| Date | 24 Sep 2026 |
|---|
After Grok 4.7’s mixed reception, the SpaceX chief points to a faster training cadence, a younger AI effort and the company’s record bringing compute online.
Elon Musk said Thursday that SpaceX could field an artificial intelligence model on par with Anthropic’s Fable line and OpenAI’s GPT-6 family within two to three months.
“I am cautiously optimistic that SpaceX will have a Fable/GPT-6 level model in 2 to 3 months,” Musk wrote on X, replying to a post that called last week’s Grok 4.7 release a stumble and said the company had rushed it under pressure from rivals.
The remark landed three days after SpaceXAI, the company’s AI division, shipped Grok 4.7. That model is priced at $2 per million input tokens and $6 per million output tokens and is aimed at coding and longer knowledge-work tasks. Independent and company-reported scores put it ahead of some competitors on a handful of software-engineering and electrical-engineering tests and behind them on others. Several developers who had used earlier Grok versions as daily drivers said the new cut did not close the gap they expected.
Musk answered a follow-up that declared the race already lost.
“1. We will keep accelerating. Our AI efforts are only 3 years old, vs 6 and 10 years old for Anthropic and OpenAI. If our second derivative remains strong, SpaceX will reach pole position in about 6 months.
- Once you far exceed the caliber of intelligence needed for a class of tasks, additional intelligence is pointless. You don’t need (and it would be cruel to put) Newton-level intelligence in your toaster!
- Hardware is hard. Bringing massive compute online rapidly is incredibly difficult. SpaceX has demonstrated exceptional ability in this regard and will only get better.”
SpaceX absorbed xAI in February. The Grok models, the Memphis Colossus cluster and related training work now sit inside the rocket company. Cursor, the coding tool SpaceX agreed to buy for $60 billion, feeds both data and engineers into the same pipeline. Musk has said for months that new foundation models trained from scratch would land on a monthly cadence. That timetable has slipped in public view more than once.
For executives buying inference at scale, the hardware line is the part that travels. Training runs at this size live or die on whether chips, power and networking show up on schedule. SpaceX has spent a decade iterating factories and launch cadence. Musk is applying the same argument to data centers: the lab that can turn silicon on fastest sets the pace, not the lab that published first.
Developers reading the same posts will notice the toaster line. It is a bet that most production workloads do not need the next full generational jump. A model that is “good enough,” cheaper and faster to call can still win the daily work. That is the slot Grok 4.6 and 4.7 have tried to occupy. The two-to-three-month target is Musk saying the next cut is meant to leave that slot.
None of the posts included benchmarks, parameter counts or a ship date firmer than the window he gave. Anthropic’s Fable 5 and Fable 5.1 and OpenAI’s GPT-6 Astra, Sol and Luna are already in customers’ hands. Those labs have longer product histories. Musk’s public case is that the slope of improvement, not the starting year, is what matters now.
Hardware remains the constraint he keeps naming. Power, packaging, memory and the time it takes to stand up a cluster do not compress as neatly as a software release note. SpaceX’s claim is that it has already shown it can do that work at unusual speed and that the same muscle now sits under the model factory.
What executives and engineering leads can do with the comments is limited and specific. Watch the next Grok drop against the two-to-three-month clock. Compare price-performance on the tasks that actually run in production, not the leaderboard screenshot. Treat compute delivery dates as first-class risk, the same way launch windows already are. And do not assume the next model from any of the three labs will be the last word. Musk said the company will keep accelerating. The posts do not say by how much.