Musk: Grok 4.7 trails Opus 5.5, expects to catch up next year
The most candid AI scorecard of the weekend came from the head of one of the labs in it — and he gave it to Chinese state television.
Elon Musk told China Media Group that SpaceXAI's flagship model, Grok 4.7, is "not as good as, say, Opus 5.5 that just got released," and put a timeline on closing the gap: "I would expect that we probably catch up to Frontier sometime next year, most likely." The interview was conducted on 23 September at Tesla's Global Engineering Headquarters and published by CMG's business channel; a full transcript and English coverage both carried the comparison. Musk's explanation for the deficit is tenure: SpaceXAI has been doing AI for about three years, he said, while Anthropic has been at it for six. The admission lands a week after Anthropic shipped Opus 5.5 at a lower price than Opus 5 — the release we covered here — Anthropic's Opus 5.5 matches Fable 5.1 for 40% less than Opus 5 — and two days after SpaceXAI's own Grok 4.7 launch, which we covered in Grok 4.7 ships at 4.6's price — and a hardened safety stack. It is also his second public concession on the same theme in a month: at a Cursor all-hands in August, Musk conceded Anthropic was leading the AI race.
The growth number he offered in the same breath is the one worth keeping: GrokBot, the consumer agent SpaceXAI is pushing into its cars and phones, "has been very popular, and is growing about 100% a month." Musk did not say from what base, which is the whole question — a monthly doubling is either the fastest-growing consumer AI product on the market or a rounding artifact on a small denominator, and he only gave us the rate. His stated plan for catching the frontier is not more parameters but better data: Anthropic made AI excellent at software engineering, he said, and nobody has made AI excellent at real-world engineering, which is what SpaceX's and Tesla's engineering data is for. That is a plausible moat and an unproven one; SpaceXAI has been making the argument for a year without a benchmark attached to it.
On China, Musk was more specific than he has been in public before. Chinese models, he said, are "doing by far the best in terms of performance per unit of compute," and his rough guess is that China closes its lithography and chipmaking constraints "in about 2 or 3 years" — a timeline far shorter than what Western equipment makers say, with Zeiss's semiconductor chief recently putting the EUV gap at 15 years. He also repeated his electricity claim: China's power output passed the United States, Europe and India combined earlier this year. Both are Musk assertions, not measurements, and the second is the kind of number that is easy to check and hard to argue with.
He then asked for a mechanism. "Maybe there could be a working committee to figure out safety," he said, arguing that rules confined to one country cannot govern models built anywhere, and that China and the US are the only two that matter. That is his own proposal, not a policy in motion — but it arrives days after Washington and Beijing set up a standing Super Intelligence Dialogue and an incident hotline, which we covered in Today in AI — September 26, 2026. He also repeated his usual numbers: at least a billion humanoid robots within ten years, 10 billion by 15, 100 billion by 20, with 90% odds on a good outcome and 10% on a bad one, and a stump-speech passage praising Xi Jinping and Shanghai that reads as the price of the venue.
Read the whole interview as soft diplomacy in both directions: a US chief executive praising Chinese manufacturing and Chinese models on CCTV, in exchange for a platform to sell GrokBot into the largest EV market he sells cars in. The safety proposal is the interesting part, and it is the only line in the transcript with no product behind it.
What to watch: whether SpaceXAI publishes anything measurable behind the GrokBot growth rate, and whether Musk's working-committee idea shows up in the bilateral channel the two governments just opened.
If a model vendor tells you it will catch the frontier "next year," what would you need to see to believe it — a benchmark, a shipped product, or an independent eval? Tell us in the comments.
