[AINews] Claude Fable/Mythos 5.1: new SOTA model, 75% cache price cut but 70% more output tokens

Anthropic’s latest Claude release is being sold as more than a benchmark trophy. Fable 5.1 and Mythos 5.1 are the new flagship models for coding and knowledge work, with the company pitching Fable as the one for difficult, long-horizon tasks that can run on its own. That framing matters, because the old complaint about Fable 5 was never that it was dumb. It was that it was brilliant in a datacenter and awkward everywhere else.

The pricing picture is mixed in a very Claude way. Input, output, and cache-write prices stay at $10, $50, and $12.5 per million tokens. Cache reads, though, drop from $1.00 to $0.25 per million tokens — a 75% cut that should be a real win for long-context and agentic workflows. The catch is output usage. Artificial Analysis says 5.1 is using about 1.7 times as many output tokens, which pushes the net cost per task up by 20%.

On the numbers, the model looks formidable. Artificial Analysis puts Fable 5.1 at 66 on its Intelligence Index, ahead of Claude Opus 5 max at 63, Claude Fable 5 max at 62, GPT-5.6 Sol max at 61, and Grok 4.6 high at 61. It also reports 59.1% on Humanity’s Last Exam, 91.4% on Terminal-Bench v2.1, 62.0% on SciCode, and a 9-point gain on τ³-Banking over Fable 5. There’s a caveat, though: AA says its eval included server-side fallback, and about 4% of output tokens were served by fallback models.

The release also sharpened a weirdly important question: are Fable 5.1 and Mythos 5.1 actually different models, or just different safety and routing paths wrapped around the same weights? That idea came from community analysis, not Anthropic itself, but it fit the broader confusion around the launch. Some users loved the coding and planning gains. Others complained about rate limits, safeguard false positives, and a subscription experience that still feels like it was designed by a lawyer with a stopwatch.

And that may be the real story here. Anthropic has clearly pushed Claude closer to a usable autonomous worker, not just a clever chatbot. But the better it gets at long, delegated work, the more the product experience seems to fight it.