Claude Fable 5 Takes Both Crowns

This is the most consequential leaderboard move we have seen in months. Text moved from Claude Opus 4.6 Thinking at 1504 to Claude Fable 5 at 1510. Solid, meaningful, but not absurd.

Code is different. The previous #1, Claude Opus 4.7 Thinking, sat at 1566 after today’s scrape. Fable 5 landed at 1665. That is not a normal version bump. That is a new tier appearing on the chart.

The “safe for general use” framing matters

Anthropic’s announcement calls Fable 5 a “Mythos-class” model made safe for general use. The company is also launching Claude Mythos 5 for a much narrower trusted-access track. Same underlying model family, different access posture.

That tells you how Anthropic wants this release understood: Fable 5 is not just Opus 4.9 with a nicer name. It is the public version of a more capable class of model, with conservative routing safeguards that can fall back to Opus 4.8 on sensitive requests. Anthropic says those safeguards trigger in less than 5% of sessions on average.

The annoying tradeoff is obvious: some harmless edge cases will get over-filtered. But from a leaderboard perspective, the signal is clean. Even with the safer public variant, users are preferring Fable 5 over everything else.

What Anthropic says changed

The official launch is unusually specific about where Fable 5 pulls ahead: longer, more complex tasks. Anthropic calls out software engineering, knowledge work, vision, scientific research, memory, and long-context performance. The common thread is not “better answers.” It is longer autonomous work without losing the plot.

Early customer quotes match the leaderboard shape. Stripe reportedly used Fable 5 for a codebase-wide migration in a 50-million-line Ruby codebase. Cursor called it state of the art on CursorBench. Cognition called it the highest-scoring model on FrontierBench. Hebbia said it was the first model to break 90% on its core analytics benchmark.

Read those as directional, not gospel. Vendor launch quotes are marketing. But they line up with the arena result: Fable 5’s advantage is biggest where tasks are messy, long-running, and easy for weaker models to derail.

The price is a statement too

Fable 5 costs $10 per million input tokens and $50 per million output tokens. That is twice Opus 4.8’s listed $5 / $25 pricing on the arena table. Anthropic is not hiding the premium positioning.

The obvious question: is a 99-point code lead worth 2× the price? For casual chat, probably not. For a migration, a high-stakes agent run, or a week of engineering work compressed into a day, the math changes fast. The expensive model can be the cheap model if it wastes fewer turns and needs less babysitting.

What this does to the race

Before Fable 5, Anthropic already owned the code board. After Fable 5, the top of the code board looks almost unfair. The nearest non-Anthropic model is Alibaba’s Qwen3.7 Max at #8. OpenAI’s strongest visible Codex-harness entry is #15. Google’s Gemini 3.5 Flash is #14.

Text is more competitive, but still brutal. Fable 5 at #1, four older Claude variants immediately behind it, and Meta’s Muse Spark at #6. OpenAI’s best text entry is GPT-5.5 High at #10.

My read: Anthropic has shifted from “best coding model” to “best long-horizon work model.” That is a bigger category. If Fable 5’s arena lead holds over the next few daily scrapes, this becomes the new benchmark everyone else has to answer.