Anthropic Launches Claude Fable 5.1 and Mythos 5.1 for Scientific Research

Upon waking up, Anthropic once again rewrote the coordinate system of machine intelligence.

Today, Claude Fable 5.1 is live on all platforms, available to everyone.

Mythos 5.1, fully powered, is locked within a whitelist and exclusively available to vetted cybersecurity and life sciences teams.

From refactoring complex engineering code to tackling long-term, high-dimensional scientific research, this twin brain is redefining what it means to be a top-tier "cyber cognitive worker."

What they excel at most is tackling complex, long-chain tasks that have previously been challenging for human experts.

As usual, let’s start with the scores—this time, the new model’s performance has shattered the competition by a massive margin.

• Research Exploration (Terminal-Bench-Science 0.1)

Fable 5.1 directly achieved 52.6%. For reference, the previous generation, Fable 5 and OpenAI's GPT-5.6 Sol, scored only 24.7% and 22.4%, respectively.

• Code Engineering (Terminal-Bench 4.0)

Fable 5.1 surged from the previous generation's 42.0% to 55.8%; with the barriers removed in Mythos 5.1, this figure climbed even higher to an astonishing 60.9%.

• Complex mental effort and integrated cognition

In the knowledge work benchmark GDPval-AA, Fable 5.1 achieved a high score of 1853; on the challenging benchmark known as "Humanity's Final Exam," the tool-augmented test surpassed the 65.0% threshold.

Even if you set the "thought level" to the lowest setting, it still delivers superior performance compared to the previous generation while consuming minimal tokens.

Even more remarkable, as capabilities leap forward, the barrier to computational power has been dramatically lowered.

In terms of base pricing, Fable 5.1 is identical to the previous generation: $10 per million tokens for input and $50 for output.

But for the most resource-intensive "cache reads," it has been slashed directly from $1 to $0.25, a reduction of 75%.

In real-world tasks, overall costs can be reduced by up to 25% in standard scenarios; for long-running tasks with heavy Agent usage, savings can reach as high as 45%.

In comparison, the current short-context promotional price for GPT-5.6 Sol is: $2 for input, $10 for output, and $0.2 for cache.

The high performance and price reduction demonstrate its strength, but this is merely surface-level.

The true turning point is that, starting with this generation, AI has completely shed its guise as an auxiliary tool—

It began to step in directly and conduct science itself.

To demonstrate the real-world impact of this "double king" pair.

Anthropic directly presented three cutting-edge examples that are much more intuitive than the rankings.

Redraw one-third of Venus

More than 30 years ago, NASA's Magellan spacecraft captured a series of radar images of Venus.

But its elevation data resolution is only 10–20 kilometers, which on Earth translates to an entire city occupying just a few pixels.

In the past, humans only created elevation maps for one-fifth of the area.

Fable 5.1 used this data to train a neural network, generating a high-resolution topographic map covering approximately one-third of Venus's surface.

The new map refines details to 2–3 kilometers, increasing height estimation accuracy by up to 25%.

Molecular and protein design, hit rate of 50%

Designing biological proteins is more challenging than processing static surface data.

The principle behind many modern drugs is to design a small protein that precisely binds to a specific target in the body, thereby blocking or activating a particular response.

How tightly they bind is scientifically called binding affinity. The higher the affinity, the greater the likelihood that the drug will be effective at low doses.

After invoking the open-source design tool, the designs produced by Mythos 5.1 were sent directly to an external facility for wet-lab validation in the real world.

The result is the strongest to date among all tested. On three targets—EGFR, Nipah G, and others—the affinity directly reaches ten times that of the best solution in the Adaptyv Bio competition.

The overall hit rate across the 12 targets is approaching 50%. Keep in mind that the typical hit rate in this field today is only around 10% to 15%.

Hand-written GPU kernels accelerate bioinformatics models by 2.5x

The third case, arguably, rewrote the research institution's computing power bill.

Genome and proteome research often requires repeatedly running specialized deep learning models on GPUs.

A whole-genome analysis may require testing a large number of mutations near each gene, causing the inference to be executed tens of thousands of times. Even a slight slowdown per iteration can quickly accumulate into significant costs.

Optimizing these models and writing custom GPU kernels is typically the work of top-tier performance engineers. It can take a team several weeks to accomplish, and academic labs often can’t afford such talent.

Mythos 5.1, on the other hand, only needed the public source code and completed the task in just a few days.

Moreover, it handwrote GPU kernels for seven open-source biological models at once, achieving up to a 2.5x speed improvement.

Most importantly, the output must be identical to the original,

When translated into billing terms, running an analysis of three million variants on the Evo 2 large model reduced the cost from $18,000 to just $8,000.

Inference speedup of seven open-source protein and genomics models on NVIDIA H100, along with estimated GPU costs before and after optimization

Fable 5.1 can directly engage in scientific endeavors, thanks to its extremely robust code logic.

Today’s agents aren’t afraid of writing code; they’re afraid of losing sight of their original goal after running for hours and calling dozens of tools, or being unable to identify the root cause after an error occurs.

Fable 5.1 strengthens precisely this complete action loop:

Read the repository documentation → Break down the task → Implement changes → Run tests → Review results → Identify failures → Proceed to the next round.

In Anthropic's words, Fable 5.1 clearly identifies where it gets stuck when encountering obstacles and rarely takes shortcuts that appear completed but leave underlying issues.

On CursorBench 3.2, which measures real-world IDE coding proficiency, it set a new internal record at 73.4%; in the commercial workflow benchmark AutomationBench, its score surged to 31.4%, nearly doubling.

Thanks to this long-term stability, it even helped the hedge fund Millennium resolve a hidden bug that had persisted for four to five years.

Faced with a ghost error occurring once in a million times, it traced the issue all the way to external code, disassembled third-party libraries, and uprooted the underlying flaw in someone else’s territory.

This time, Fable 5.1 truly delivers "end-to-end completion" for long-chain tasks.

First round of coded testing

After the full rollout of Fable 5.1, a series of real-world tests followed immediately.

Fable 5.1 Tops AAAI

The real-world comparison demo from the AI/ML API still clearly shows the difference between Fable 5.1 and GPT-5.6 Sol.

Feed them all the same prompt and output it at once.

Fable 5.1 cost $5.69, while GPT-5.6 Sol used only $0.88.

Fable strives for realism, with more lifelike waves, depth of field, and sandy rings around the islands. Sol, on the other hand, embraces a minimalist style, consistently maintained across all five scenes without ever breaking form.

Anthropic researcher Alex Albert directly gave Fable 5.1 a photo of a plot of land during a test.

It can autonomously complete house design, produce high-quality renderings, and even generate a cinematic virtual tour video.

In this hardcore demo, Fable 5.1 directly built a fully interactive model of the human brain.

It not only precisely replicates every fold of the cerebral cortex in code, but also lets you witness the entire process of the phrase “pass the salt” traveling through neural pathways.

Generate an ARC survival game, Fable 5.1 passed on the first try.

Anthropic's recent security moves focus on a "loose-tight" approach.

Song, the railing is no longer so "neurotic."

Previously, programmers using Claude to review their own code were frequently flagged as hackers. Now, with Fable 5.1, permissions have been expanded to allow code vulnerability detection, reducing false positives in cybersecurity by 60%.

But the bottom line remains strict: only identify vulnerabilities—never write exploit code. High-risk operations like penetration testing must still be handled by the more strictly regulated Opus team.

Then, it turns the moat directly into an "anti-distillation mechanism."

To completely block the previous method of modifying context across multiple dialogues to extract Claude's "chain of thought," Fable 5.1 has locked it at the API level.

Starting now, the API will strictly check the context.

If the prompt you submit differs from the one used to generate the thought block, the API will either immediately reject it with an error or completely remove the entire thought block—leaving no room for loopholes.

Currently, this mechanism targets newly registered API accounts created after August 31.

However, the official has stated that this lock will be mandatory for all future new models.

Top researchers in Silicon Valley have repeatedly mentioned a vision: the greatest value of AI is to compress decades of progress in biomedicine and fundamental science into just a few years.

From redrawing the map of Venus to increasing protein hit rates manyfold in test tubes, this pair of twin brains is turning this prophecy into reality.

The threshold for scientific discovery is visibly lowering.

Just don’t forget that the hand offering the flame also quietly pulled up the ladder behind it using “reverse distillation” at the same moment.

Reference materials:

https://www.anthropic.com/claude-fable-and-mythos-5-1

https://www.anthropic.com/news/enterprise-frontier-safeguards

This article is from the WeChat public account "New Intelligence Yuan" (ID: AI_era), authored by Moses Peach.