AI GUIDEPartner content

Claude Fable 5.1: twice as good at science and a quarter cheaper

A breakdown of the Claude Fable 5.1 update: benchmark figures versus Fable 5 and Opus 5, cheaper cache reads, and practical effort-level settings.

Affiliate link: your price stays the same and the project earns a commission.

On September first, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. By version number, it is a cosmetic update; by results, not at all: on a scientific benchmark, the model doubled its previous score and beat the more powerful Opus 5, while the cost of typical work fell by about a quarter with base prices unchanged. We break down what lies behind the figures and how to use it.

Two models with one engine

Claude Fable 5.1 and Mythos 5.1 are the same model. They differ only in their safety restrictions. Fable 5.1 is available to everyone and runs on AWS, Google Cloud, and Azure. Mythos 5.1 is provided to vetted cybersecurity and life-science professionals through verification programs, and for now only to organizations in the United States: some of its restrictions that hinder research on dangerous topics have been removed.

The difference in figures is visible in coding. On Terminal-Bench 4.0, Fable 5.1 scores 55,8%, while its unrestricted twin scores 60,9%. Five points—and that is the price of the safety settings.

What the tests showed

Claude Fable 5.1 achieved its main result on Terminal-Bench-Science 0.1. This consists of 70 real-world scenarios assembled by practicing scientists: tasks from the life sciences, physics, Earth sciences, mathematics, and engineering, where the model works step by step like a researcher at a terminal.

TestFable 5Fable 5.1Competitors
Terminal-Bench-Science 0.124,7%52,6%Opus 5 — 29,0%, GPT-5.6 Sol — 22,4%
Terminal-Bench 4.0 (coding)42,0%55,8%Opus 5 — 52,3%, Mythos 5.1 — 60,9%
AutomationBench (business processes)17,1%31,4%—
Humanity's Last Exam, without tools57,8%60,9%with tools 65,0%
CursorBench 3.2.070,5%73,4%—

Notice the range of improvements. Where the task requires completing a long piece of work—in science and process automation—the increase is several-fold. Where the model needs to answer a difficult question here and now, the gain is modest: three points on Humanity's Last Exam, and less than three on CursorBench. The model has become smarter not so much in reasoning as in endurance.

Why it became cheaper

Token prices have not changed: 10 dollars per million input tokens, 50 per million output tokens. Cache reads have become cheaper, now costing 0,25 dollars per million, four times less than before. For ordinary workloads, the savings amounted to around 25%; for agentic scenarios with constant context rereading, they reach 45%.

The savings are not abstract; they are directly tied to how an agent works. It runs the same large context around in a loop dozens of times, and previously each such request cost a noticeable amount of money. The model has a context window of one million tokens and can produce up to 128 thousand in output.

Effort levels: the main practical lever

The model has five modes: low, medium, high, xhigh, and max. The cost increases with the level, and this is where it is easiest to waste money for no reason. The habit of setting the maximum for every task consumes limits without any return.

The working logic is simple. Everyday tasks such as analyzing a document, drafting an email, or editing code from a description are handled perfectly well on low and medium. The higher modes are for genuinely difficult tasks: lengthy debugging with an unclear cause, research involving a dozen steps, or analysis of a large body of data. Test your typical tasks to see where the boundary lies. Most likely, it is much lower than the level you are currently setting.

Working unsupervised

Anthropic emphasizes separately that the model behaves more calmly during long tasks without a person nearby: it keeps its own records of what it has done and reprioritizes when circumstances change. This is where previous generations broke down. After several hours of work, an agent would lose the thread and start going around in circles.

In practice, this changes how tasks should be assigned. Previously, work had to be cut into small pieces and the results stitched together manually. Now it is more effective to hand over the entire project, with all its context, and retrieve the result together with a report on what was done.

Science as an indicator of a change in task class

The two examples Anthropic gives as demonstrations are worth noting even if you are far from research.

The first concerns Venus. The model trained a small neural network on radar images from NASA's Magellan mission, collected more than thirty years ago, and built an elevation map for roughly a third of the planet. Resolution improved from 10–20 kilometers to 2–3, with elevation accuracy improving by about a quarter. The map is being released freely so that future missions can choose landing sites.

The second concerns protein design. The model designed binding proteins for 12 targets and was successful in about half of the cases, whereas 10–15% is considered normal in this field. For three targets—EGFR, Nipah G, and 15-PGDH—the affinity was ten times higher than that of the best work submitted to Adaptyv Bio competitions.

This is where the difference between “answering a question” and “conducting research and delivering a result” lies. The tool has stopped being a correspondence assistant and become an executor of work that previously required weeks from a specialist.

Who needs it and who does not

  • Developers — yes, especially for debugging. The model looks for the cause of a failure instead of covering up the symptom.
  • Analysts and researchers — yes. Long documents, spreadsheets, presentations, and multistep analyses.
  • Those building agents — yes, for endurance on long tasks and cheaper cache reads.
  • For short chats and minor questions — no. A simpler model will be enough here, and it will cost less. Choosing the right tool for the task is a separate skill that saves more than any settings.

What to do in practice

  1. Rework your default effort levels. Set low or medium as the default, and enable the higher modes deliberately.
  2. Give it entire tasks. The model performs best on large, complicated projects, not on a chain of small questions.
  3. Ask it to check its own work. Requiring it to run the code it has written before submitting it noticeably improves the result and costs almost nothing.
  4. Take advantage of the fact that it has started writing more simply. The direct request “shorter and more human” now actually works instead of being ignored.
  5. If you have an agentic workload, recalculate the budget. Cheaper cache reads change the economics of scenarios that previously did not pay off.

Frequently asked questions

How does Claude Fable 5.1 differ from Mythos 5.1?

The engine is the same; the safety restrictions differ. Mythos is available only to vetted cybersecurity and life-science professionals from organizations in the United States, and its coding result is five points higher.

Is it worth switching from Fable 5?

If you have long-running agentic tasks or research work, the switch pays off immediately: performance on specialized tests doubles, while cache reads have become four times cheaper. For short requests, the difference is barely noticeable.

Which effort level should I choose by default?

Start with medium and increase it only on tasks where the result is clearly worse than expected. The maximum mode is justified for lengthy debugging and multistep research; for everything else, it simply uses up your limits faster.

The main takeaway from this release is not the specific percentages. The improvement occurred where the model works for a long time on its own—and that is the class of tasks that was still considered inherently unreliable just a year ago.

Compare models before you start

The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.

Browse models

Affiliate link: your price stays the same and the project earns a commission.

claude fable 5.1 claude mythos 5.1 anthropic terminal-bench science claude effort levels claude cache reads claude benchmarks agentic tasks

SEO Mind42 editorial team

We explore SEO and neural networks in practice: test services on our own projects, verify prices and limits against primary sources, and share things you can put to use the same day.

📚 Reference guide to SEO and AI 🔄 Materials are updated 🕐 Updated: 4 October 2026

Related reading

All in this section →