Anthropic has released Fable 5, and the AI research community is paying close attention. The company built its reputation on safety-first AI development, but Fable 5 signals something new: Anthropic is no longer content being the safety-focused alternative. They want to win on capability too — and the benchmarks suggest they might be doing exactly that.
What Is Fable 5?
Fable 5 is the latest model in Anthropic's Claude family, sitting above Opus 4 and Sonnet 4 in capability and cost. It is designed for the most demanding tasks: multi-step research, complex code generation across entire repositories, scientific reasoning, and nuanced long-form writing.
Anthropically describes it as their first model where safety and capability reinforce each other rather than trade off. The constitutional AI training has been significantly revised so that the model's refusals and guardrails are grounded in explicit reasoning rather than pattern matching — meaning it can explain why it won't do something, and can often find a safer way to accomplish the underlying goal.
How Does It Compare to GPT-5 and Gemini 3?
On coding benchmarks (HumanEval, SWE-Bench), Fable 5 leads the field with a 92.4% pass rate — edging out GPT-5 at 91.1% and Gemini 3 Ultra at 89.7%. On mathematical reasoning (MATH benchmark), Fable 5 scores 91.8%, compared to GPT-5's 90.2%.
Where GPT-5 still holds an edge is in multimodal image understanding and real-time web browsing. Gemini 3 maintains advantages in Google Workspace integration and video understanding. But in pure text reasoning — the core of most enterprise use cases — Fable 5 has pulled ahead.
- HumanEval coding: Fable 5 92.4% vs GPT-5 91.1% vs Gemini 3 89.7%
- MATH benchmark: Fable 5 91.8% vs GPT-5 90.2% vs Gemini 3 87.4%
- Long-context retention (200k tokens): Fable 5 leads by significant margin
- Safety benchmark (AdvBench): Fable 5 99.1%, the highest ever recorded
What It Means for Developers
For developers building on the Anthropic API, Fable 5 is available now via the model ID `claude-fable-5`. Pricing is $15 per million input tokens and $75 per million output tokens — comparable to GPT-5 Turbo and significantly cheaper than Gemini 3 Ultra for most workloads.
The extended 200k context window handles with noticeably less mid-context degradation than previous models. Developers working on code review, document analysis, and agentic workflows report that Fable 5 maintains task focus across far longer contexts than Claude Opus 4 or GPT-5.
Fable 5 is available now via the Anthropic API. Model ID: claude-fable-5. Context window: 200k tokens.
The Bigger Picture
Fable 5's release marks a pivotal moment in the AI landscape. For over a year, the narrative has been that OpenAI leads on capability while Anthropic leads on safety. Fable 5 challenges that framing directly.
The real question isn't which model scores highest on benchmarks — it's which model you can trust to run autonomously for hours on important tasks. On that dimension, Anthropic's track record and Fable 5's transparent reasoning may prove to be the decisive advantage. We'll be watching developer adoption closely over the next 90 days.
