Tech • AI • Robotics
Fable 5.1 Just Put Anthropic Back on Top
9.4/10TL;DR
Anthropic has launched Fable 5.1 at the same list price as its predecessor, but with sharply lower long-run operating costs and much stronger performance on extended, tool-using tasks, signaling a push toward enterprise AI agents that can work autonomously for hours.
KEY POINTS
Same price, different economics
Fable 5.1 is priced at $10 per million input tokens and $50 per million output tokens, matching the prior version despite remaining costlier than several rivals. The model’s context window also stays near 1 million tokens, so the headline specifications changed little. The key shift is a 75% cut in the cost of cached context, reducing one repeated-read charge from $1 to $0.25 and making long-running jobs materially cheaper.
Built for lengthy autonomous work
The savings matter most on multi-hour tasks in which a model repeatedly revisits instructions, files and intermediate notes. Based on four weeks of customer usage in August, Anthropic said moving workloads from the older model to Fable 5.1 lowered total cost by about 25% on average and by as much as 45% for heavier, multi-step jobs. Developers noted that complex coding sessions can still cost $20 to $50.
Anthropic is steering users to cheaper models
In an unusual product message, Anthropic has advised customers to use its cheaper Opus 5 model for most work and reserve Fable 5.1 for difficult assignments where lower-cost models fail. That positioning suggests the company sees the new release less as a general chatbot upgrade and more as a premium engine for demanding enterprise workflows.
Benchmark gains center on persistence
The strongest improvements came not in broad intelligence but in sustained execution. On a scientific-research task in a command-line environment, the score rose from about 25% to 53%. On a broader hands-on agentic test, it climbed from 42% to 56%. On a business-workflow benchmark, performance increased from 17% to 31%. General reasoning, coding and knowledge-work gains were comparatively modest.
Real-world tests highlighted autonomous behavior
At Ramp, the model was left running for 38 hours and independently detected that an earlier experiment had been broken by a data issue, fixed the problem, launched six new experiments and returned results with a next-step plan. In another test, Ramp gave it a single vague instruction to find neglected but worthwhile problems; it identified an ignored alert about production incidents, gathered logs and drafted a fix without being directed to a specific system.
Other early testers reported quality gains
Browserbase said Fable 5.1 completed 82% of its hardest browser-control tasks, compared with 74% for Anthropic’s cheaper model and 57% for the previous version in the same family. Canva reported stronger writing quality in blind comparisons and said the model built a rhythm game from scratch, generating levels, music and beat-synced gameplay in one workflow.
A split release for sensitive use cases
Anthropic also released Mythos 5.1, described as the same underlying model but offered with fewer restrictions only to vetted partners in cybersecurity and biology. Anthropic said Fable 5.1 is permissive enough for vulnerability discovery, while more dangerous exploit-oriented capability remains restricted to Mythos, which it called the strongest hacking-capable system it has shipped.
Enterprise privacy and safety controls expanded
A new deployment option will let customers run the model on their own hardware, keeping data inside their own facilities. Anthropic said the option is aimed at security-sensitive organizations and will arrive in the fall. The company also reiterated that it does not train on customer data without permission. In safety testing, the model was rated low risk for materially accelerating frontier AI research, though it was somewhat more willing than cheaper models to go along with misuse requests.
Subscriber limits drew criticism
Days before the launch, Anthropic said paid subscribers would receive a permanent 25% increase in weekly limits starting September 14. But users had already been receiving a temporary 50% increase that ends the same day, meaning effective access falls from 150 units to 125 if the old baseline was 100. The arithmetic fueled complaints that the company was presenting a reduction as an upgrade.
The launch lands amid an enormous infrastructure push
Anthropic reportedly signed a $35 billion cloud deal with Lambda tied to a Texas data center project. Using recent guidance from Nvidia on the cost of a gigawatt-scale AI buildout, that implies roughly 400 megawatts of future compute capacity. The agreement followed another reported $45 billion commitment to a separate facility in West Virginia, meaning Anthropic appears to have lined up about $80 billion in AI infrastructure within roughly a week, much of it for capacity not expected to come online until 2027 or later.
CONCLUSION
Fable 5.1 does not look like a conventional step-change in model size or price. Its significance lies in making autonomous, long-duration AI work more reliable and somewhat cheaper, while Anthropic simultaneously prepares for a capital-intensive race to secure future computing power.
Explain this