Grok 4.8 Leak: The 2.5T Model Musk Just Named Is Finishing Training This Week

Grok 4.7 still does not have a model page on the x​AI developer docs, and on September 13 Elon Musk was already talking about the model after it. In a post on X he said that Grok 4.8, a 2.5-trillion-parameter model trained on x​AI's new C++ software stack, "will finish training this week and start RL," and roughly twelve hours later he added that it "will be a noticeable improvement" over the still-unreleased Grok 4.7. This is a what-we-know-so-far piece, so the framing is deliberate: Grok 4.8 has not been released, has no model ID, no pricing, no context window and no benchmark card, and every number attached to it so far is a founder's claim rather than a shipped spec. The one-line answer for anyone who just wants it: Musk expects Grok 4.8's initial training run to wrap up in the week of September 14 with reinforcement learning to follow — and on x​AI's own track record, a training-completion announcement still leaves a release weeks out.

What is actually known about Grok 4.8 (it is a short section)

The confirmed record on Grok 4.8 is exactly as long as the word "confirmed" allows: the model does not exist anywhere except in Musk's posts. The model list at docs.x.ai still ends at grok-4.6 — no grok-4.7 entry, no grok-4.8 entry, no price card, no release note — and Artificial Analysis, which tracks x​AI's shipped models down to the effort level, lists nothing newer than Grok 4.6 either. What is on the record is two posts from the x​AI CEO twelve hours apart, plus a string of earlier comments about the software stack the model is said to be trained on. None of it has been followed by an x​AI announcement, and the single strongest thing actually known about Grok 4.8 is that it exists as a named training run, that it is reported to be in its final days of initial training, and that it is not released.

What Musk has said — and the honest label on each claim

Here is everything circulating about Grok 4.8, with the sourcing label on each item rather than a number masquerading as a fact.

• A roughly 2.5-trillion-parameter design — stated by Musk on September 13. That is about two-thirds larger than the 1.5-trillion-parameter Grok 4.6 and roughly 19% above the ~2.1-trillion figure long attached to Grok 4.7. It has never appeared on an x​AI specification sheet, so treat it as a widely relayed claim, not a confirmed spec.

• "Trained with our new C++ software stack" — Musk, September 13. This is the first numbered Grok that Musk has publicly tied to the C++ stack, and the stack itself is worth its own section below.

• "Will finish training this week and start RL" — Musk, September 13. Read precisely: initial training ends, then reinforcement learning begins. That matters because RL is exactly the phase Grok 4.7 has been stuck in for weeks, and it is where a launch can still slip.

• "Grok 4.8 will be a noticeable improvement" — Musk, September 14, the deliberate counterweight to his claim two sentences earlier that "Grok 4.7 should be roughly on par with Opus 5.0, not 5.1." Noticeable is not the language of a frontier jump, and the same post points further out for that: "Grok 4.9 is probably Astra/Fable class. Grok 5 maybe better than anything. We shall see." All of it is positioning — none of it has a benchmark behind it.

• Grok 4.7 ships before Grok 4.8 — not a Musk quote, but the only reading the evidence supports. Grok 4.7 finished initial training in August and is in late RL tuning; Grok 4.8 has not finished training. Whether Grok 4.7 gets a long life is the open question, and the precedent is not reassuring: Grok 4.6 replaced Grok 4.5 within about a month.

Why the C++ stack is the interesting part

The parameter count is easy to mock and impossible to verify; the software stack is the thing with real substance behind it, and it has a paper trail. In late May Musk said the C-based training stack was "almost finished." On June 29 he said the whole stack would be in C/C++ "in ~3 months." On July 8, the day Grok 4.5 was announced, he said Grok 4.5 was "not yet using our internally developed C/C++ inference software that exactly maps to the GB300 hardware" — and that "doubling or more of the current speed is probably achievable" once it did.

Put together, the claim on Grok 4.8 is that this is the first Grok trained end-to-end on that stack: training software written in C++ instead of the Python-adjacent tooling the rest of the industry uses, designed to map exactly onto NVIDIA's GB300 hardware, with a promised inference-speed payoff on top. It is a genuinely different claim from "bigger model." It is also entirely x​AI's account: no independent lab has measured the stack, no benchmark has been run against it, and "doubling or more" was Musk's estimate, not a measurement. Worth watching for exactly the reason that makes it hard to confirm.

What ships first — and the pattern you should notice

The ordering is settled by the evidence even though neither model has a date. Grok 4.7 is the nearer one: it finished initial training in August, Musk pointed at around September 12, that window expired, and on September 11 he said it needed "a few more days" of reinforcement-learning tuning — describing a model that "wraps up prematurely" on hard tasks and does not check its own answers strictly enough, with a suspected cause he hedged with "possibly." Grok 4.8 then lands in the middle of all that: named while 4.7 is still unpinned, bigger by ~400 billion parameters, and reported to be about to start the same RL phase that has already held 4.7.

The pattern worth noticing is x​AI's cadence, not its deadlines. Musk's verbal windows have a poor record — July's "about 4 weeks," August's "3 to 4 weeks," September 2's "about 10 days," and then "a few more days" on September 11. His training-completion announcements have a better one, but they still precede release by weeks: Grok 4.6's initial-training completion in late July led to an August 12 ship date, and Grok 4.5's completion led to a mid-July ship, roughly three to six weeks in each case. By that arithmetic, a training finish in the week of September 14 puts a Grok 4.8 release somewhere in October — and the honest version of that sentence is that nobody at x​AI has said so.

Where Grok 4.8 sits relative to the model you can actually call

To reason about what Grok 4.8 must beat, anchor on the model x​AI has actually shipped. Grok 4.6 launched August 12: $2 per million input tokens and $6 per million output below 200K input (doubling to $4/$12 above it), a 500,000-token context window, a February 1, 2026 knowledge cutoff, and an Artificial Analysis Intelligence Index of 44 on the current v4.3 scale, which ranks it 20th of 200 models.

• Status — in training, unreleased as of September 15 versus released August 12, 2026.

• Parameters — ~2.5T (per Musk, not on a spec sheet) versus 1.5T for Grok 4.6.

• Price per 1M tokens — none published versus $2 in / $6 out.

• Context window — none published versus 500K tokens.

• Independent score — none versus an AA Intelligence Index of 44 on the v4.3 scale (20th of 200).

• Serving stack — the new C++ stack (per Musk) versus today's production stack.

And the field Grok 4.8 is aimed at, on the same v4.3 scoreboard: GPT-6 Astra at 61, Claude Opus 5 at 51 (7th of 200), Claude Fable 5 at 50 (8th), GPT-5.6 Sol at 47 (14th), and Grok 4.5 at 39. Two notes on reading those numbers. First, they are the current scale — the older 61/62/63 values from July and August launch coverage belong to a retired index and cannot be compared with these. Second, Musk's own framing has already set expectations: 4.8 is "a noticeable improvement," while the Astra/Fable-class claim is parked one release further out at Grok 4.9. That is a CEO managing expectations toward a two-step arc, and it is worth reading that way.

What this means for teams calling Grok today

For developers, the practical ceiling on the API is unchanged by any of this: the model you call is Grok 4.6, not the one in the teasers, and a named training run changes nothing you can do today. On OrcaRouter's catalog the Grok series currently runs Grok 4.6 and Grok 4.5, billed at the provider's list price with 0% markup — so the $2/$6 you put in your cost model is the $2/$6 on your invoice. When Grok 4.8 ships and a provider onboards it, we will list it the same way at list price: one key, no new integration, no second contract.

That same routing layer is the low-risk way to move when Grok 4.8 does land, and the C++-stack story makes the caution more pointed rather than less. A brand-new model's scores are vendor claims until independent labs reproduce them — the first-week numbers are precisely the ones to distrust, and a model trained on a stack nobody else has used needs that scrutiny even more. Automatic failover lets you point a fraction of traffic at it without betting a production path on a vendor card: if a regression shows up on real requests, the next request falls back to a provider that still answers. That is the difference between trying an unproven model and betting your pipeline on it.

How to know first (better than the rumor mill)

Everything above goes stale the day x​AI flips a switch, so here is the check that beats every tracker:

• Watch the model list, not the tweets. The moment Grok 4.7 or Grok 4.8 is real, docs.x.ai/docs/models gains a model ID with pricing and a context window attached. That is how Grok 4.6 appeared — the model list first, the announcement after.

• Watch the OrcaRouter model catalog. A grok-4.7 or grok-4.8 model page on orcarouter.ai is the "you can call it today" signal — when a provider has onboarded it and we are listing it, it is routable through one API at list price.

• For consumer-side visibility, the Grok app gains the model as an entry point. API availability and app availability can diverge, so which signal you watch depends on whether you are building or just curious.

The state of play

Grok 4.8 is a named training run, not a shipped model: as of September 15 it does not appear on x​AI's model list, has no price, no context window and no independent score, and everything known about it comes from two Musk posts. What those posts claim is substantial — a 2.5-trillion-parameter model, the first numbered Grok trained on the new C++ stack, finishing initial training in the week of September 14 and entering RL — and what they promise is deliberately modest: "a noticeable improvement" over Grok 4.7, with the Astra/Fable-class leap reserved for Grok 4.9. The independently checkable facts remain negative ones: no grok-4.8 model ID in the documentation, no pricing, no benchmark entry on Artificial Analysis. Even the read that Grok 4.7 ships first is inference from the evidence, not an x​AI commitment. It is also worth noting the same week produced a wrinkle in the opposite direction: Musk publicly backed Anthropic's Dario Amodei's call for the industry to slow capability development, which makes for an odd backdrop to a roadmap of four named frontier releases. If you came here for a release date, the honest answer is still that there isn't one. The winning move for anyone building today is unchanged: build on what is live at $2/$6, and switch the moment the model list changes.