On September 1st, Anthropic introduced Claude Fable 5.1 — an updated version of the Fable 5 model — along with a parallel variant, Claude Mythos 5.1, featuring a different set of constraints. Both models are built on the same architecture but are intended for different scenarios: Fable 5.1 is aimed at complex tasks in coding, intellectual work, and extended agentic processes, while Mythos 5.1 is available only to participants in trusted access programs.
Improved Performance on Benchmarks
According to the company, Fable 5.1 demonstrated a noticeable performance improvement on agentic and scientific tests compared to its predecessor and the competing model GPT-5.6 Sol:
Terminal-Bench-Science 0.1 — 52.6% vs. 24.7% for Fable 5 and 22.4% for GPT-5.6 Sol
Terminal-Bench 4.0 — 55.8% for Fable 5.1 and 60.9% for Mythos 5.1 vs. 42.0% for Fable 5 and 37.3% for GPT-5.6 Sol
GDPval-AA v2 — 1853 points vs. 1723 for Fable 5
The base cost for using the model remains unchanged — $10 per million input tokens and $50 per million output tokens. However, Anthropic reduced the price for cache reads to $0.25 per million tokens, which, according to the company's estimates, saves about 25% on typical workloads and up to ~45% on agentic scenarios. The model's context window reaches 1M tokens, and the maximum response length is limited to 128K tokens.
Venus Map as a Demonstration of Capabilities
In the announcement, Anthropic provided a telling example of the model's use in science: Fable 5.1 trained a neural network and constructed a new topographic map of approximately one-third of Venus's surface. The work was based on over-30-year-old radar data from NASA's Magellan mission and an existing partial map of the planet. As a result, the resolution improved from 10–20 km to 2–3 km, and the accuracy of elevation measurements increased to 25%. The company released the finished map under a Creative Commons license.
System Prompt Leaked into the Open
Almost immediately after the release, Fable 5.1's system prompt — the document that defines the model's behavior rules, memory handling, tool usage, artifacts, computer use, and request routing — surfaced online. According to September 2nd posts on social media platform X, its size exceeds 270,000 characters. The file was extracted and published in a GitHub repository (Anthropic section, file claude-fable-5.1.md, updated September 1st). One example discussion of the find was posted by a user on X.
Where the Model is Available
Fable 5.1 can be invoked via the Claude API under the identifier claude-fable-5-1, as well as through Amazon Bedrock, Google Cloud, and Microsoft Foundry. Mythos 5.1 remains a restricted version — access is only open to participants in trusted access programs.
This release is the second major update to the Fable and Mythos line since their initial launch in June 2026. The main changes affected pricing, the speed of solving long agentic tasks, and accuracy on scientific benchmarks, while the core model architecture remained the same.
AI Perspective
From a machine data analysis perspective, the story with Fable and Mythos writes the second episode of regulatory risk this year: in June 2026, the U.S. Department of Commerce temporarily restricted access to these models, citing national security risks, and only lifted the ban three weeks later. The technical gains on benchmarks and the almost simultaneous leak of the system prompt exist within the same context: the more powerful and "chatty" a model becomes, the closer the scrutiny from regulators and security researchers on its internal architecture. On a macroeconomic level, this cycle of "ban–lift" adds an element of uncertainty for corporate clients choosing an AI infrastructure provider, comparable to sanction risks in the semiconductor industry. Will such turbulence remain a one-off episode or become a regular practice for the most powerful AI models?
