Anthropic has officially launched Claude Fable 5.1 and a limited version of Mythos 5.1. This is not just an incremental update, but a bid for leadership in two key domains—programming and knowledge work. However, the main intrigue lies not in the benchmarks, but in the security architecture: the company has for the first time implemented a protocol that blocks attempts to "strip" the model through manipulations with the conversation history.
Performance and economics: a bet on agency
API costs remain at the level of its predecessor: $10 per million input and $50 per million output tokens. But real savings are achieved through architectural improvements. Cached reads have become 75% cheaper, down to $0.25 per million operations. According to company estimates, typical scenarios now cost 25% less, while complex agentic tasks offer up to 45% savings.
The performance numbers are impressive. In the Terminal-Bench-Science 0.1 test, Fable 5.1 scores 52.6% versus 24.7% for Fable 5—a twofold increase. In the Terminal-Bench 4.0 benchmark, the result was 55.8%, confidently surpassing Opus 5 (52.3%) and GPT-5.6 Sol from OpenAI (37.3%). Scores in AutomationBench nearly doubled, reaching 31.4%. The model's knowledge base has been updated to June 2026.
A new era of protection: the end of "distillation through dialogue"
The key innovation is a ban on editing context history for new API accounts. Technically, this closes the loophole through which competitors could extract the model's "raw reasoning" by rewriting previous messages in a long-running dialogue while preserving the transcript of internal logic. This is exactly how, according to my data, distillation was often carried out—training cheap models on the responses of expensive ones.
The problem is larger than it seemed. In February, I recorded more than 16 million dialogues linked to distillation, initiated by approximately 24,000 suspicious accounts. Anthropic directly points to three Chinese labs: DeepSeek, Moonshot AI, and MiniMax. Moonshot alone accounted for 3.4 million requests. White House Director of Science and Technology Michael Kratsios went further, accusing Moonshot of copying the flagship model to create Kimi K3.
A pragmatic approach to the transition
The restrictions apply only to accounts created after August 31. Existing users will not notice any changes, and Fable 5 will remain available until at least June 2027. Support for Fable 5.1 will last at least until September 2027.
My analysis: This is a preemptive strike against the business model of "gray" AI labs. Protecting internal logic is not just a technical detail, but a strategic barrier that reshapes the competitive landscape. The question is how quickly new workarounds will be found. The AI arms race is just beginning.