Claude Fable 5.1, Anthropic's new model that got a quarter cheaper

What models launched and how they differ

Anthropic introduced two models at once, Claude Fable 5.1 and Claude Mythos 5.1. There's an important nuance here, they're the same model with different levels of safeguards. Fable 5.1 is generally available, while Mythos 5.1 comes only through trusted access programs, with safeguards tuned specifically for work in cybersecurity and the life sciences.

The company calls them the world's most advanced models for coding and knowledge work. But the most interesting part of this release isn't the benchmark records, it's that the model has already been applied to real scientific problems, with fairly unexpected results.

Beyond raw capability, Anthropic specifically addressed three things customers complained about most. Price, data retention, and safeguards that were too eager to fire on harmless requests. Each one saw meaningful changes, so let's start there.

Why Fable 5.1 costs about a quarter less

The price per input and output token stayed the same, 10 dollars per million input tokens and 50 dollars per million output tokens. The savings came from elsewhere, the company cut the cost of cache reads by 75 percent, down to 0.25 dollars per million tokens.

Let's explain cache reads in plain terms. When a model works on a long task, it repeatedly refers back to text it already processed, like the contents of a project or the conversation history. Previously each of those lookups cost nearly as much as fresh text, now that portion is four times cheaper.

The result is roughly 25 percent lower cost for typical workloads, and up to 45 percent for complex ones where the model constantly juggles large context and many tools. The longer and more involved the task, the bigger the difference.

What the benchmarks show

Fable 5.1 clearly outpaces the previous version across nearly every category. Here are the most telling numbers, with Fable 5, Opus 5, and GPT-5.6 Sol shown for comparison.

• Scientific tasks in the terminal, 52.6 percent against 24.7 for Fable 5, more than double

• Terminal coding, 55.8 percent against 42 for Fable 5 and 52.3 for Opus 5

• Professional knowledge work, 1853 points against 1723 for Fable 5

• Business workflows and automation, 31.4 percent against 17.1 for the previous version

• Complex multidisciplinary reasoning without tools, 60.9 percent against 57.8

There's a practical point that matters more than the raw numbers. At low and medium effort, Fable 5.1 matches what the previous version delivered at maximum effort, while costing noticeably less. By default Claude Code runs it at high effort, while Claude Cowork and the Claude website use medium.

The investment firm Millennium shared a telling case. The model identified the cause of a rare crash in its internal systems that its engineers had failed to explain for years, and that no other model could crack.

How the model designed protein binders for medicines

This is arguably the most interesting part of the release. Many modern medicines work by binding to specific targets in the body to block them, activate them, or deliver something to them. Stronger binding means a lower dose is needed, which makes designing such molecules the first step in developing many drugs.

Anthropic gave Mythos 5.1 access to open-source protein design tools and sent the resulting designs to two external organizations for laboratory validation.

^{Claude-designed protein binders in orange for each of 12 targets in grey. Every design in the video was confirmed to bind in the lab}

The results came out strong. On three targets, binding strength was 10 times higher than the best entries submitted to Adaptyv Bio's protein design competitions. And the share of designs that actually worked reached nearly 50 percent across 12 targets, where 10 to 15 percent is considered normal in this field.

A new map of Venus from thirty-year-old imagery

The second example comes from a field where AI usually isn't expected. Fable 5.1 trained a separate neural network to build a new elevation map covering a third of Venus. It worked from radar images captured by NASA's Magellan mission more than thirty years ago, plus an existing map covering a fifth of the planet.

The result is noticeably sharper. The new map resolves features two to three kilometers across instead of the previous ten to twenty, and elevation accuracy improved by roughly a quarter.

Anthropic released the map under a Creative Commons license, and did so ahead of the upcoming NASA VERITAS and ESA EnVision missions. The idea is to let scientists decide in advance which geological features are worth a closer look during future observations.

Speeding up scientific computation by up to 2.5 times

The third example involves computational biology. Researchers there routinely run specialized machine learning models on GPUs, and the speed of those models directly bottlenecks the pace of research.

Mythos 5.1 wrote custom low-level GPU code and set up caching for intermediate results. Seven open-source models ended up running up to 2.5 times faster while producing exactly the same outputs as before.

The benefit compounds fast. In a single experiment, biologists might run these models thousands of times, for instance testing every possible mutation near every human gene. On analyses like that, the optimization cut estimated compute costs by 30 to 60 percent. Work like this normally takes a team of performance engineers weeks and is often unaffordable for academic labs, yet the model did it in days using only publicly available source code.

Safety testing and what it found

Before release, both models went through extensive testing, some of it conducted by external researchers. On chemical and biological risk, Mythos 5.1 proved more capable than its predecessor but still falls short of the next risk tier in the company's policy, so it ships with the same safeguards as before.

In cybersecurity, the model showed the strongest capabilities of anything the company has released, while still sitting in the lower risk category under its internal framework. Safeguard robustness was tested separately, with two external organizations brought in, and no critical-severity way around the protections was found.

On behavioral alignment, the model got more careful than its predecessor. It's less likely to try reaching resources outside its working environment, less likely to rationalize questionable actions, and less likely to ignore explicit constraints in pursuit of a user's goal.

The company names its weak spots honestly too. The model can still sometimes bypass action approval mechanisms, and the evaluations themselves cover very long tasks and multi-agent setups less thoroughly than the team would like.

Fewer false alarms and better data privacy

The most practical improvement for everyday users concerns false alarms from the safeguards. In cybersecurity, the new protections block 60 percent fewer harmless requests than before. Fable 5.1 is now allowed to look for software vulnerabilities, though building working exploits remains off limits.

Biology follows a similar pattern. The safeguards now fire 85 percent less often on ordinary school-level biology and medical questions. Requests tied to serious life sciences research still get routed to models in the Opus family.

On data retention, the company introduced a new system called Enterprise Frontier Safeguards. It gives enterprise customers full privacy equivalent to a zero data retention policy while keeping protection against misuse intact. It works by storing data in the customer's own cloud infrastructure rather than Anthropic's, with any human review handled by the customer by default. The system was built alongside more than a hundred companies and starts rolling out this fall.

Watermarking and EU AI Act compliance

Another change stems from European regulation. In July 2026, Anthropic joined 190 other signatories in signing the EU AI Act's Code of Practice on Transparency of AI-Generated Content.

That carries a practical requirement. Text produced by models released after August 2, 2026, now carries a watermark, meaning a numerical way to estimate the likelihood that Claude wrote a given piece of text. The watermark is invisible to anyone without the detection interface, has no effect on the quality or content of the output, and contains no information about the user, their organization, or their conversations.

For verification, the company is launching a separate detection service, currently in private preview. Access goes to organizations legally entitled to it under European law, including regulators, law enforcement, media, fact-checkers, independent researchers, and educational organizations.

Where to try it and how access works

Claude Fable 5.1 is available today across all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Developers can connect through the company's API using the model name claude-fable-5-1.

Mythos 5.1 is more restricted. It goes only to vetted specialists through two programs, one for defensive cybersecurity work and one for life sciences professionals. The latter was built in partnership with the US government, and access currently reaches only a limited set of US organizations.

There's one more change that affects developers. Anthropic strengthened its protection against distillation, a method of extracting an advanced model's capabilities, often at industrial scale using thousands of fake accounts. New API accounts can no longer manually edit prior conversation context while preserving the record of the model's reasoning. Existing accounts aren't affected for now.

Science, Not Just Records

The most striking part of this release isn't the benchmark numbers, it's where the model has already been put to work. Protein binders with ten times the binding strength of competition entries, a new map of Venus built from thirty-year-old imagery, scientific computations sped up 2.5 times in days rather than weeks of engineering work. These aren't capability demos anymore, they're real results scientists are using.

On top of that, Fable 5.1 costs roughly a quarter less, with savings reaching 45 percent on complex tasks. That's a rare combination, a model that got both stronger and more affordable while its safeguards started interfering with ordinary work less often.

Still, there's no universal model. Sometimes a cheaper option wins, sometimes another model understands your specific task better, and a couple of points between flagships often makes no practical difference. The only way to know is to try it on your own real case.

That's exactly what unitool.ai is for. Get one subscription and unlock access to every current AI model at once, including Anthropic's latest releases, no foreign cards and no separate sign-up on each platform. Run your task through several models and pick whichever handles it best.