output Price | $50/M | $75/M | $10/M |
Fable 5 is cheaper than Opus 4.8 but performs better. This pricing strategy likely aims to compete with OpenAI's GPT-5.5—the latter is much cheaper but has a clear gap in coding capability.
Who Should Use Fable 5
Professional developers working on complex problems: architectural design, large codebase debugging, multi-file refactoring, and greenfield system design. For routine code completion and simple bug fixes, Sonnet or Haiku are sufficient.
Over 80% of code merged into Anthropic's production codebase is generated by Claude. Claude Code's annualized revenue is close to $6.3 billion, capturing 54% of the AI coding agent market.