Claude Fable 5.1 and Mythos 5.1 launch with lower cache costs and fewer restrictions
Anthropic has launched Claude Fable 5.1 and Claude Mythos 5.1, just about 3 months after the original Fable 5 release. The update improves speed and benchmark performance while cutting cache read pricing by 75% from the previous $0.25 per million tokens. Since cache reads let Claude reuse previously processed context, the reduction is especially relevant for coding agents and other long running workloads. Anthropic estimates Fable 5.1 is around 25% cheaper than Fable 5 for an average workload, with savings reaching up to 45% for tasks that rely more heavily on cached context.
Fable 5.1 is faster and improves across several coding and terminal benchmarks, scoring 55.8% on Terminal Bench 4.0, up from 42%, 73.4% on CursorBench 3.2.0, up from 70.5%, and 52.6% on Terminal Bench Science, up from 24.7%.
Following criticism of Fable 5's restrictions, Fable 5.1 can now identify software vulnerabilities for defensive research and analysis. Anthropic expects around 60% fewer safeguard interventions during an average Claude Code session. Fable 5.1 is available through Claude, the Claude API, Amazon Web Services, Google Cloud, and Microsoft Azure under the claude-fable-5-1 model ID.
Comments
Open weight model that have no restriction are just taking over frontier closed models becasue they let security expert do their job!
Large open weight models do have restrictions mostly because the literacy around sensible subjects (like building a nuclear bomb or a bioweapon) is much sparse than mainstream ones, or are cyclic reasoning like conspiracy theories, which is greatly reducing the quality of answers.
Some projects like heretic-project.org can reduce the refusal rate but models are rarely becoming relevant on sensible subjects. It's just a long list of instructions to suppress model own rules to refuse some prompt, it doesn't add any new information.
Most large model companies are now trying to AI-generate more content to feed their model (even Anthropic is stopping scanning hardcover books because it cannot keep up with the giant amount of new tokens need to train multi trillion parameters models), which could lead to "model collapse" (i.e. greatly reducing the overall quality) as the researchers proved, but could also improve some very specific fields of knowledge by synthesizing more content about it.