Peter Diamandis opened the robotics segment of Moonshots with one number. The cost for a frontier AI model to complete a robot subtask, he said, fell from $2.36 in July to $0.38 in September, an 84 percent drop. Then he projected it forward: halving in cost every three weeks "means it effectively goes to zero, you know, within the year."
Diamandis, the founder of XPRIZE and Singularity University, recorded the episode on October 16, 2026, with Salim Ismail, Emad Mostaque and XPRIZE chief executive Anousheh Ansari. It was published the next day.
A frontier model is one of the most capable general-purpose AI systems from the leading labs. In this setting it does not write text for a person. It looks at camera images and sends movement commands to robot arms, one step of a physical chore at a time. Diamandis credited the figures to RoboCurve, which he described as a Y Combinator startup, and to its benchmark, RoboDojo RC Tier 1, which he said measures how well a model can drive a physical robot through actual subtasks.
The arithmetic in his claim holds up. A drop from $2.36 to $0.38 is 84 percent, which is a little more than two and a half halvings. That works out to roughly one halving every three weeks if the two measurements were about two months apart.
What RoboCurve's published test shows
RoboCurve's September 23 evaluation does not report the $2.36 and $0.38 figures. It describes a narrower comparison. The benchmark tier is made up of six adapted tasks, picked as the easiest of 18 tested with a model called Astra: handling fruit, stacking bowls, storing an item safely, capping a pen, standing a bottle up and sorting objects. Two-armed robots receive camera views, and human graders award partial credit from zero to one depending on how far the robot got.
Over 120 trials each, Astra averaged 36.7 percent progress and fully completed five tasks. Opus 5.5 averaged 36.0 percent and fully completed one. Their estimated inference costs, meaning the cost of running the model to produce its decisions, were $1.14 and $0.90 per attempt. Those costs leave out the cost of operating the robot, and Astra's figure assumes a caching discount applied after the fact to recorded conversations. RoboCurve also noted that the operators knew which model they were grading and that resetting objects by hand introduced variation.
The page therefore shows two models scoring about the same on a set of easy physical tasks, with one about 21 percent cheaper per attempt. It does not chart a trend over time. The three-week halving rests on the two figures Diamandis quoted, extended into the future.
"Blows that out of the water"
Salim Ismail, founder of Open ExO, set the figure against the AI capability trend the hosts had been following, which he described as doubling every 10 weeks and already "mind-boggling." Halving every three weeks, Ismail said, may be the fastest-moving capability they had seen. He compared it with the 18-month doubling of Moore's Law, the long-running observation that chips gain transistors at a steady pace: "This blows that out of the water."
The case that generalist models win
Diamandis turned to Emad Mostaque, founder of Intelligent Internet, who had sent him the story. Mostaque framed robotics as a series of open questions. First it was whether robot bodies were good enough, and they seemed able to do "just about anything." Then the hands got good. The remaining question was the mind: whether robots need specialist models built for particular jobs, or whether general-purpose models can pick up new skills. "This shows that generalist models can actually win," he said.
The answer went on to mention that the chief executive of the robot maker Unitree had just said the company was no longer sure it needed specialist models. It recalled Unitree's early robots "stumbling about like little drunkards" and said they now do kung fu and synchronized routines. If the cost of the model's thinking keeps halving this fast, the argument ran, there is effectively "no cost for robotic intellect and robotic locomotion," and what is left is the cost of hardware, which would lead to "an explosion" in robots.
The same line of argument held that SpaceX and Tesla do not build customized robotics models and have chosen to "play the generalization game." It pictured one robot being an engineer, a lawyer and a plumber, on the reasoning that "humans are generalized. We're not like lobotomized and specialized." It left room for very specialized machines but called humanoid robots "a different game."
Diamandis added that he had discussed robot surgery with Elon Musk at the start of the year, and that Musk predicted robots would do perfect surgery within three years. He also said the show was due to interview Brett at Figure, the humanoid-robot company, in a couple of weeks. Figure, he recalled, started building its own robot models because the early large language models did not work, and he wanted to ask what Figure uses now.
Advice for hardware founders
Diamandis then asked what an entrepreneur in robot hardware should be thinking. The answer was to build as many robots, actuators (the motors that move a robot's joints) and spare parts of the robotic supply chain as possible.
When Diamandis asked whether founders should also collect as much training data as they could, the reply was that data would matter for many specialized scenarios. The open question was who buys it. One possibility was robotics companies themselves, with Figure's billion-dollar commitment as the example. The other, favored in the reply, was that every major foundation-model company, having hoovered up the world's text, would now want all the robotics data too. The reply added that robotics data is an order of magnitude larger than language data, and that embodied intelligence needs it.
Figure's own August 25 announcement puts the commitment at more than $1 billion over twelve months for data and compute together, not data alone. Its Index platform collects videos of people doing physical tasks such as cooking, laundry and stocking shelves. After four months of private operation, Figure reported more than 16 million uploaded videos, over 44,000 weekly active contributors and $15 million paid to contributors.
Ansari asked whether the performance depends on the robot's shape, or whether it would work on, say, three-armed robots. The response was that this test covered arms and humanoid-type machines in specific settings, but that the results should generalize quickly. Language models got good enough that people no longer have to "baby" them, the reply said, and robots were about to get there too. RoboCurve's published trials used two-armed robot setups.
The video precedent
The discussion then turned to whether the curve was a surprise. After Diamandis asked, the answer was no: robotics, it was argued, would "suddenly be cooked," just as video generation was. High-quality AI video, the argument went, had gone from about 50 cents per second to 0.1 cents per second with new models due to be released the following week, faster than real time. That is a 500-fold drop, and the lower figure is a projection for models that had not yet shipped.
Once a model crosses the threshold where "it can do the work," the argument continued, the price collapses, and the job becomes building the actual robots. It pointed to a claim from an earlier episode that Musk is building a million Optimus humanoid robots for next year, rising to 10 million, and that at 1.7 million units Optimus would overtake Tesla in revenue. "And it's still not enough," the discussion concluded.