Grok V9 'done training' — what Musk's claim is actually worth

There is a difference between a model exists and you can use it, and xAI announcements have a habit of living in the gap between the two. On 25 May, Elon Musk said a new foundation model — V9-Medium, reportedly around 1.5 trillion parameters, roughly triple the current production model — had finished training, with internal evaluations described as positive. A mid-June release was floated in the coverage that followed. By mid-June: no firm date, no model card, no independent benchmarks. The gap is still the gap.

In the public discourse, 'finished training' tends to be conflated with 'released' or 'ready'. It isn't. Training is the first of several phases: the model learns from data, which is compute-intensive and time-bounded — so 'training complete' is a real milestone. What follows is typically weeks or months of evaluation, safety testing, fine-tuning, and the engineering work of turning an artefact into a product. For a model at the scale Musk described — 1.5 trillion parameters — that pipeline is not trivial. A 'finished training' statement says the learning phase is over. It says nothing about whether the resulting model is coherent, safe, aligned, or better than what's already deployed.

The V9-Medium announcement exists alongside at least one other unshipped major model: Grok 5, which has been hyped for considerably longer and which, per Fello AI's tracker as of early June, remained in training with prediction markets openly skeptical of a June ship. That creates an unusual positioning: a company promoting two separate unreleased models simultaneously — neither of which has an independent benchmark to point to — while the current flagship, Grok 4.3, is the model developers can actually access today.

The parameter count of V9-Medium deserves a note. 1.5 trillion parameters would place it in rarefied company — roughly the scale of the largest reported frontier models — but parameter counts are not reliable proxies for real-world performance. A well-trained, smaller model often outperforms a larger one on the tasks that matter. The claim is notable as a signal of intent and investment; it is not a quality claim without benchmarks to support it.

None of this makes V9-Medium fiction. xAI is a real lab that ships real models: Grok 4.3 exists, Grok Imagine 1.5 Preview shipped in June, and the company moves fast. The pattern being observed is not capability fraud — it is a consistent habit of announcing timelines that compress in the press release and expand in reality. Prediction markets, which aggregate the expectations of informed participants who have financial skin in the game, were giving a June ship for Grok 5 notably short odds as of early June. That skepticism is a useful calibration, not a verdict.

The actionable read for anyone making model selection or product decisions: V9-Medium is an announcement about an internal training run, not a shipped option. The model to evaluate today is Grok 4.3. When a model card appears — with real benchmark numbers, an API endpoint, and third-party independent testing — that is the moment to form an opinion on V9's quality and fit. Until then, the announcement is worth noting in a timeline and worth not building a roadmap on.