OpenAI Folds Robotics Into Core Research as General Models Beat Specialized Brains
Claude Fable 5.2 leads RoboDojo at 64.9%; Physical Intelligence seeks $20B with no revenue

OpenAI disbanded its standalone robotics research group on November 13, 2026, folding its robotics engineers directly into the broader research organization. The timing was significant: in the same week, two independent robotics leaderboards showed general-purpose foundation models outperforming purpose-built robotics AI by significant margins — a convergence that made OpenAI's move look less like a retreat and more like an early read on where the technology is heading.
What the Benchmarks Actually Show
Four days before the OpenAI reorganization, Anthropic released Claude Fable 5.2 and published its performance on RoboDojo, a benchmark developed by the Multimedia Laboratory at the University of Hong Kong and researchers from nearly 20 institutions worldwide that evaluates robot manipulation across both simulated and real-world tasks. Fable 5.2 scored 64.9% on the simulated component of the benchmark — more than double the 27.6% of its predecessor Fable 5.1, released just two months earlier. It also outpaced Claude Opus 5.5 at 39.4% and GPT-6 Astra at 29.8%. Neither Fable 5.2 nor Astra has ever been designed or marketed as a robotics model: both are general-purpose systems used primarily for coding, reasoning, and multimodal tasks.
The same week, Artificial Analysis published updated results on its AA-Robotics v0.1 benchmark, which evaluates models controlling a physical robot arm through code generation. Fable 5.2 scored 75 on that measure, taking first place ahead of GPT-6 Astra at 69 and GPT-6.1 Sol at 65. It was also the first model to record a successful score on the benchmark's measure-box task — which requires estimating object dimensions from video without physical measurement tools — a spatial-reasoning challenge every model evaluated prior to Fable 5.2 had failed to complete.
Earlier in November, NVIDIA's Seattle Robotics Lab published Astranomy, a study that tested frontier AI models on a Franka robotic arm across a task-completion benchmark. Using GPT-6 Astra directly as a robot controller, the lab recorded a 66.3% task success rate — higher than the best dedicated robot model in the study at 51.0%, and far ahead of Physical Intelligence's π0.5 at 28.0%.
Why General Models Keep Outperforming Specialized Ones
The gap has a structural explanation. Physical Intelligence's π0 and π0.5 models are built on PaliGemma 3B — a 3-billion-parameter vision-language model — with a separate 300-million-parameter action expert. The architecture is designed for responsiveness: the action expert operates at around 200 Hz, far faster than a large language model could run.
The problem is knowledge contamination. When you train the action expert alongside the language backbone, gradients from the robotics training data interfere with the web-scale knowledge encoded in the language model. Physical Intelligence identified this issue and introduced a technique called knowledge insulation, formalized in a May 2025 paper. The key mechanism is a stop-gradient that prevents the action expert's learning signal from flowing back into the VLM backbone. The result is a model that learns faster, runs faster, and generalizes better across unseen tasks.
That fix works at the action layer. What it cannot solve is the knowledge gap at the planning layer — which is where general models are winning.
The Three-Layer Stack and Where the Value Lives
Modern robotics pipelines operate at three time scales. The planning layer — handled by a large language model — thinks at the second scale: it interprets instructions, maintains context across a task, and decides what to do next. The action layer operates at roughly 200 Hz, issuing the specific motor commands that move a joint from position to position. The motor layer runs at kilohertz rates, handling low-level physical control.
The value at stake is in the planning layer. The NVIDIA Astranomy study confirmed this explicitly: when Astra was used as an orchestrator — intervening in just 14.5% of steps during task execution — it raised a small strategy model's success rate from 36.8% to 66.5%. Most of its interventions corrected grasping location (64.4% of interventions) or wrong target or plan (20.2%). Those orchestrating interventions — comprising less than one in seven steps — contributed nearly thirty percentage points of improvement in task success. Running Astra as an orchestrator increased LLM inference costs per episode; operating it in a full agentic configuration raised those costs further.
Google DeepMind's Gemini Robotics 2 uses the same logic at the system level. Gemini 3.5 Flash sits at the top of the stack, handling planning. A dedicated action VLA operates in the middle layer. A smaller Gemma model runs on-device for the motor layer. The frontier model provides the cognition; the specialized components provide the speed.
The Doubling Problem for Dedicated Startups
A paper published in October 2026 — titled "The Bitter Lesson Comes to Robotics" — tested nine frontier models on physical robot arms across six tasks and found performance doubling roughly every 53.9 days, based on runs against real hardware from July through October 2026.
That doubling rate creates a structural problem for companies selling products built around today's specialized models. Physical hardware trials by Robocurve underlined the gap: in a block-in-bowl placement task, OpenAI's GPT-6 Astra completed 19 of 20 runs successfully — a 95% success rate. Claude Fable 5.1 scored 40% on the same test. Astra cost $0.94 per completed run; Fable 5.1 cost $2.12. The evaluation also revealed the limits of the planning-layer argument: when precision mattered — placing a puzzle piece into a groove — Astra's success rate fell to 10%, consistent with what dedicated action models handle today.
Physical Intelligence's own progression shows how fast the specialized models are themselves moving. π0, released in October 2024, was trained for specific manipulation tasks. π0.5, released in April 2025, generalized to mobile manipulation in novel home environments. The company has signaled a next iteration, internally referred to as π*0.6, that remains unannounced.
The Valuation Question the Benchmark Data Creates
Bloomberg reported in October 2026 that Physical Intelligence is in talks to raise funding at a $20 billion valuation. That would be up from $11.5 billion in an earlier round and from $5.6 billion in 2025. The company has not disclosed revenue.
A $20 billion valuation with no disclosed revenue is a bet on three claims: that the specialized action layer is durable, that Physical Intelligence has a defensible lead in physical world data collection, and that general-purpose foundation models will not absorb the planning-layer value that the action layer depends on. Each of those claims looked more complicated in the week those numbers were circulating.
What the OpenAI Restructuring Actually Signals
According to Fortune, Guanya Shi — who now leads robotics within the broader organization — described the transition in terms familiar to anyone who watched specialized NLP systems give way to general language models. Shi told Fortune that general models have become the substrate for robotics work and that robot-specific development has become primarily a problem of prompting and orchestration, not building specialized models from scratch.
That framing is imprecise in one important way: physical robot control still requires the action and motor layers that general language models do not natively provide. But at the planning layer — the layer that determines what a robot is trying to do and in what order — the analogy holds. And it is at the planning layer where Physical Intelligence, Figure, and every other company selling a foundation model for robotics is building their case.



