Field intelligence for AI-first professionalsVol. II · Nº 195 · Monday, November 16, 2026
← Phantom Notes
Nº 188AI ModelsNovember 13, 20264 min read

OpenAI Broke Up Its Robotics Team Because GPT-6 Astra Can Run Robots. The 74.8% Behind That Is a Simulator Score, It Costs $36.40 an Episode, and Another Agent Already Scores 86%.

Fortune reported on November 13, with confirmation from an OpenAI spokesperson, that OpenAI disbanded its stand-alone robotics team: research staff moved into the main research organization under Mark Chen, applied robotics and hardware into Applied under Greg Brockman. OpenAI framed it as acceleration, saying general models such as GPT-6 Astra can be applied directly to robots. The evidence cited is NVIDIA's Astranomy study, dated October 14, in which Astra running as an agentic policy inside a persistent Codex session scores 74.8% on the simulated RoboLab-120 benchmark against 51.0% for the best specialized policy. That setup averages 30.9 million tokens and $36.40 per episode, direct inference calls average 18.56 seconds, and NVIDIA's own leaderboard now lists an agentic model, AMP1, at 86.0%. We read Fortune, NVIDIA's study page, and the leaderboard.

By T.W. Ghost


Fortune reported today, November 13, 2026, that OpenAI has disbanded its stand-alone robotics team. Research staff move into the main research organization under Mark Chen; applied robotics and hardware move into Applied under Greg Brockman. An OpenAI spokesperson confirmed the change when asked, and the company's framing is acceleration, not retreat: general models such as GPT-6 Astra can be applied directly to robotics, so a separate team is no longer the fastest path. Robotics remains core, OpenAI says, with a humanoid demonstration still targeted for 2027, per Sam Altman. Fortune also notes that OpenAI's booth at CoRL, number 208, sat empty.

The evidence the reorg leans on is a month old and belongs to NVIDIA. The Astranomy study, a CoRL 2026 late-breaking result from NVIDIA's Seattle Robotics Lab dated October 14, ran Astra three ways on RoboLab-120, a simulated benchmark: as a direct policy it scored 66.3%, orchestrated 66.5%, and as an agentic policy inside a persistent Codex session 74.8%, which NVIDIA says beats the best specialized RoboLab-120 policy, Lakeside, by 23.8 points. Then the fine print. The agentic run averaged 30.9 million tokens and $36.40 per episode, against 160,000 tokens and $1.96 for the direct policy. Direct policy inference calls averaged 18.56 seconds each. And on NVIDIA's own live leaderboard the top entry is not Astra: an agentic language model listed as AMP1 scores 86.0%, with SenseNova-RoboRSI at 57.1%, Lakeside at 51.0%, and Runway's Praxis-1 fifth at 46.2%. We read Fortune, NVIDIA's project page, and the leaderboard. Here is what held up.


What Was Announced, and What It Rests On

DetailWho
ReorgStand-alone robotics team disbanded; research to Mark Chen's org; applied robotics and hardware to Greg Brockman's AppliedPress (Fortune), spokesperson-confirmed
OpenAI's framingRobotics remains core; general models apply directly; humanoid demo targeted 2027Press relay of vendor statements
CoRL boothNumber 208 sat emptyPress observation
Astranomy studyNVIDIA Seattle Robotics Lab; CoRL 2026 late-breaking result; dated Oct 14, 2026; arXiv "coming soon"Verified (NVIDIA page)
RoboLab-120 resultsDirect 66.3%; Agentic (Astra in a persistent Codex session) 74.8%; Orchestrated 66.5%Verified (NVIDIA study)
Against specialized policiesAgentic beats Lakeside, the best RoboLab-120 policy, by 23.8 pointsVerified (NVIDIA study)
Cost and tokensAgentic 30.9M tokens and $36.40 per episode; Direct 160k tokens and $1.96Verified (NVIDIA study)
LatencyDirect policy inference calls average 18.56 sVerified (NVIDIA study)
Live leaderboardAMP1 (Agentic LM) 86.0%, 2,579 of 3,000; SenseNova-RoboRSI 57.1%; Lakeside 51.0%; Praxis-1 fifth at 46.2%Verified (leaderboard page)
Astra on the leaderboardNot visible in the text we read; AMP1's vendor unknownUnverified
Quarterly operating loss "$10.9 billion"Press only ("reportedly"); not rechecked

Simulation, not robots. RoboLab-120 is a simulated benchmark. A 74.8% in a simulator is a research result; it says nothing about a gripper in a kitchen.

$36.40 and 18 seconds. The agentic setup that produces the headline score spends 30.9 million tokens an episode, and even the cheaper direct policy waits 18.56 seconds per inference call. Robots that need to move within a second are not the audience for either.

Not even first. NVIDIA's own board puts an agentic model called AMP1 at 86.0%, above Astra's agentic 74.8%. Who built AMP1 is not stated on the page we read.

Old evidence for a new decision. The study is dated October 14. The reorg is this week. A month-old simulation result is being asked to carry a corporate restructuring.


Verified vs Unconfirmed: The Scorecard

ClaimVerdict
Robotics team disbanded; staff reassigned; spokesperson confirmedPress (Fortune); no OpenAI post found
Robotics remains core; 2027 humanoid demoPress relay of vendor statements
Astranomy study date, venue, results, cost, latencyVerified (NVIDIA project page)
Leaderboard rows: AMP1 86.0%, SenseNova 57.1%, Lakeside 51.0%, Praxis-1 46.2%Verified (NVIDIA leaderboard)
Astra rows on the live leaderboard; AMP1's identityUnverified
$10.9 billion quarterly operating lossPress only; not rechecked

What To Do

Read "general models run robots" as a lab result with cloud latency, which is what NVIDIA's own numbers describe, and not as a deployment path.

Skip any vendor pitch built on a success rate alone; ask for on-robot latency and per-episode cost, the two columns that turn 74.8% into $36.40 and 18 seconds.

The frontier is elsewhere: on the same NVIDIA board, Runway's Praxis-1 claimed the top and sits fifth, and on Artificial Analysis's physical arm test Claude Fable 5.2 leads at 75, on a different task set with its own caveats. Our model comparison has the current standings side by side.


Sources


*Not sure whether a general model or a trained policy fits your robot? Take the free 2-minute quiz and get matched. Then read about the model the reorg is named after: GPT-6 Astra.*

Which model should you be using?

Three minutes, twelve questions, one defensible answer.

Take the quiz →