Anthropic Releases Claude Fable 5.2 and Mythos 5.2…

Anthropic is stepping on a very delicate moment in time.

On one side is a highly uncertain IPO exam paper. Even though the valuation has converged to about $1.5 trillion, potential investors' doubts about the return on Anthropic's massive investments remain unresolved. Just now, Reuters disclosed a telling episode: Anthropic had sent a key document to underwriting banks, listing an offering size of about 867 million shares, a price range of $70 to $80 per share, corresponding to a maximum total offering of about $69.36 billion.

However, after the document was sent, the company urgently withdrew the relevant pricing information, saying it was "not accurate and will be updated separately."

On the other side, under close scrutiny from the capital markets, Anthropic has unveiled a long-prepared technological trump card: the new-generation flagship-tier models Claude Fable 5.2 and Mythos 5.2. Among them, Fable 5.2 is open to ordinary users through safety restrictions in the cybersecurity and biology fields, while Mythos 5.2 is mainly provided to vetted cybersecurity and life sciences research institutions.

But beyond Anthropic's traditional strength in autonomous coding engineering capabilities, this time a new variable has been added—accelerating penetration into the physical world and embodied fields.

On RoboDojo, an industry-recognized authoritative benchmark for robotic arm manipulation simulation, Fable 5.2 achieved an average success rate of 64.9%, not only leaving behind the previous generation Fable 5.1 (27.6%) and Opus 5.5 (39.4%), but also directly reaching more than twice that of GPT-6 Astra (29.8%).

In addition, this capability is rapidly penetrating into reality—in a subsequent real-machine blind test conducted by third-party developers, it took over a real robotic arm and completed a full 62-move game of checkers, and demonstrated a rare ability to correct deviations when frequently encountering physical contact errors and hardware deformation.

At the same time, Bingjie Tang, a former dexterous manipulation engineer on the Tesla Optimus team, confirmed that she is joining Anthropic as an MTS (Member of Technical Staff). During her time at Tesla, she long worked on learning-based dexterous manipulation for humanoid robots, covering data collection, policy training, model evaluation, and real-robot deployment; previously, at the NVIDIA robotics lab, her research focused on manipulation in contact-rich environments and sim-to-real transfer.

In fact, this talent flow is not an isolated move. Looking back over the past year, Anthropic's testing chain in the physical world has actually been very clear:

In 2025, Project Fetch was launched, initially having Claude assist non-robotics experts in programming and controlling robot dogs, and subsequently further testing the model's ability to autonomously connect devices, read sensors, and execute tasks;

In July 2026, the research "Claude Plays Robotics" was released, expanding the testing scope to robotic arms, quadruped robots, and humanoid robots, exploring how general large models can control robots through visual perception, task planning, and invoking motion policies. Project Pilot, launched the same month, expanded the experimental subjects to drones, testing autonomous navigation and target tracking capabilities.

In August, the Model Hardware Standard (MHS) was released, attempting to establish a unified interface for AI to operate physical devices such as microscopes and robotic arms.

It can be seen that Anthropic is clearly no longer satisfied with having Claude execute tasks only in software environments. And this probing into the physical world is essentially a boundary breakthrough forced by commercial reality.

At this critical juncture of preparing for an IPO, the capital market's scrutiny of Anthropic's valuation of up to $1.5 trillion is becoming increasingly stringent. Although overall revenue is still racing ahead, its third-quarter quarter-over-quarter growth rate has significantly slowed from 144% in the previous quarter to 47%. Along with the gross margin pressure brought by massive computing investment, relying purely on API calls and enterprise software subscriptions within the screen is visibly hitting the ceiling of growth.

The strategic weight of Physical AI has thus been rapidly elevated.

From partnering with engineering services provider UST to embed Claude into actual R&D in chip manufacturing and automotive production lines, to launching the MHS specification to directly orchestrate scientific instruments and industrial equipment; and now to Fable 5.2 doubling its robot simulation score and former Tesla Optimus dexterous manipulation engineer Bingjie Tan joining quickly—Anthropic is trying to prove to investors that Claude will not stop at being the brains in a chat box; it has the potential to become the execution core of physical productivity.

Although the embodied business is unlikely to become a performance pillar in the short term, the shift in the focus of industry competition is also an indisputable fact: large models are no longer satisfied with "answering questions"; they want to "take over work."

This logic of accounting by the "final delivery result" has likewise extended to the software development side. A most dramatic contrast then appears with Fable 5.2: even though its token list price is a full two times that of Opus 5.5, on high-difficulty engineering tasks the actual cost bill it delivers is instead half as expensive.

Benchmark Score More Than Doubles GPT-6; Real-World Tests Show It Learns to Correct Itself

Remember the discussion that GPT-6 Astra sparked in the robot simulation field back in September this year?

At the time, the paper "An Unexpected Robot Policy" presented a rather striking direction: there is no need to train a policy model specifically for each robot task; a general-purpose large model can also directly participate in robot control through visual observation, reasoning, and action interfaces. This led many people to begin re-examining the relationship between general-purpose large models and specialized embodied models.

Now, Anthropic has pushed this result substantially further forward.

In the officially published RoboDojo Sim test, Fable 5.2 reached 64.9%, GPT-6 Astra was 29.8%, its own previous generation Fable 5.1 was 27.6%, and Opus 5.5 was 39.4%.

Converted, Fable 5.2's score is 2.18 times that of GPT-6 Astra and 2.35 times that of Fable 5.1.

△ RoboDojo Sim scores of various models. Image source: Anthropic. This set of results was run internally by Anthropic, using the highest available reasoning intensity for each model.

RoboDojo assesses not conceptual understanding but end-to-end task execution: using only the continuous frames input from a camera, the model must autonomously determine the spatial trajectories of each joint of the robotic arm and the grasping force. An average score of 64.9% means it has moved beyond the trial-and-error category of "occasional attempts" and begun to possess the stability to plan continuous actions across scenarios.

In the official real-machine demonstration, Fable 5.2, without any targeted fine-tuning for specific tasks, completed fine movements such as manipulating puzzle blocks, unscrewing a bottle cap, and holding a pen to draw simple sketches on cardboard.

In addition, an unfiltered real-world test by a third-party developer further probed the upper limit of Fable 5.2's generalization ability.

In the test, developer Karolina Dubiel (@karolina_dubiel) connected Fable 5.2 directly to a real physical robotic arm and had it play a complete game of checkers against a human in a real desktop environment.

This seemingly simple adversarial test actually challenges the roughness and uncontrollability of the real physical world:

- Fault tolerance: Across the 62 moves of the entire game, due to reflections on the tabletop, slight displacement of the pieces, and insufficient friction in the gripper, the robotic arm experienced 14 instances of grasp slippage or offset. But without manually rewriting code, the tester merely input "try again" in natural language, and Fable 5.2 was able to combine the visual feedback to proactively review the cause of the previous step's failure, recalculate the grasp angle, and complete the action retry.

- Physical feedback perception and dynamic compensation: When the robotic arm extended under load, its end effector drooped by about 7 millimeters due to gravity and mechanical clearance. Faced with this non-preset physical deformation, Fable 5.2 did not mechanically execute the original theoretical spatial coordinates; instead, it read the torque and displacement telemetry data returned by the joint sensors and autonomously superimposed a compensation offset in subsequent actions, completing accurate placement of the pieces.

At the same time, Dubiel's test also exposed clear limitations of the current approach.

She added in the comments that the entire game actually took about 46.5 minutes, averaging about 45 seconds per move. The public video was sped up several times to even a hundred times, and the actual operation was not as smooth as the video appears.

Nominal Token Price Doubles; Why Is Delivery Cost Halved?

On the commercial side, Fable 5.2 creates an apparent pricing paradox.

Its API is priced at $8 per million input tokens and $40 for output. Although this is 20% lower than the previous generation, it is still a full two times that of the flagship model Opus 5.5. Under conventional logic, doubling the unit price means R&D costs rise accordingly.

However, in multiple tests involving real engineering, the ledger moves in the opposite direction:

Code merge benchmark FrontierCode 2.0: Fable 5.2 needs only the High reasoning tier to score 45.3%, matching Opus 5.5's top tier of 45.5%, while cutting total cost per task by about 50%; if pushed to the Xhigh tier, its score climbs directly to 49.1%, leaving GPT-6 Astra (42.8%) behind.

Knowledge and chart tasks: In the Perplexity WANDR large-scale information gathering and Chartography chart benchmarks, Fable 5.2 achieved the same or even higher accuracy at about 40% and less than 10% of Opus 5.5's per-task cost, respectively.

The essence of this cost inversion lies in the crushing blow that "first-pass yield" deals to compute consumption.

Real-world tests by community developers exposed the key gap: when building a 3D submarine scene, Opus 5.5 took 6 hours and 40 minutes and required frequent human intervention to correct errors and fill in details, whereas Fable 5.2 took only 1 hour and 26 minutes, taking shape almost in a single pass.

In complex engineering with multiple nested rounds, the most expensive part of traditional large models has never been the listed Token price of a single output, but rather the endless context rework triggered by logic breakpoints, debugging backtracking, and the hours of human engineer intervention.

Of course, this advantage does not apply to all scenarios: in the OSWorld 2.1 system operation test, Fable 5.2 (89.3%) led Opus 5.5 by only 2.1 percentage points; in the AutomationBench business workflow test, it even trailed Gemini 4 Carbon's 53.9% with 51.4%.

In other words, when handling basic spreadsheets and shallow system clicks, Fable 5.2 may be an overpriced tool with excess performance; but in deep waters requiring high fault tolerance and deep logical self-consistency, it is a skilled worker that can save you dozens of rounds of futile debugging.

The agent tipping point: ultra-long reasoning cycles

If code and robotics tests demonstrated localized productivity, then in extreme engineering experiments, Fable 5.2 for the first time made the sustained operation of large-scale agent clusters industrially viable.

The most typical validation is the formal transcription of Fermat's Last Theorem. Anthropic orchestrated about 1,000 Fable 5.2 agents and, in an environment with little human intervention, completed the entire formalization process of the theorem in the Lean language in only about 30 hours, whereas the system previously required 11 days.

A more critical comparison occurred in the middle-to-late stages of the task:

After 18 hours of continuous operation, Fable 5.2 reached 86% completion, while Opus 5.5 was at only 26%;

Continuing to 30 hours, Fable 5.2 converged smoothly to 100% completion, while Opus 5.5 barely crawled to 27%, falling into a logic deadlock.

This ability to resist context drift and avoid logic infinite loops in extremely deep reasoning networks was directly reproduced in real engineering development:

Building a system from a blank sheet: engineers had it start from an empty folder and orchestrated more than 1,000 sub-agents to run continuously without rest for 8 days, ultimately delivering a complete usable laptop operating system prototype that includes a kernel, drivers, a file system, window interaction, and an embedded Claude.

Underlying architecture migration: a group of agents directly rewrote a key production service entirely from Python to Rust, and after launch the service startup speed increased 26-fold.

End-to-end penetration across software and hardware: based solely on a requirements description, it worked with Diode to design three physical electronic products, seamlessly covering everything from component selection, circuit board design, and 3D modeling of the enclosure to firmware flashing in one pass; after hardware prototyping, the devices powered on and ran without any circuit modifications.

In the past, the industry often talked about multi-agent collaboration, but after dozens of tool-chain calls, accumulated errors often plunge the entire system into a hallucination avalanche. The technical increment Fable 5.2 demonstrates is precisely that it pushes the fault-tolerance threshold of long-horizon tasks outward by a big step.

Conclusion

From 8 consecutive days of low-level system reasoning on screen to adaptive correction of robotic arm end-effectors in the physical world, Anthropic, which is making waves in the secondary market, is clearly no longer content to be a high-premium software Copilot.

Under the spotlight of the capital market, the story it is reconstructing is that of an autonomous execution unit moving toward physical productivity.

This article comes from the WeChat public account "AI Frontline" (ID: ai-front), authors: Feiye, Qinghe, published by 36Kr with authorization.