Anthropic says Claude Fable 5.2 scored 64.9% on RoboDojo, a public benchmark in which an AI controls a simulated two-arm robot from camera images, against 29.8% for OpenAI’s GPT-6 Astra on the same harness. That is 2.2 times GPT-6 Astra’s score, by our arithmetic, on Anthropic’s own runs, as shown on its launch page.
Fable 5.1 scored 27.6% on the same test, so the new model added 37.3 percentage points, or 2.4 times the score, in one release, by our arithmetic. Four days later, Fortune reported that OpenAI had disbanded its stand-alone robotics team because its general models can be applied directly to robots.
Business Pill · Physical AI
An evergreen Business Engineer explainer on physical AI: models that act in the physical world through robots and devices. It teaches the general idea only and does not discuss Fable 5.2 or OpenAI.
The key insight: As we read the two sources together, the robotics result came from a general model rather than a robotics model: Anthropic says Fable 5.2 improved even on robotics tasks it was never explicitly trained on, and its RoboDojo score of 64.9% is 2.2 times GPT-6 Astra’s on Anthropic’s runs, by our arithmetic, while OpenAI told Fortune its general models can be directly applied to robotics.
Anthropic says that “compared with earlier Claude models, Fable 5.2 shows substantially improved skills in controlling physical robots—even for robotics tasks on which it has never explicitly been trained”. Its video shows robots controlled by Fable 5.2 solving a puzzle cube, unscrewing a bottle cap and drawing a picture of a robot.
On RoboDojo, Anthropic says Fable 5.2 “outperformed other frontier models that were run on the same harness, as well as most AI systems built specifically to control robots”. The company’s post on X on 9 November made the same point.
The scores in Anthropic’s chart, all at the highest available effort level and all from Anthropic’s internal runs: Fable 5.2 64.9%, Opus 5.5 39.4%, Gemini 4 Argon 34.2%, GPT-6 Astra 29.8% and Fable 5.1 27.6%.
By our arithmetic, Fable 5.2’s score is 1.9 times Gemini 4 Argon’s and 1.6 times Opus 5.5’s, and it leads GPT-6 Astra by 35.1 points. Our launch piece and physics and operating-system piece cover the model’s other results.
RoboDojo Sim average success rate at the highest available effort level, from the chart on Anthropic’s Fable 5.2 launch page; all five scores are Anthropic’s internal runs on one harness.
What OpenAI Changed
OpenAI confirmed the reorganisation when Fortune asked, the magazine reports. Its robotics research team now sits in the main research team under chief research officer Mark Chen, and the applied robotics and hardware teams moved to the Applied organisation under president Greg Brockman, Fortune says.
A company spokesperson told Fortune that robotics remains a core part of OpenAI’s long-term vision, and that general-purpose frontier models, including GPT-6 Astra, “have shown that can be directly applied to robotics”. Our earlier piece covers the move.
Fortune also cites an Nvidia study in which an agent powered by GPT-6 Astra reached a 74.8% success rate on a simulated manipulation benchmark, against 51.0% for the best specialised robotic models on that leaderboard. That is a different benchmark and harness from RoboDojo, so the two scores for GPT-6 Astra are not comparable.
What the Wider Field Says
Fortune says OpenAI’s reorganisation reflects “a bet that is gaining traction across the AI industry”: that general-purpose models built for language, code and vision can also control robots. Fortune quotes Carnegie Mellon’s Guanya Shi as reading OpenAI’s move as a signal that it sees robotics as “no longer a special problem”.
Fortune also quotes MIT professor Russ Tedrake, chief executive of Walden Robotics, saying robotics companies keep an advantage the labs lack: robots already working on factory floors and gathering data from those sites. Nvidia’s own robot model is covered in our GR00T 2 piece.
The Structural Read
As we read Anthropic’s chart, the gain sits in one release: Fable 5.2 added 37.3 points over Fable 5.1 on the same benchmark and the same effort setting, by our arithmetic, a larger gap than the 11.8 points between Fable 5.1 and Opus 5.5.
As we read the scores, no independent leaderboard yet confirms the ordering: all five come from Anthropic’s internal runs on one harness, and the GPT-6 Astra result Fortune cites from an Nvidia study used a different benchmark.
Anthropic, Claude Fable 5.2 launch page
“Fable 5.2 shows substantially improved skills in controlling physical robots”
Three Implications
FOR ROBOTICS START-UPS Fortune reports that the general-model approach would seem to challenge the assumptions behind start-ups that raised money on robots needing brains of their own, and quotes MIT’s Russ Tedrake on the deployment data those companies hold.
FOR READING BENCHMARK CLAIMS RoboDojo scores here are Anthropic’s internal runs at the highest effort level; per-task results and real-robot success rates are not on the launch page.
FOR OPENAI’S ROBOTICS PLANSOpenAI told Fortune robotics remains a core part of its long-term vision; its research team now reports to Mark Chen and its applied robotics and hardware teams to Greg Brockman’s Applied organisation, Fortune says.
The framework records: “The compounding is not uniform across capability dimensions.”
As we read it, robot control is one such dimension: the RoboDojo score moved 37.3 points in a single release, by our arithmetic; the framework was published before Fable 5.2 and does not discuss robotics.
What Is Not Established
All five RoboDojo scores come from Anthropic’s internal runs on its own harness, not from an independent leaderboard, and Anthropic does not give per-task results on the launch page. The robot video shows selected tasks; it is not a measure of success rates on physical hardware.
OpenAI did not comment on Anthropic’s results, and neither company was contacted for this piece.
This Week In The AI Supercycle
AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source
This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.
On Anthropic’s own RoboDojo runs, Fable 5.2 scored 64.9%, 2.2 times GPT-6 Astra and 2.4 times Fable 5.1, by our arithmetic, in the same week that OpenAI disbanded its stand-alone robotics team and moved its researchers onto other teams, as Fortune reported.
A note on sourcing. The RoboDojo scores and quotations come from Anthropic’s Claude Fable 5.2 launch page and its 9 November post on X; all five scores are Anthropic’s internal runs on one harness. OpenAI’s reorganisation, its spokesperson’s statements, the Nvidia study and the outside comments come from Fortune’s report of 13 November, read in full. Multiples and point gaps are our arithmetic. Neither company was contacted for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
