Anthropic made an unexpected but logical move on Tuesday by releasing two new models at once — Claude Fable 5.1 and a limited version of Mythos 5.1. This is not just an iterative update with improved benchmarks. The key feature of the release is a fundamental change in security policy that will hit unscrupulous competitors using distillation to copy others' work.
The main innovation concerns protecting the model's internal logic. Now, new developer API accounts physically cannot rewrite dialogue history to preserve Claude's original "thinking." Previously, this loophole allowed attackers to extract the model's "reasoning" and use it to train their own, cheaper algorithms. Anthropic directly links this change to competitive struggle, emphasizing that the restriction is aimed at other companies.
Price and performance: a bet on savings
The cost of basic tariffs remains unchanged: $10 per million input tokens and $50 per million output tokens. However, savings are now built into the architecture. Cache reads have become 75% cheaper — down to $0.25 per million operations. According to the company's estimates, typical tasks now cost 25% less, and complex scenarios for AI agents save up to 45%.
Test results are impressive. In Anthropic's own ranking, the biggest leap was shown by the "scientific" work of AI agents: Fable 5.1 scored 52.6% on the Terminal-Bench-Science 0.1 test, while Fable 5 had this figure at 24.7%. In programming, the advantage is slightly more modest: 55.8% on Terminal-Bench 4.0 versus 52.3% for Opus 5 and 37.3% for GPT-5.6 Sol from OpenAI. In business process automation tasks, scores nearly doubled, reaching 31.4% on AutomationBench. The model's knowledge base now includes data up to June 2026 — five months fresher than its predecessor.
War on distillation: names named
This step is a response to a systemic problem. Back in February, Anthropic tracked over 16 million dialogues related to distillation through approximately 24,000 fake accounts. The company directly named three Chinese laboratories: DeepSeek, Moonshot AI, and MiniMax. Moonshot alone accounted for 3.4 million requests. Later, in July, White House science advisor Michael Kratsios publicly accused Moonshot of copying Anthropic's flagship model to create its own Kimi K3.
The restriction applies only to accounts created after August 31. Existing users of Claude Code, Cowork, and Claude.ai will not notice any changes. Fable 5 will remain available until at least June 2027, and support for Fable 5.1 will last at least until September 1, 2027. Anthropic promises that in future versions, all accounts without exception will undergo verification.
My view: This is a precedent-setting decision. Protection against distillation at the API level is not just a technical patch, but a signal to the market that the AI arms race is moving into the legal and infrastructural arena. For investors and developers, this means that "cheap clones" of leading models will soon lose their main advantage, and companies with their own research cycles will gain an additional barrier to competition.