Claude Fable 5.1: The World's Most Advanced Model for Coding and Knowledge Work, at a Lower Cost

Claude Fable 5.1: The World's Most Advanced Model for Coding and Knowledge Work, at a Lower Cost

On 1 September 2026 Anthropic released two models: Claude Fable 5.1, generally available, and Claude Mythos 5.1, available only through trusted access programs. The company describes them plainly as the world's most advanced models for coding and knowledge work.

The surprise is that the release is not only about capability. It answers three complaints Anthropic says it heard from customers: price, data retention, and over-sensitive safeguards. Those are exactly the things that matter to anyone running a model in real work rather than in a demo.

The numbers: first place on every published benchmark

These are Anthropic's published results, against its predecessor Fable 5 and its rivals Opus 5 and GPT-5.6 Sol:

• Agentic scientific research (Terminal-Bench-Science): 52.6% against 24.7% for its predecessor, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol — more than double its own previous score.
• Agentic coding (Terminal-Bench 4.0): 55.8% against 42.0% for its predecessor and 52.3% for Opus 5. The Mythos 5.1 variant reaches 60.9%.
• Knowledge work (GDPval-AA): 1853 against 1723 for its predecessor and 1824 for Opus 5.
• Computer use (OSWorld 2.0): 77.9% on the partial measure and 41.7% on the strict one, ahead of the rest.
• Multidisciplinary reasoning (Humanity's Last Exam): 60.9% without tools and 65.0% with them.
• Business workflows (AutomationBench): 31.4% against 17.1% for its predecessor — close to double.

In fairness: these are the vendor's own measurements, published with a stated standard error of 3.5 to 4.5 points on the scientific research benchmark. Independent verification takes weeks. But consistent gains across eight different benchmarks say more than a lead on any single one.

The feature your accountant will notice: cheaper without a price cut

The base token price did not change. Even so, Anthropic says real cost drops by about 25% on typical workloads and up to 45% on step-heavy agentic work, because it cut the price of cache reads by 75%.

For anyone unfamiliar with the distinction: when the model re-reads context it has already processed — your project files, your instructions, the conversation so far — you used to pay a relatively high rate for it. Agentic work re-reads the same context dozens of times within a single task, which is why it benefits most. The discount goes to whoever runs the model seriously, not to whoever asks it a passing question.

Anthropic also added effort levels: low, medium and high. It says Fable 5.1 at low or medium effort matches or beats Fable 5 at much lower cost. The default is high effort in Claude Code, and medium in Claude Cowork and on claude.ai.

Data privacy: the shift that matters to institutions

Anthropic announced a system called Enterprise Frontier Safeguards, giving customers complete privacy equivalent to a zero data retention policy while still preventing adversarial use. It works by storing data in cloud infrastructure the customer controls entirely, not Anthropic. The rollout is phased, beginning this fall; until then, eligible customers can use the model with zero data retention.

That single point opens a door that was closed to sensitive Saudi sectors — government, financial and healthcare bodies whose data-protection obligations stop them sending data to an external service that retains it.

More precise safeguards: fewer refusals of legitimate work

Few things frustrate a developer more than a model refusing a legitimate request because it looked dangerous. Anthropic says the newest safeguards block 60% fewer false positives in cybersecurity. Fable 5.1 can now be used to discover software vulnerabilities, though not to develop exploits for them — a distinction worth keeping: finding a vulnerability is defensive work, weaponising it is not.

What shows this is more than benchmark scores?

Three published results are worth citing because they are concrete outcomes rather than test percentages:

• At the investment firm Millennium, the model found the cause of a rare crash in internal systems that its engineers, and every other model, had failed to explain over several years.
• It trained a neural network that produced a new elevation map of a third of Venus from radar imagery captured by NASA's Magellan mission over thirty years ago, resolving detail down to two or three kilometres instead of ten to twenty, with heights up to 25% more accurate. The map was released under a Creative Commons licence.
• In computational biology, it wrote custom GPU kernels that sped up seven open-source models by as much as 2.5 times with identical outputs, cutting estimated GPU costs by 30 to 60%. Work that would normally take a team of performance engineers weeks was done in days.

The Origami view

For a Saudi business owner the real story is not the leaderboard. It is that one release combines three things: higher capability, lower cost specifically for agentic work, and a privacy option that keeps data in infrastructure you control. Capability alone was already available; cost and privacy were the two things blocking these models from moving out of experiments and into production systems.

Our practical advice: do not start from the model, start from the task. Most of what a mid-sized company needs today runs well at medium effort for less money. Test your task at the lowest level first and raise it only when you can show the need — the difference on a monthly invoice is larger than the price table suggests.

Conclusion

Fable 5.1 leads every benchmark Anthropic published, more than doubles its predecessor on agentic scientific research, and costs less for serious work. More telling is that the release addressed customer complaints rather than capability alone — a sign of a maturing market, where competition has moved from who scores highest to who makes deployment possible inside a real organisation.

Sources

• Anthropic — the official announcement of Claude Fable 5.1 and Claude Mythos 5.1, source of the benchmark table, the pricing, privacy and safeguards claims, and the scientific examples.
• Anthropic — the published note on standard error for Terminal-Bench-Science and the public leaderboard figures.
• The Verge, Decrypt and The New Stack — independent coverage of the 1 September 2026 launch and the pricing figures.