Claude Mythos 5.1 and Fable 5.1: Capabilities

Anthropic’s Fable 5.1 lands in a very strange moment: it’s being pitched as a major model release, but it’s arriving beside GPT-6 Astra, which people are also calling the world’s most powerful model. That makes this less a victory lap than a side-by-side stress test. And on the evidence so far, Fable 5.1 looks like a real upgrade, even if the bigger leap may be Astra’s.

The official pitch is straightforward. Anthropic says Fable 5.1 is its best model yet for coding, data analysis, computer use, design, presentations, and long-running agentic work. Boris Cherny says it writes better, has a better tone, and has less Claude-speak. The company also cut cache-read pricing from $1 to $0.25 per million tokens, which lowers effective costs, especially for agentic use. It says benign biology requests now trigger safeguards 85% less often than with Fable 5, and Claude Code users should see about 60% fewer cyber interventions per session.

Users seem to like the change in the model itself. Reports describe Fable 5.1 as more well-rounded, better at writing, more willing to admit mistakes, and better at simplifying code. It also appears more proactive: if you give it a high effort level, it tends to spend those tokens on extra useful work. Some of that may be exactly what people want. Some of it may just mean it won’t sit still.

The benchmark picture is mixed but generally positive. Anthropic’s own results show modest gains in several places, including Terminal-Bench-Science, CursorBench, Humanity’s Last Exam, Chartography, BenchCAD Vision2Code, OSWorld 2.0, OfficeQA, and GDPval-AA v2. But there are regressions too, including in some legal work measures, HealthBench, BioMysteryBench, and a few tool-heavy tests where Fable 5 still holds up better. The pattern is not clean. It’s more like a model that got stronger in a lot of useful ways and occasionally tripped over its own helpfulness.

The bigger practical shift may be pricing and access. Anthropic says Fable 5.1 can be used with zero outside data retention for eligible customers through a system that lets the customer store the data, and full zero data retention is already available for those customers. That matters because Fable 5 reportedly never got above about 11% of Anthropic dollar spend on Ramp, despite being the best model there. If 5.1 is cheaper, less annoying, and easier to adopt, that could change fast. The test now is whether users actually move, or whether they keep whining about headline price while the per-token math quietly does the job.