OpenAI’s GPT-6 Astra can now control robots well enough to approach a perfect score on easier robot benchmarks, but a single long simulated task consumed “roughly nine million tokens and nearly two hours,” according to a research paper posted to arXiv on 15 October 2026 by a SenseNova joint research team. The paper, titled “Has GPT-6-Astra Achieved Embodied Intelligence?”, tested the model across 12 robot benchmarks and 8 real robot configurations.
Its own answer is no, not yet. The authors conclude that GPT-6 Astra “has largely crossed the threshold of embodied competence, but not yet that of practical embodied intelligence.” The paper is a preprint and has not been peer reviewed, and its authors are a SenseNova research team, not OpenAI.
Business Pill · TOKENS PER WATT
A short Business Engineer explainer on why AI costs are measured in tokens and power. It teaches the general idea only and says nothing about OpenAI, SenseNova or the paper in this piece.
The key insight: As we read it, the number that travels is not the benchmark score but the bill: the paper says a near-solved robot benchmark still cost “about a million tokens and nearly an hour” per episode, and a harder one “roughly nine million tokens and nearly two hours.”
The team ran GPT-6 Astra, without retraining it, as the brain of simulated and real robots: robotic arms on tabletops, humanoids, mobile platforms and dexterous hands. Under the paper’s protocol, the model could learn in context from successful demonstrations in each benchmark’s training set, and it got up to 10 times the usual number of action steps to finish a task.
On that basis the paper reports that GPT-6 Astra “approaches saturation on the LIBERO series and SIMPLE, while achieving 81.9% on RoboDojo and 87.3% on DexJoCo.” A benchmark is saturated when the best systems score close to its maximum, so the test stops telling them apart.
GPT-6 Astra controlling simulated robots: about a million tokens per episode on LIBERO-Long++, where success already exceeds 97%, and roughly nine million tokens and nearly two hours on LIBERO-Ultra-Long. Source: arXiv 2610.22471, Table 4.
What Was Posted on X
New paper finds that GPT-6-Astra nearly saturates easy robot benchmarks but still fails to verify physical success and recover from irreversible mistakes. And at a cost too high for real use.
Robot manipulation and navigation have been dominated by task-specific models. The question was whether 1 general model could handle them all without weight updates.
The authors tested GPT-6-Astra on 12 robot benchmarks. It studied training demos into a frozen knowledge package and got up to 10 times the usual action steps to recover.
It neared saturation on the LIBERO suite and reached 81.9% on RoboDojo and 87.3% on DexJoCo. But a single LIBERO-Ultra-Long episode used about 9M tokens and nearly 2 hours, and on real robots it could not trade more steps for success.
Treat general models as capable robot controllers for research, but expect task-specific demos, heavy compute, and physical limits to block deployment for now.
Rohan Paul posted about the paper, arXiv 2610.22471, on X on 22 October. His post says GPT-6-Astra nearly saturates easy robot benchmarks but still fails to verify physical success and recover from irreversible mistakes, at a cost too high for real use, and cites 81.9% on RoboDojo, 87.3% on DexJoCo and about 9M tokens and nearly 2 hours for one LIBERO-Ultra-Long episode. The piece checks every figure against the paper itself.
OpenAI introduced GPT-6 with 47-year-old footage of a guy at MIT with a sensor from a military helicopter helmet strapped to his wrist, just so a computer could figure out what he was pointing at.
The room he’s sitting in held about a million dollars of equipment. All of it was there to teach the computer two words, “that” and “there.”
The project was called Put That There. Chris Schmandt and Eric Hulteen built it in 1979 with Richard Bolt at MIT’s Architecture Machine Group, the lab that later turned into the Media Lab.
Here’s how it worked. You sat in a leather Eames chair in a soundproof room, facing a screen as big as the wall, lit from behind by one of the first video projectors ever built. The sensor on your wrist tracked where your hand was aimed while a voice recognizer listened. That recognizer, an NEC DP-100, cost $100,000 by itself, which is over $400,000 today.
So you’d say “put a yellow circle there,” point at the wall, and a yellow circle showed up right where you pointed.
That sounds easy until you think about what the computer actually has to know. “That” and “there” don’t mean anything by themselves. When you say them to a friend, they look where you’re looking and get it instantly. The computer in 1979 couldn’t see a thing, so the MIT team basically lent it your arm. Your voice told it what to do and your hand told it where, and it put the two together on the fly.
If you forgot to point, it would ask you “Where?” and wait. Schmandt says it was one of the first two conversational systems anybody ever wrote.
A year later they swapped the shapes for a map of the Caribbean, and the lab’s military sponsors would sit in that chair pushing battleships around Cuba.
Then the actual computer business went a different way. Voice recognition stayed too unreliable for everyday use for decades, so we ended up learning the computer’s language instead, getting used to menus and folders and dragging a little arrow around with a mouse. Pretty much every screen you’ve used since the 1984 Mac assumes your hand is doing the pointing.
Now look at what OpenAI bragged about. Astra set a new record on a test called ScreenSpot Pro, which checks whether an AI can find the exact button you mean on a crowded screen full of pro software. Schmandt and Hulteen were attacking that same problem with a $100,000 microphone and a helicopter sensor.
This time the computer is the one moving the cursor.
For 47 years we built every computer around the human hand, because no machine could handle a sentence as simple as “put that there.” Astra’s whole pitch is that you can go back to just saying it.
Aakash Gupta (@aakashgupta) notes that OpenAI introduced GPT-6 with footage of Put That There, a 1979 project at MIT’s Architecture Machine Group in which a wrist sensor and a $100,000 NEC voice recognizer let a computer work out what “that” and “there” meant when a user pointed at a wall-sized screen.
The Cost of a Robot Episode
The paper is blunt about the price of that performance. “Strong performance comes at a substantial cost in tokens and wall-clock time, even on benchmarks that GPT-6-Astra nearly solves,” it says. On LIBERO-Long++, where success at the standard budget already exceeds 97%, “each episode still consumes about a million tokens and nearly an hour of inference and tool use.”
On the harder LIBERO-Ultra-Long, “these costs rise to roughly nine million tokens and nearly two hours,” and on RoboDojo, “a single episode under the extended budget takes more than three hours on average.” The authors’ rule of thumb for cost is that 1M tokens costs about 1 USD, because roughly 98% of the tokens were cached input. On that rule, one Ultra-Long episode would cost roughly $9 (our arithmetic).
The paper also says “even simple tasks incur inference cost and execution time far beyond what would be practical for commercial deployment,” and that because thinking time is not counted in the step budget, the cost “also limits deployment in settings that require real-time control.”
Where It Fails
The paper lists the model’s recurring failure modes. Among them:
It “often struggles to verify physical success and identify which task condition remains unmet.” In one puzzle task it reported the pieces were in place while its corrections pulled them apart again.
It “cannot reliably recover from consequential or irreversible failures,” such as a spill or a dropped part.
“Long-horizon execution remains difficult, as the model struggles to preserve and track progress across many subtasks.”
The demonstration packages it learns from are “largely task-specific, which limits transfer” to new tasks.
On Real Robots
The real-world tests ran on 8 robot configurations, from single arms to a humanoid. The same model drove all of them without platform-specific training, the paper says, but “what limits success is the nature of the task.”
Task type on real robots
Successes / trials
Platform(s)
Rigid pick-and-place
17 / 18
not broken out in the finding
Cloth folding
2 / 4
Franka 3 arm
Navigation and search
9 / 23
AgileX Tracer
Plug insertion
0 / 11
three platforms
Source: arXiv 2610.22471, section 4.2 (“Finding 1”), real-world tests on 8 robot configurations. Counts as given in the paper.
The abstract adds two gaps for real deployment: “robot demonstrations required for in-context learning are expensive to collect,” and “it is implausible to trade longer execution for higher success rates with physical constraints.” Its failure analysis makes the same point: an early mishap, “such as a spill or a dropped part, cannot be undone.”
Connecting the Coverage
The paper lands in a week of robotics claims. Our Optimus piece reported Elon Musk’s stated expectation of 20,000 Optimus robots a week, which Axios reported as a year or maybe a little longer away. Our Cosmos piece covered Nvidia citing Porsche’s expectation of up to 90% fewer repeated physical test drives through simulation.
The cost finding echoes our cost-to-outcome piece, in which Satya Nadella said customers will “measure the cost of tokens against the outcomes they produce.” Our KLS conjecture piece reported Microsoft saying unguided AI runs on roughly the same token budget fell short of the goal, while a human-directed process succeeded.
What It Means for Businesses
As we read it, the paper separates two questions that are often merged: whether a general AI model can control a robot at all, and whether it can do so cheaply, quickly and reliably enough to deploy. The paper answers the first mostly yes and the second not yet.
As we read it, the real-robot table is the practical guide: rigid pick-and-place worked in 17 of 18 trials in the paper’s tests, while precise contact work, such as plug insertion, did not succeed once in 11.
The Structural Read
As we read it, the paper separates capability from deployability: the model can drive many robot types without retraining, while the time, the tokens and the physical failure modes it describes keep it out of production.
As we read it, the step budget explains the gap between simulation and the real world: in simulation, extra steps can fix a mistake; the paper says that on real robots it is “implausible to trade longer execution for higher success rates with physical constraints.”
As we read it, the real-robot table is a ranking of task difficulty: rigid pick-and-place worked in 17 of 18 trials, while plug insertion, which needs precise contact, worked in none of 11, per the paper.
arXiv 2610.22471, SenseNova joint research team
“GPT-6-Astra has largely crossed the threshold of embodied competence, but not yet that of practical embodied intelligence.”
Three Implications
FOR COMPANIES EVALUATING ROBOT AI As we read it, the paper suggests asking for cost per completed task and wall-clock time alongside success rates; it gives both, and they change the picture.
FOR ROBOT MAKERS As we read it, the paper’s finding that a general model drove 8 different robot configurations without platform-specific training points to the model as a shared control layer, with the limits set by the task, not the hardware.
FOR BUYERS OF AI GENERALLY As we read it, this is the cost-to-outcome question in physical form: the paper measures the tokens and hours spent per episode against whether the task was actually done.
The framework’s core line: there are two curves, “what your underlying technology can do” and “what your users can actually do with it,” and “the gap between them — what the platform makes possible versus what the product surfaces — has a name.”
As we read it, the paper measures that gap for robots: competence on benchmarks on one curve, cost, time and physical reliability on the other. The framework was published in June 2026 and does not discuss this paper.
What Is Not Established
The paper is an arXiv preprint and has not been peer reviewed. It was written by a SenseNova joint research team, and we did not contact OpenAI or the authors. Its dollar figures rest on the authors’ own rule of thumb rather than on an invoice. We read the abstract, the introduction, the limitations section and the real-world findings, not every appendix table.
Business Engineer Framework
The Product Overhang Doctrine
A Business Engineer read on the gap between what a technology can do and what a product actually delivers to users.
A research team found GPT-6 Astra close to saturating easier robot benchmarks, but one long simulated task took roughly nine million tokens and nearly two hours, and on real robots precise contact tasks still failed. The authors’ verdict: embodied competence, not yet practical embodied intelligence.
A note on sourcing. Quotations and figures come from arXiv preprint 2610.22471, posted on 15 October 2026 by a SenseNova joint research team: its abstract, introduction, limitations section and real-world findings. The roughly $9 per episode is our arithmetic from the authors’ own rule of thumb. Other figures come from our earlier pieces. The paper has not been peer reviewed; we did not contact OpenAI or the authors. Nothing here is investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
