Musk concedes Grok trails Anthropic by years — while the 100%-a-month agent claim measures distribution, not capability

Elon Musk has done something unusual for a frontier-AI executive: he conceded the gap. In an interview with China Media Group, Musk said that Grok 4.7, the latest publicly available model from his AI company, is "a solid workhorse of a model" but not as good as Anthropic's recently released Opus 5.5. His explanation was chronological rather than technical: xAI has been "around for three years during AI, whereas Anthropic's been around for six years." In the same conversation he said his agent product, Grok Bot — described as a personal digital assistant — is "growing about a 100% a month."

The admission is not a one-off. According to the report, Musk said in mid-September that the unreleased Grok 4.8, which he says relies on 2.5 trillion parameters, would improve performance significantly but still would not be at par with OpenAI's Astra and Anthropic's Fable class of models. His stated timeline for closing the distance: "we probably catch up to frontier sometime next year, most likely." Worth translating one number for readers: a parameter count is an input — roughly, the size of the model's adjustable weights — not an outcome. Bigger models are more expensive to train and run; they are not automatically better. We have consistently flagged Grok's quality scores as estimated rather than measured, precisely because leaderboard Elo can reward models tuned to be likeable rather than correct.

Now the other number. "100% a month" means doubling. Compounded for a year that is roughly a 4,000x increase, which is a useful way to see why no serious product sustains that rate for long. More important: no absolute base was given. A growth rate without a denominator tells you the slope, not the size — doubling from 50,000 users and doubling from 50 million are the same percentage and wildly different businesses. And note what it measures: adoption, not capability. That is entirely consistent with a thesis we have argued before — the competitive front has moved from "who has the smartest model" to who controls the plumbing and the distribution surface where assistants actually reach people. A model that trails the frontier can still win the habit.

Musk's substantive bet is the most interesting part and the least testable. He argues Anthropic has built an AI "excellent at software engineering" but that no one has built one "excellent at real world engineering," and that data from SpaceX and Tesla gives him the path to it. That is a genuinely differentiated position — physical engineering is badly served by today's text-and-code benchmarks — but it is also, for now, aspiration rather than demonstrated capability, and there is no accepted way to score it. A related report cited by the same outlet says Musk is working to bring some 1.2 million NVIDIA GPUs online in a mega data center. Compute is being bought at extraordinary scale; the results are still a promise.

On China, Musk offered a forecast that deserves an asterisk. He said China will "address the compute limitation faster than most people expect," with a rough estimate of two or three years to solve lithography and chipmaking. Frank Rohmund, semiconductor CEO at Zeiss — which makes the high-end optics inside EUV lithography machines — has said publicly that China may be 15 years behind the West on the latest EUV equipment. Two interested parties, two numbers an order of magnitude apart, one of them delivered to Chinese state media. Neither should be taken as settled fact.

**Our reading:** the most reliable signal in this interview is the concession, not the growth rate. Admitting your model trails a rival costs a CEO something; quoting an unaudited monthly percentage costs nothing. Read accordingly. The strategic picture is that xAI is buying its way toward the frontier with capital and compute while betting its differentiation on physical-world data — plausible, unproven, and impossible to verify until there is a benchmark for it. The China question matters less than the framing suggests: Chinese open-weight labs are already competitive on the silicon they have, which is the part of the story that has actually been measured. And the long view stays intact. A third well-funded contender that is honest about trailing is better for everyone than a second one that pretends otherwise — competition at the frontier is what drags the price of capability down, and cheap capability is the mechanism by which AI eventually reaches medicine, materials and everything else that makes abundance more than a slogan. The near term is noisier claims; the direction of travel is still good.