Musk says Grok V9 is trained. We've heard this tune.

On 25 May, Musk said V9-Medium — reportedly ~1.5 trillion parameters, ~3× the current production model — had finished training with 'positive' internal evals. A mid-June release got floated in the tech press. By mid-June: no firm date, no model card, no independent numbers. The counter on 'announced but not shipped' xAI models is now at two: V9-Medium and the much-louder-hyped Grok 5, which as of early June remained in training and whose prediction-market odds of a June ship were unimpressive. The model developers can actually use is Grok 4.3. File that somewhere useful.

Let's be clear about what kind of skepticism this is. It is not 'xAI doesn't ship real models' — they do. Grok 4.3 exists. Grok Imagine 1.5 Preview shipped in early June. The lab moves fast by industry standards. The specific pattern being called is announced-timeline slippage, which has a track record, combined with unverified internal claims, which is the least informative category of AI announcement. 'Positive internal evals' is the most unfalsifiable phrase in the business — every model has positive evals before it goes out the door, because that's what you need to ship it. It is not a signal of anything external.

The 1.5 trillion parameter claim is worth a separate beat. That would make V9-Medium a genuine frontier-scale model — comparable to the largest reported models from any lab. If accurate, that's a real investment. But parameter counts are the vanity metric of the AI era: they tell you what the training run cost, not whether the model is good. A well-trained 70B model regularly beats a sloppily trained 500B on the benchmarks that matter. Until there's a benchmark you can read and an API you can hit, '1.5 trillion parameters' is a press number, not a quality claim.

What makes the V9-Medium situation stranger is the context: xAI is simultaneously promoting two major unreleased models. V9-Medium (25 May, 'done training') and Grok 5 (ongoing, longer hyped, even bigger claimed ambition). Running parallel announcements for unshipped models is an unusual communications strategy — it doesn't obviously serve the developer community, who want to know what they can build with today. What it does serve is a narrative of relentless forward momentum, which has its own value in recruiting and funding circles even if it's frustrating if you're trying to decide which model to integrate.

The grown-up move is to note this and move on. If you're making a build decision: Grok 4.3 is what exists. When a model card appears for V9-Medium — with third-party benchmark numbers, a safety evaluation summary, and an API endpoint you can hit — cover it then, because at that point you have something to evaluate. Not the parameter number. Not the 'positive' internal eval score. The actual performance on real tasks. Right now we're covering a claim. A claim from a credible lab with real infrastructure and a real engineering team, yes — but a claim. The test is in the artefact, not in the tweet. Come back when the card drops.