Anthropic Releases Claude Fable 5.1 to Combat AI Imitation Competitors

Anthropic has released Claude Fable 5.1 with new guardrails designed to prevent model distillation, a technique competitors use to train cheaper systems on expensive AI’s outputs. The move follows documented campaigns where rivals extracted millions of Claude conversations through fake accounts to copy the model’s capabilities.

• Anthropic blocked new API accounts from editing Claude’s reasoning history while preserving prior thinking transcripts, closing a known distillation route.
• Cache read costs dropped 75% to $0.25 per million tokens, making typical workloads 25% cheaper and complex agent tasks up to 45% cheaper.
• Fable 5.1 scored 52.6% on agentic science benchmarks, more than double Fable 5’s 24.7%, while maintaining coding leadership at 55.8%.

• 16M+ Claude exchanges traced to distillation campaigns across roughly 24,000 fake accounts in February
• 3.4M exchanges attributed to Moonshot AI alone among distillation threat actors identified
• 75% reduction in cache read pricing compared to prior Fable model pricing tier
• 52.6% Terminal-Bench-Science score, up from Fable 5’s 24.7% on agentic work

Anthropic introduced Claude Fable 5.1 and its restricted sibling Mythos 5.1 on Tuesday, positioning both as the world’s most advanced models for coding and knowledge work. The releases arrived across Anthropic’s platform, Amazon Bedrock, Google Cloud and Microsoft Foundry simultaneously. The update prioritizes cost efficiency alongside capability gains, with headline per-token pricing holding steady at $10 per million input tokens and $50 per million output tokens.

Blocking the Distillation attack vector

The most consequential change operates quietly: new API accounts can no longer manually rewrite conversation history while preserving Claude’s stored reasoning transcripts. Anthropic explicitly framed this restriction as a defense against model distillation, the practice of training a cheaper model to replicate an expensive one’s behavior by learning from its outputs.

Model distillation has become a widespread concern in the AI industry as development costs for frontier models have climbed into the billions of dollars. Smaller companies and research teams have increasingly turned to extracting outputs from expensive models to train their own systems at a fraction of the cost. This approach, sometimes called “knowledge extraction” or “API scraping,” allows competitors to achieve comparable performance without bearing the computational and research expenses of training from scratch. The technique exploits a fundamental asymmetry: once a model generates an output, that output becomes training data that can be repurposed indefinitely.

The company detailed the scope of the threat in February, having traced over 16 million Claude exchanges to distillation campaigns run through approximately 24,000 fake accounts created by competitors. Three Chinese laboratories appeared in Anthropic’s analysis: DeepSeek, Moonshot AI and MiniMax. Moonshot accounted for 3.4 million of the extracted exchanges alone, representing a systematic effort to harvest Claude’s capabilities at scale.

The concern escalated in July when White House science adviser Michael Kratsios publicly accused Moonshot of copying Anthropic’s flagship model to construct its Kimi K3 system. Moonshot has not issued a public response to the accusation. The timing of the distillation protection announcement may reflect mounting pressure from both the U.S. government and commercial interests seeking to enforce intellectual property boundaries in AI development.

Exemptions and Rollout Timeline for Distillation Protection

The protection applies only to accounts created on or after August 31, leaving older accounts exempt from the restriction. This selective rollout creates a compliance challenge for Anthropic, as determined competitors could potentially continue using legacy API credentials to extract model outputs. The company faces a tradeoff between retroactively restricting existing accounts, which could disrupt legitimate developer workflows, and tolerating a temporary window where distillation remains feasible through older credentials.

Claude Code, Cowork and Claude.ai users see no changes to their experience. Anthropic stated that every account will face the editing check when future models launch, establishing the restriction as a permanent feature of its API rather than a one-time measure. Fable 5.1 will remain available until at least September 1, 2027, giving developers a full year before deprecation and allowing time to migrate to newer systems.

Performance Gains in Agentic and Coding Tasks

Fable 5.1 delivered its sharpest performance improvement in agentic science work, reaching 52.6% on Terminal-Bench-Science 0.1, a score more than double Fable 5’s 24.7%. Agentic capabilities represent an emerging frontier in AI development, where models must plan and execute multi-step tasks with minimal human intervention. The doubling of agentic science performance suggests that Anthropic has made fundamental architectural improvements to reasoning and task decomposition.

Coding performance widened its lead over competitors, with Fable 5.1 achieving 55.8% on Terminal-Bench 4.0, compared to 52.3% for Opus 5 and 37.3% for OpenAI’s GPT-5.6 Sol. Business workflow automation capabilities nearly doubled to 31.4% on AutomationBench, signaling broader strength across enterprise use cases. Software development remains the most commercially valuable application domain for large language models, and maintaining a significant performance edge in coding translates directly to enterprise adoption.

Knowledge cutoff moved forward five months to June 2026, keeping the model’s training data more current than prior releases. This refresh improves the model’s ability to answer questions about recent events and developments while maintaining consistency with the broader training approach.

The 75% reduction in cache read costs creates material savings for developers running repeated or long-context queries, addressing a pain point that has limited adoption of extended reasoning and context windows. The distillation protections target a specific vulnerability that has already cost Anthropic millions in copied exchanges, though the restriction’s applicability only to new accounts leaves a window for rival labs to continue extracting from older API credentials.

BlockWest is a news publication. Nothing here is investment advice. Read our disclaimer and editorial policy.