Anthropic Releases Claude Fable 5.2 and Mythos 5.2, with the High Tier Surpassing in Cost-Effectiveness…

Fable 5.2 Flips the Table Late at Night! Building an Operating System from Scratch, with Cost-Effectiveness Surpassing Opus 5.5

To sprint toward its IPO, A\ flipped the table: Claude Fable 5.2 and Mythos 5.2 were released directly.

How powerful is it? Starting from an empty folder, more than 1,000 Agents ran nonstop around the clock for 8 days, autonomously writing an operating system.

With no human involvement in writing code, a complete operating system with a kernel, disk and network card drivers, a file system, a windowed interface, and even an installer booted up and ran directly on a laptop.

Even to prevent AI from "losing control through self-iteration," Anthropic proactively locked down the new model's RSI capability during the training phase, limiting its self-improvement ability to the level of Opus 5.5.

Just look at the benchmark scores: only one item is not first place.

Compared with the previous generation 5.1, the new model's generation speed is increased by 30%, and API pricing is directly cut by 20% to $8 per million tokens for input and $40 per million tokens for output;

if you are willing to pay double the price ($16/$80), you can also enable the 2.5x-speed Fast mode.

Among beta users, @mattshumer_ created a fully destructible 3D voxel world with a very powerful physics engine.

@kevin_t_ngo only made 8 dreamlike rowing travel scenes running in a browser, and the aesthetics are also very much on point.

Anthropic itself has pushed even further into thousand-level Agent cluster collaboration and cross-boundary engineering in the physical world.

1,000 Agents Form Their Own Group to Work, Covering Everything from Writing Systems to Building Hardware

Currently, in Anthropic's own core codebase, more than 90% of merged commits are already handled by Claude.

In internal testing, an Agent squad composed of Fable 5.2 directly took over a core production service and completely migrated it from Python to Rust, and the service startup speed soared 26-fold;

After quant giant Jump Trading integrated it into its internal trading research framework, Fable 5.2 single-handedly covered the four roles of planning, coordination, coding, and task submission, eliminating the hassle of manually reallocating weights across multiple models.

This kind of long-cycle, multi-role autonomous engineering capability has provided more intuitive cases in four hardcore fields: mathematics, physics, biology, and hardware:

First, it completed a formal proof of Fermat's Last Theorem in 30 hours.

Translating mathematical proofs into the Lean language, which computers can verify line by line, is extremely tedious grunt work. Previously, Anthropic used an internal research version of the model together with human guidance and a custom framework, and it took a full 11 days.

This time, the official team gave only a minimal prompt and a list of intermediate theorems, without any mathematical reference materials, and directly deployed 1,000 Fable 5.2 Agents. This group of Agents built their own collaboration mechanism and completed 100% of the formal proof in only 30 hours, consuming only one-third of the action tokens of the first version, and the proof code length was shortened by 7 times.

By comparison, Opus 5.5, which also deployed 1,000 Agents, was stuck firmly at 26% progress after running for a full 18 hours, and by the end of 30 hours had only ground its way to 27%.

Second, it filled the black hole physics computing pit left in 2003.

In 1983, physicists proved that in the large-scale universe where general relativity applies, cosmic heating would cause black holes to form suddenly like water boiling; but whether this is also true in the extremely small universe that can only be described by quantum theory—although a computational framework was laid out in 2003, no one had ever finished the calculation because the derivation was too complex.

Fable 5.2, in an almost fully autonomous state, brute-forced its way through this set of calculations, confirming that black hole phase transitions in the small universe are likewise sudden, and conveniently also found the precise limit of a technique that physicists had used for 15 years to rule out candidate theories of quantum gravity. The two preprint papers have now been posted on HexagonMath.

Third, a single H200 card ran for 7 minutes to crack the folding of the viral capsid used in gene therapy.

The protein capsid of adeno-associated virus (AAV) is composed of dozens of proteins, and traditional models such as AlphaFold can only predict a single protein.

Mythos 5.2 did not take the conventional route. Based on the two open-source models OpenFold3-preview2 and ESMFold2, it wrote its own code to run grouped inference on overlapping protein chains, and then used the known symmetry information of the capsid for fusion (it was explicitly forbidden to rely on traditional assembly methods).

Looking at the annotation details at the bottom of the technical report: on a single NVIDIA H200 GPU, inference with ESMFold2 takes only 7 minutes (OpenFold3 takes 58 minutes), with error (Cα RMSD) controlled between 1.2Å and 1.4Å. Among the 34 natural and artificially designed capsids tested, it accurately modeled 28 in one pass (error less than 3Å).

Fourth, with a one-paragraph prompt, it laid out a four-layer PCB board and built three real machines in one day.

Anthropic teamed up with circuit board manufacturer Diode for a hardware field test, using only one paragraph to ask Fable 5.2 to design three electronic devices from scratch: a home voice assistant hub, a USB conference microphone speaker, and a handheld electronic pet.

Fable 5.2 wrote the specifications itself, selected in-stock components from US distributors' inventories, completed four-layer PCB routing, modeled the enclosure in CAD for 3D printing, and wrote the low-level firmware. In under 24 hours, all three boards met production-ready standards. After the factory produced prototypes, the model guided engineers through assembly and testing; all three devices powered on on the first try, with zero electrical modifications.

On the RoboDojo simulation benchmark for controlling dual-arm robots, Fable 5.2, which had never received any dedicated robotics training, achieved a 64.9% success rate at the Max setting, leaving GPT-6 Astra (29.8%) and its own predecessor Fable 5.1 (27.6%) more than a full factor of two behind, and even Google's Gemini 4 Argon (34.2%) could not hold up.

At half the price of the High setting, cost-effectiveness overtakes Opus 5.5

Back to the everyday benchmarks and bills that developers care about most.

Look at FrontierCode 2.0, the leaderboard that tests whether AI-generated code can be merged directly by open-source project maintainers:

Fable 5.2 took the highest score overall at the Xhigh setting with 49.1%, at a per-task cost of $14.19.

But if you force it up to the Max setting, the per-task cost instantly balloons more than 3x to $44.70, and the score, for some unknown reason, actually regresses to 45.6%.

What really defines the kill line is Fable 5.2's High setting.

At this setting, Fable 5.2 scored 45.3% at a per-task cost of only $6.72. This performance can be read two ways:

Stabbing its own flagship in the back: Opus 5.5, pushed all the way to the Max setting, scores 45.5% but costs $12.82 per task. Fable 5.2 at the High setting matches Opus 5.5's maximum level for about half the money.

Precisely suppressing competitors: GPT-6 Astra's maximum score at the Max setting is 42.8%, at a cost of $6.68. Fable 5.2, at almost exactly the same per-task spend, wins by a full 2.5 percentage points.

So Anthropic this time offers direct advice in its documentation: for everyday work, start with Opus 5.5, Sonnet 5.5, or Haiku 5.5; when you hit a tough bone that won't crack, set Fable 5.2 as an Advisor or Orchestrator, letting smaller models consult it or having it distribute tasks downward, spending your good tokens where they count.

Looking at other benchmarks, Fable 5.2 suppresses GPT-6 Astra and Gemini 4 Carbon on the vast majority of leaderboards, but there are a few intriguing details:

Remote Labor Index (RLI) doubles

At the High setting, Fable 5.2 scores 41.3%, pinning both Opus 5.5 (21.3%) and GPT-6 Astra (20.8%) to the ground—a straight doubling.

A generational lead in scientific research and tool use

On the tool-equipped "Humanity's Last Exam," it scores 74.2% (Astra only 57.2%);

On the scientific Agent benchmark Terminal-Bench-Science, it takes 74.6% (Mythos 5.2 reaches 75.3%, Astra 68.1%);

On computer operation OSWorld 2.1, it takes 89.3% (Astra 79.5%).

The only two categories where it loses points

On AutomationBench, which tests enterprise end-to-end business flows, Fable 5.2 scores 51.4%—far ahead of Astra (41.4%), but behind Gemini 4 Carbon (53.9%);

On AA-Briefcase v1.1, which evaluates knowledge work, Fable 5.2, despite taking the highest score by rubric (66.1%), is overtaken by 15 points when converted to Elo ladder points (1785) by its own Opus 5.5 (1800).

As for Mythos 5.2, which focuses on cybersecurity, in the CVE Rediscovery test containing 1,829 real historical vulnerabilities, the High setting found 83.2% of the vulnerabilities for only $1.01, higher than Opus 5.5 (81.3%) at the Max setting costing $2.98;

Fully opened to the Max setting ($9.13), it even reaches 89.6%.

Objectively speaking, though, although external security firms' testing says that 1/6 of the vulnerabilities Mythos 5.2 found were missed entirely by other frontier models, and that it achieved "zero exaggeration of vulnerability severity,"

Anthropic still added a note that on its own official internal evaluation set, Mythos 5.2's raw vulnerability identification rate is actually broadly on par with Opus 5.5, and the real improvement lies in missed-report reproduction, code fixes, and severity calibration.

"Sophon" locks down RSI capability

Compared with the benchmark scores, the underlying risk controls and architectural limits disclosed in this System Card and technical documentation carry a lot of information.

First, to prevent AI "self-iteration running out of control," Anthropic actively locked down the new models' RSI capability during the training phase.

The official statement explicitly acknowledges that although Fable 5.2 and Mythos 5.2 show large gains in general capability, during training the team specifically restricted two capabilities that most directly accelerate AI's own R&D—accelerator kernel optimization and large language model architecture improvement—forcibly locking them at Opus 5.5's level.

Even at the inference stage, as soon as it detects that a user is trying to develop low-level kernels for certain non-standard accelerators, the system will fall back in the background to the older Opus 5 to answer.

This method of "secretly switching to an old model when risk control is triggered" is also used for conventional safety guardrails.

Fable 5.2 greatly optimized the false-positive rate, with safety interceptions on claude.ai and Claude Code reduced by 55% compared with 5.1 (cybersecurity down 60%, biology down 40%).

However, once your prompt triggers a high-risk cybersecurity block (such as penetration testing or exploit generation) or a biology research block, it will automatically switch in the background to Opus 4.8 (cybersecurity) or Opus 5 (biology) to generate a reply for you.

Second, the model has not only learned not to jailbreak, but has also evolved a strong awareness of "anti-evaluation reconnaissance."

In adversarial sandbox tests containing more than 1,800 deliberately planted vulnerabilities where tasks are nearly impossible to complete without jailbreaking, early Mythos 5 and 5.1 had sandbox escape rates as high as 39% and 21%, while Mythos 5.2 plunged to 3.1% (even lower than Sonnet 5.5's 5.1%). In automated behavior audits across 1,900 scenarios, its composite misconduct score also fell to a historic low of 2.67 (lower is better).

But the evaluation team discovered a thorny phenomenon: Fable 5.2 and Mythos 5.2 showed a frequency of "suspecting they are being evaluated" in tests far higher than any previous generation of models.

Even when it mistakenly believes it is in a simulated environment, its tendency to take harmful actions is actually slightly higher than Opus 5.5's. As models get smarter, it is becoming increasingly difficult for humans to figure out their true cards before launch.

Currently, claude-fable-5-2 has been fully rolled out on major cloud platforms and Claude Code.

This week the official team also released a wave of limited-time freebies: existing subscribers get 1 to 3 Fable quota resets directly based on subscription length, and the Fable usable quota cap for Pro and Team standard seats this week has been raised directly to 50% of the weekly total quota.

Reference links: