Anthropic has released Claude Fable 5.1 and Claude Mythos 5.1, around three months after Fable 5 and Mythos 5 first appeared.
Fable 5.1 is the version most people will encounter. Mythos 5.1 uses the same underlying model with more permissive safeguards for approved cybersecurity and life sciences work. Access to Mythos is currently limited to vetted individuals and organizations.
The 5.1 update focuses heavily on complicated work that can take many steps to complete. Anthropic reports gains in coding, scientific research, business workflows, and computer use, alongside changes to pricing, privacy, and safety controls.
Fable 5 arrived in June with an emphasis on long-running tasks. Anthropic designed it to keep working through large coding projects, research assignments, and other jobs that might require many actions before reaching a result.
Fable 5.1 improves performance across several of those areas. Anthropic also says it can reach similar or better results than Fable 5 at lower "effort" settings in some tests. Effort controls how much processing the model uses before producing an answer. Lower effort can reduce the time and cost involved in completing a task.
Several early-access customers reported improvements in long tasks as well. Jane Street said the model remained easier to follow as assignments became more complicated, while Red Hat reported more concise updates during coding work. Shopify said it was able to run for long periods while keeping track of its own work and adjusting priorities.
These are customer assessments published by Anthropic, so they provide useful examples rather than independent proof of performance.
Anthropic has also adjusted some of the safeguards introduced with Fable 5. The original model sometimes redirected harmless cybersecurity or biology questions because its safety systems were deliberately cautious.
The company says its updated biology safeguards now trigger 85% less often on benign elementary biology and medical questions. Claude Code users should also see around 60% fewer cybersecurity interventions per session.
Fable 5.1 can now help identify software vulnerabilities for defensive security work. Some activities, including exploit generation and penetration testing, continue to be handled by other models or restricted.
Anthropic's comparison chart contains several different tests, or benchmarks. Each benchmark gives the models a particular type of task and measures how well they perform.
Source: https://www.anthropic.com/claude-fable-and-mythos-5-1
The easiest way to read the table is row by row. Compare the models within the same row because each test uses its own scoring method.
On Terminal-Bench-Science, which tests how well an AI can work through scientific research tasks using tools, Fable 5.1 scored 52.6%. Fable 5 scored 24.7%, Opus 5 scored 29.0%, and GPT-5.6 Sol scored 22.4% in Anthropic's evaluation. Anthropic also reports a margin of error of roughly 3.5 to 4.5 percentage points for each model on this test.
For AutomationBench, Fable 5.1 scored 31.4%, compared with 17.1% for Fable 5, 26.9% for Opus 5, and 19.6% for GPT-5.6 Sol. This benchmark focuses on business workflows where the AI has to complete several connected steps. A simple example might involve finding information, updating a system, and producing a final result without a person guiding every individual action.
The GDPval-AA v2 row looks different because it uses points. Fable 5.1 received a score of 1,853, ahead of Fable 5 at 1,723, Opus 5 at 1,824, and GPT-5.6 Sol at 1,711. This test is designed around professional knowledge work. The point total only makes sense within that benchmark, so a score of 1,853 cannot be compared directly with a percentage elsewhere in the table.
OSWorld 2.0 tests computer use. The model has to interact with software and work through realistic tasks on a computer. Anthropic shows two scores for Fable 5.1: 77.9% partial and 41.7% strict.
Partial scoring gives credit when the AI completes some of the required steps. Strict scoring requires full completion under the benchmark's rules. The difference between those two numbers gives a useful picture of current computer-using AI. The model can often make substantial progress through a task, while complete success remains less consistent.
Another row tests multidisciplinary reasoning using Humanity's Last Exam. Fable 5.1 scored 60.9% without tools and 65.0% with tools. In this context, tools can include things such as search, code execution, or other systems the model can use while solving a problem.
Across the results Anthropic published, Fable 5.1 scored above the other generally available models shown in each reported comparison. Benchmark results still depend on the specific test, settings, and tools being used, so they give a useful comparison rather than a complete ranking of which model will perform best for every job.
Anthropic says Fable 5.1 and Mythos 5.1 are identical at the model level. Mythos has access to capabilities that are restricted for general users because of the potential risks involved in advanced cybersecurity and life sciences work.
Anthropic has created separate access programs for those areas. Its Cyber Verification Program is intended for vetted security professionals, while its Life Sciences Verification Program provides approved researchers with access to more advanced biology capabilities. Mythos 5.1 is currently available only to a set of US organizations, with broader access planned.
The scientific examples Anthropic published give some indication of why those capabilities are being treated separately.
In protein design tests, Mythos 5.1 created candidate molecules that were sent to external organizations for laboratory testing. Anthropic says nearly half of its designs successfully bound to their intended targets across a group of 12 targets.
Fable 5.1 was also used to create a higher-resolution elevation map covering about a third of Venus using radar data from NASA's Magellan mission. According to Anthropic, the resulting map can show details at a scale of around two to three kilometers, compared with 10 to 20 kilometers in the older elevation data.
In another experiment, Mythos 5.1 optimized seven open-source machine-learning models used in biology. Anthropic reports speed improvements of up to 2.5 times and estimated computing-cost reductions of 30% to 60% for some large analyses.
These are results reported by Anthropic and its collaborators. They show the kinds of research tasks being tested with the new models, while broader independent testing will give a clearer picture over time.
Fable 5.1 keeps the same main API price as Fable 5: $10 per million input tokens and $50 per million output tokens. If you only look at those numbers, the price appears unchanged. The savings come from how Claude handles information it has already processed.
During a long task, Claude may need to look at the same instructions, documents, or previous work many times. Anthropic can store some of that information temporarily in a cache, which is essentially a reusable copy of material Claude has already read.
With Fable 5, reading that cached information cost more. For Fable 5.1, Anthropic has cut the price of those cache reads by 75%, to $0.25 per million tokens.
Imagine asking Claude to work with a 100-page company document throughout a project. It may need to refer back to that document repeatedly. The first time it processes the information, normal input pricing applies. When it can reuse the stored version later, those repeated reads are much cheaper.
That difference becomes more noticeable on long coding projects, research assignments, and AI-agent workflows because the model may repeatedly refer to the same information while completing dozens or hundreds of steps.
Anthropic estimates that this change reduces the total cost of a typical Fable 5.1 workload by around 25% compared with Fable 5. For complex coding and highly autonomous tasks, it estimates savings of up to around 45%.
There is a second factor as well. Anthropic says Fable 5.1 can achieve results similar to or better than Fable 5 at lower effort settings on some tasks. Lower effort means Claude uses less processing to reach the result, which can reduce the amount spent on the task even though the published price per token has not changed.
Fable 5.1 also falls under Anthropic's new text-watermarking requirements for models released after August 2, 2026. The watermark is invisible in normal writing and is designed to allow authorized detection systems to estimate whether Claude was involved in creating a piece of text.
For people already using Fable 5, the 5.1 release brings improvements in several of the tasks the model was originally designed for, along with cheaper reuse of previously processed information and fewer unnecessary safety interruptions. Mythos 5.1 extends the same underlying capabilities into specialist research areas where Anthropic continues to control access.
© 2026 SVA Consulting