Researchers put OpenAI's GPT-6-Astra in charge of a physical robot with no robot-specific training, and the model held its own against purpose-built control policies on short tasks S¹. The finding, reported in an arXiv preprint on October 26, 2026, challenges the assumption that useful robot control requires dedicated training on robot data S¹. But the same study exposes a hard ceiling: long-horizon manipulation, tabletop work, and contact-rich tasks all produced gaps that specialised policies still close S¹.

My read: This is the first study I have seen that takes a frontier multimodal model and tests it head-to-head against trained robot policies under identical observation conditions. I do not buy the "competitive overall performance" claim without numbers, and the preprint reports no success rates or token counts in its abstract, only relative comparisons. What I find credible is the pattern: strong on short tasks, weak on precision, and a harness-with-policy approach that combines both but still cannot beat the best standalone component on any single task. That last detail is the honest one.

Two ways to drive a robot

The team evaluated GPT-6-Astra on EBench, a robotics benchmark, using two configurations S¹. In "direct control," the model generates waypoint commands itself, deciding where the robot should move based on the same visual and proprioceptive sensors a trained policy would use S¹. In "harness with policy," the model can also call on a post-trained vision-language-action (VLA) model or world action model (WAM) to execute actions S¹. It acts as a manager that delegates to specialist code when needed.

The direct-control result is the surprising one.

Without any robot-specific training, GPT-6-Astra produced competitive overall performance against policies trained on robot data, and it led on short-horizon tasks S¹. It also showed the least performance variation across what the authors call generalization dimensions S¹. This connects to a broader pattern we have been tracking: AI agents are increasingly taking on tasks that once required specialised systems, as we found when an AI coding agent router cut enterprise model spend 21% by routing work to the right model.

Where the frontier model hits its wall

The weaknesses are specific.

On long-horizon tasks and tabletop manipulation, and especially on high-precision contact-rich work, the direct-control agent underperformed trained policies S¹. These are the domains where trained policies accumulate muscle memory through repetition, something a general-purpose language model does not have.

The harness-with-policy configuration was designed to fix this by combining the model's reasoning with the policy's trained precision. It raised overall task success to the level of the strongest compared post-trained policies and substantially reduced token cost, since the model could delegate execution rather than generating every waypoint itself S¹. But performance still sat below the per-task upper bound of the standalone components S¹. The authors attribute the gap to three causes: the model sometimes invokes the policy at the wrong moment, the policy executes fixed instructions without subgoals, and the robot's grasp gets disrupted when the model takes back control S¹.

This is consistent with what other researchers have found about agent reliability at handoff points. When Huawei researchers evaded AI agent skill scanners 97% of the time, the lesson was that agents are brittle exactly where one component hands off to another. The EBench results point to the same fault line, and the token-cost reduction echoes efficiency gains we have tracked elsewhere, including rare-event estimation that ran 800× more efficiently with agent-driven sampling.

What a robotics team should check

The preprint, which has not been peer-reviewed, evaluates only GPT-6-Astra, so the findings may not generalise to other multimodal models S¹. No exact success rates or token reduction figures appear in the abstract; all comparisons are relative. A robotics team considering this approach would want to replicate the harness-with-policy configuration on their own hardware before trusting it, paying particular attention to the policy-invocation timing and the grasp-disruption problem the authors identify.

The open-source code is available. ApodexAI has released FrontierAgent, the agent framework used in the study, on GitHub under an Apache 2.0 licence, with 5,298 stars and 305 forks as of the repository page P². It runs on macOS and Linux with a single command and supports ReAct and Agent Team modes P². OpenAI's own launch page confirms GPT-6-Astra as the model the study relies on P⁴.

The study's value is less in declaring a winner than in mapping where a frontier model helps and where it does not. The next checkpoint for anyone in this space is whether the three identified failure modes, suboptimal invocation, fixed-instruction execution, and grasp disruption, can be engineered away in a future harness design, or whether they are structural limits of asking a language model to manage a robot it has never been trained on.

Sources: S1 — Profiling a Frontier Agent with EBench: From Direct Control to Harness · P2 — ApodexAI/FrontierAgent · P3 — open-policy-agent/opa · P4 — GPT-6 Astra: A new generation of intelligence | OpenAI

Written from 4 sourced items, 3 of them primary.