OpenAI folds robotics team into research and applied orgs as general models run robots

OpenAI breaks up its standalone robotics team

OpenAI has confirmed to Fortune that it is disbanding its stand-alone robotics group and redistributing its people across the company. Robotics researchers move into the research organization run by Chief Research Officer Mark Chen, while applied robotics and hardware staff join the Applied organization led by President Greg Brockman. According to Fortune, these are internal reassignments rather than exits: roles and the product roadmap stay intact, and the company still aims to hold its first public robot demonstration in 2027.

The company presents the change as an acceleration of its robotics effort rather than a pullback. Its position, as reported by Fortune, is that its general-purpose models — GPT-6 Astra in particular — are now strong enough to drive robots directly, making a separate specialist organization the wrong structure.

The results behind the decision

The argument has shifted over the year. According to the-decoder, OpenAI's robotics plan had previously been built around Sora, its video generation system, but the current claim is that the general models have caught up.

Two findings are cited in support. In NVIDIA research described by the-decoder, an agent running on GPT-6 Astra completed 74.8% of simulated DROID robotics tasks, against 51% for specialized robotics models. Separately, on StationeryBench, a benchmark hosted on GitHub under robocurve, Astra was tested against Ai2's MolmoAct2 on five tabletop-object tasks performed on a dual-arm YAM robot over 200 trials: Astra fully completed 7 of 100 tasks while MolmoAct2 completed none, with average progress scores of 46 versus 12.

Both numbers come with caveats. The DROID result is simulation only, and even the StationeryBench figure shows a model fully completing fewer than a tenth of its tasks.

Specialized models still have a case

Papers quoted by the-decoder indicate that dedicated robotics models retain real advantages in cost, reliability, speed and energy consumption — which makes OpenAI's restructuring a wager rather than a settled conclusion.

The competitive field is also wider than one lab, with Anthropic advancing Claude-based robotics research and Google pushing robots that run Gemini 4 Argon, per the-decoder. One Fortune detail captures how quickly the call was made: at the CoRL conference in Austin, OpenAI still held a booked but empty booth, number 208. As Fortune relays, CMU professor Guanya Shi argues that robotics has simply stopped being a special problem.

Not OpenAI's first robotics shutdown

OpenAI shut its original robotics team in 2020, citing insufficient training data and a judgment, described by co-founder Wojciech Zaremba, that large language models looked like the faster route to AGI. The effort returned in 2024. What separates this reorganization from the 2020 closure is the evidence base: general models now have measurable robotics results to point to, from the NVIDIA study to StationeryBench. Folding the team in therefore reads as a bet on speeding the work up, not leaving the field.

Why it matters

For automation and robotics teams, the signal is directional. Embodied AI is being absorbed into the same general-model stack already used for text and vision — one model family potentially spanning all three — and the reported 2027 demonstration is the checkpoint to watch. Treat the simulator numbers with care, though: success on simulated DROID tasks is not a robot on a factory floor, and the cost, speed and energy advantages of specialized models remain the strongest counterargument to track. If you are planning robotics integrations, the practical question is whether OpenAI ships developer-facing interfaces as the research organization absorbs the team.