In a controlled benchmark published on September 4, 2026, the research group Robocurve put OpenAI's GPT-6 Astra in control of a pair of robot arms and tested it on two manipulation tasks, pitting it against Anthropic's Claude Fable 5 and Fable 5.1 under identical conditions. The result splits cleanly in two: on a simple pick-and-place task the newer model is dramatically better; on a precision insertion it is not better at all.
The hardware was a bimanual I2RT YAM arm setup — two six-degree-of-freedom arms, each fitted with a parallel-jaw gripper. Rather than training a model end-to-end on raw pixels, the team used the Inspect Robots harness (version 0.58.0). In that setup the model acts as an "agent policy": it issues absolute end-effector poses through a move_to command — x, y, z, yaw, pitch, roll, plus the gripper state, for each arm — and the robot's inverse-kinematics (IK) layer converts those poses into joint angles. Each model received 20 trials per task, was allowed up to 20 large-language-model (LLM) calls per run, was capped at 25% of the robot's speed, and ran with the harness's default safety guardrails. The observation consisted of three camera views — overhead, left wrist, right wrist — plus the robot's proprioceptive state. Every trial was graded by a human on a five-stage rubric running from "no purposeful approach" (0) to "placed in its final position" (4); a run that fails still records how far it got.
On the simpler task — "pick up the red block from the table and place it inside the bowl" — GPT-6 Astra placed the block in 19 of 20 trials, a 95% success rate with a mean stage of 3.95. Claude Fable 5.1 managed 8 of 20 (40%, mean 2.40); the earlier Fable 5 managed 1 of 20 (5%, mean 1.30). Astra was also faster and cheaper: an average 2.5 minutes and an estimated $0.94 per run, against Fable 5.1's 6.8 minutes and $2.12, and Fable 5's 8.2 minutes and $2.69. It also used far fewer output tokens — about 2,100 per run, compared with roughly 12,900 for Fable 5.1 and 19,200 for Fable 5.
The second task — "pick up the round blue puzzle piece by the knob at its center and place it into the matching circular groove in the board" — tells a very different story. Astra completed the insertion only 2 times in 20 trials (10%), essentially matching Fable 5.1's 2 of 20 and beating Fable 5's 0 of 20 only on a technicality. Its mean stage was 2.00, actually below Fable 5.1's 2.35. The report's own wording is blunt: Astra "reaches the groove and stalls at the same final step Fable does." All three models could maneuver the piece to the opening; all three then got stuck on the same last motion — the precision insertion itself. Astra did it more cheaply ($1.36 per run, about 2,700 output tokens, 3.4 minutes, against Fable 5.1's $2.18, 10,500 tokens, 5.9 minutes), but cheap stalling is still stalling.
The chart visualizes the article's central finding: the models diverge sharply on the simple pick-and-place task but stall together on precision insertion.
That stall is the most informative part of the benchmark. A coarse pick-and-place task stresses spatial reasoning and grasp selection; a knob-into-grove insertion additionally stresses fine positional control and the ability to feel for alignment. The fact that a newer, stronger model reaches the same bottleneck as its predecessors — and with a lower mean stage, because it more often failed before the final step — suggests the barrier lives in the last centimeters of manipulation, not in getting the arm into the right general area. What exactly causes the stall, the report does not say. It could be a limitation of the model, of the move_to interface (which asks for a single absolute pose rather than letting the model feel its way in), or of a setup in which the model never receives tactile feedback. The benchmark only establishes that under these conditions, the insertion step remains unsolved.
The authors were candid about the study's constraints, and they shape how far the results should be generalized. Astra's 20 trials per task were run two days after the Fable trials and were not interleaved with them, so nothing rules out drift in the setup over time. The bowl comparison is not apples-to-apples: the Fable bowl trials ran on rig-3, which was unavailable for Astra, so Astra's bowl runs happened on rig-1. The puzzle comparison, at least, was on the same rig for all models. Scoring was operator-judged with the model known to the grader, which opens the door to unconscious bias toward the newer model. Costs are list prices; Anthropic's requests were sent without prompt caching, while OpenAI cached roughly a fifth of Astra's input automatically, and that discount was not applied — so Astra's cost figure is, if anything, an overstatement. Objects were reset by hand between trials, and every model ran at medium reasoning effort only. Twenty trials per model per task is a small sample, and two tasks is a narrow slice of manipulation; the report does not claim more than that.
The cost figures are worth pausing on, because they are easy to misread. At list prices, the cheapest Astra run cost about $0.94 and the dearest Fable 5 run about $2.69 — for a task a human could perform in seconds. Even taking the numbers at face value, putting away a block at a dollar-odd per attempt is not competitive with any plausible labor cost, and the per-run figure conflates token price, reasoning effort, and the harness's fixed overhead. What the numbers do show clearly is the newer model's relative efficiency: roughly 2,100 output tokens per run on the bowl task against about 12,900 and 19,200 for the two Fable models. Token throughput, though, is a proxy for controller cost inside one specific harness — not a forecast of deployed-system economics, and not a measure of physical capability.
One more thing the benchmark does and does not test. The setup is an LLM agent driving a robot through a symbolic interface — poses, not pixels-to-joints. This style, in which a general-purpose model plans at a high level while a conventional controller handles the low-level geometry, is sometimes called "code as policy," and it is a deliberate choice. It also means the result says nothing about end-to-end vision-language-action models, which map sensor input directly to joint commands and were not part of this comparison. The puzzle failure is therefore evidence about LLM-planned manipulation under this particular harness, not about robotic capability in general. The report's technical appendix lists the interface precisely — move_to with x, y, z, yaw, pitch, roll and gripper per arm, the robot's IK converting poses to joint angles — so the setup is reproducible, and whether a different interface closes the insertion gap is a question this experiment leaves open.
So what can be said confidently? GPT-6 Astra, driven by an agent policy through a standard manipulation harness, is markedly better than the two Fable models at a simple pick-and-place task, and not better — at least not yet — at a precision insertion that defeats all three models. The margin on the simple task is large; the gap on the hard task is zero. Whether the insertion problem yields to a better interface, more trials, tactile sensing, or a different model class is, for now, an open question this benchmark does not answer.