Grok 5 Release Date: Separating What xAI Said From What Is Circulating

As of 2026-08-20 there is no Grok 5 release date, no Grok 5 on any API we can call, and no xAI announcement naming a day. What there is instead is a large amount of confident writing about a model nobody outside xAI has used. The most-repeated specification — that Grok 5 is a 6 trillion parameter mixture-of-experts model — traces back to no named source at all, while the parameter count actually reported as coming from Elon Musk is 1.5 trillion. That is a four-fold gap between the popular number and the sourced one. This page is a ledger: what xAI has said on its own channels, what is attributed to Musk by named outlets, and what is simply circulating. Then, because waiting is not a strategy, the measured numbers for the Grok you can call today.

Almost every Grok 5 page you will find states its parameter count as fact. Very few say where the number came from. When you chase the citations, the picture changes shape.

The ledger: confirmed, attributed, unsourced

Three tiers, because they are not the same thing and treating them as one is how a rumour becomes a specification.

Note which tier the famous numbers land in. The GPU count, the 10T model, the seven-way parallel training run — the details that make the coverage feel authoritative — are the ones with nobody's name on them.

The 1.5T versus 6T gap

This is the part worth carrying away.

The sourced figure is 1.5 trillion. The reporting says Musk announced in early May 2026 that a Grok base model designated V9-Medium, at 1.5 trillion parameters, had finished training and would reach the public in two to three weeks. V9-Medium is described as an internal xAI version code, and the inference that it corresponds to the product named Grok 5 is the outlet's inference, not an xAI statement. That distinction matters: it is entirely possible V9-Medium ships under some other name, or as one member of a family.

The popular figure is 6 trillion, and it has no author. It appears across many pages, usually stated flatly, sometimes with a 10 trillion variant alongside it. We could not find an original.

So the two candidate answers to how big is Grok 5 differ by roughly 4x, and the larger one is the one with less evidence behind it. If you are making a capability or cost projection off a parameter count, that is not a rounding error.

What happened to the release windows

The two-to-three-week window from early May would have closed in late May 2026. It is now 2026-08-20, more than three months past that, with no release. Read that as a data point about how to weight the next forecast, not as a criticism of the model: training timelines slip everywhere, and a model that is late is not thereby bad.

What xAI did ship while Grok 5 did not

Focusing on the model that has not arrived hides the fact that the lineup kept moving. On OpenRouter as of 2026-08-20 you can call x-ai/grok-4.20 and x-ai/grok-4.20-multi-agent at $1.25 / $2.50 with a 2,000,000-token context window, x-ai/grok-4.3 at the same price with a 1M window, and x-ai/grok-4.5 and x-ai/grok-4.6 at $2 / $6 with 500K.

That 2M context figure is worth pausing on: it also circulates in the unsourced tier for Grok 4.20, and it happens to be one of the few such claims that is independently checkable. It checks out. Unsourced does not mean false — it means unverified until you verify it, which is a different job from repeating it.

Note also that Grok 4.6 is already live and got almost no coverage, while a model that does not exist got a great deal. If you want the current pricing picture across the family, our Grok API pricing page tracks it, and our note on price volatility explains why any figure here needs a date attached.

The Grok you can call today, measured

These are our numbers, from our executed nine-task Python benchmark: generated code run against hidden asserts in an isolated subprocess, temperature 0, cost derived from measured token counts at list price. Full method on the methodology page.

For context on what that price buys elsewhere, GPT-5.4 mini also went 9/9 on the same tasks at $0.53 per thousand and 2.3 seconds, and DeepSeek V3.2 went 9/9 at $0.08. Grok 4.5 is a capable model that is not competing on price. Our Grok 4.5 review has the full run, and the message limit page covers the consumer-tier quotas that catch people out.

We have not benchmarked Grok 4.6 yet. It is on the list.

What nobody can tell you, including us

• The release date. There is no announced day. Anyone giving you one is forecasting.
• The parameter count. 1.5T is the sourced figure and it is attached to an internal codename, not to the product name Grok 5.
• The benchmarks. There are no Grok 5 scores because there is no Grok 5 to score. Any table you see comparing Grok 5 to other models is projection.
• What we did not do: we did not see Musk's original posts. Everything in the attributed tier above is us reading named outlets reporting on them, on 2026-08-20.

When Grok 5 is callable it goes through the same nine tasks as every other model here and we publish the result, good or bad.

FAQ

When is Grok 5 coming out? Unannounced as of 2026-08-20. Two forecast windows — a September 2025 training start and an early-May 2026 estimate of two to three weeks to public release — have both passed.

How many parameters does Grok 5 have? The figure with a source behind it is 1.5 trillion, for a base model designated V9-Medium. The widely repeated 6 trillion figure has no named source we could find.

Is Grok 5 available on any API? No. As of 2026-08-20 the newest Grok models callable on OpenRouter are Grok 4.5 and Grok 4.6.

What is the best Grok model right now? Of the ones we have measured, Grok 4.5 — 9 out of 9 on our executed tasks at $2.93 per thousand. We have not yet measured Grok 4.6.

Will Grok 5 be AGI? Musk is reported to have described a rising probability of it. That is a statement of belief by an interested party, not a result.