Skip to content
FrontierDot

Grok 4.8: What xAI Has Actually Said, and the Grok 4.5 to 4.7 Release Timeline

Live timeline · last updated · 1 entry in the update log

As of today Grok 4.8 is something Elon Musk has talked about on X, once in September and again on Oct 7, and xAI’s own docs list no such model. The sequence below runs from Grok 4.5 in July to Grok 4.7 in September, and every entry points to a primary page or post.

As of Oct 18, 2026, Grok 4.8 is not released and has no xAI announcement, docs page or API model name; the only primary evidence is Elon Musk’s X posts of Sep 14 and Oct 7, while the latest shipped model is Grok 4.7 (Sep 21).

Read on 2026-10-18: xAI’s news pages for Grok 4.5, 4.6 and 4.7, the xAI developer release notes, models page and Grok 4.7 model page, plus official posts on X by @SpaceXSI, @grok and @elonmusk. Times in the timeline are UTC as X reports them. We cite no third-party articles, did not run any model, and add nothing that is not on those pages.

Timeline

  1. xAI, which now publishes under the SpaceXSI name, announces Grok 4.5 and calls it its first model trained specifically for coding and agents, trained with Cursor. The news page says it was trained across tens of thousands of NVIDIA GB300 GPUs, and prices it at $2 per million input tokens and $6 per million output tokens. It is available the same day in Grok Build, Cursor and the xAI console. The company’s post on X, at 17:57 UTC, says EU availability is expected later in the month.

    xAI: Introducing Grok 4.5

  2. Grok 4.6 is released, described as building on Grok 4.5 with a focus on long-running agents and more ambitious interactive and visual work. xAI’s post at 15:32 UTC says it is available in Grok Build, Cursor, Grok Bot and the API at the same $2 and $6 prices, with doubled included usage in Cursor and Grok Build for the first week.

    xAI: Introducing Grok 4.6

  3. The official @grok account says Grok 4.6 is now available across grok.com, iOS and Android, extending the model from developer tools to the consumer apps about sixteen days after the first release.

    @grok on X: Grok 4.6 across web, iOS and Android

  4. Replying to a user at 17:22 UTC, Elon Musk writes that Grok 4.7 “needs a few more days to cook,” adding that xAI might have penalized response length too much in reinforcement learning, so the model still gives up early on hard tasks and is not yet rigorous enough in checking its work. This is the first dated statement in our sources about Grok 4.7 being unfinished.

    Elon Musk on X: Grok 4.7 needs a few more days

  5. Grok 4.8 is named in public. At 01:24 UTC, answering a question about what to expect from xAI in September besides Grok 4.7, Musk writes: “Grok 4.8, which is a 2.5T model trained with our new C++ software stack, will finish training this week and start RL.” Later that day, at 11:19 UTC, he says Grok 4.7 should be roughly on par with Opus 5.0, that Grok 4.8 “will be a noticeable improvement,” that Grok 4.9 is “probably Astra/Fable class,” and that Grok 5 may be “better than anything. We shall see.” At 14:07 UTC he adds that xAI’s pre-training software is now an internally developed C/C++ stack written primarily by the Starlink software team, while the reinforcement-learning stack is still somewhat of a legacy pile.

    Elon Musk on X: Grok 4.8 is a 2.5T model

  6. Grok 4.7 is released. xAI’s post at 16:17 UTC says it is a notable improvement over Grok 4.6 at the same price and speed, available in Cursor, Grok Build and the Grok API. The announcement says it uses a new, larger base model than Grok 4.6, was trained with a longer reinforcement-learning run weighted toward problems that take many hours, and was trained to understand the Grok Bot harness. Pricing starts at $2 and $6 per million input and output tokens, with a fast variant at twice the price.

    xAI: Introducing Grok 4.7

  7. xAI’s official account says Grok 4.7 is now available on the Gemini Enterprise Agent Platform, the first of two cloud-platform availability posts we found for the model in early October.

    @SpaceXSI on X: Grok 4.7 on Gemini Enterprise Agent Platform

  8. Two things happen on one day. At 00:04 UTC xAI announces that Grok 4.7 is live on Microsoft Foundry. At 18:02 UTC Musk, discussing how Grok Bot picks models, writes that most Bot requests are simple and “will be handled by a lightning-fast version of Grok 4.8 when that comes out.” This is the latest primary mention of Grok 4.8 we could find, and it is phrased as a future event.

    Elon Musk on X: Grok 4.8 and Grok Bot

  9. Status check. xAI’s developer release notes list Grok 4.7 under September as the newest text model, and the models page, which says to use Grok 4.7 for code and chat, lists grok-4.7, grok-4.6, grok-4.5 and older models with no grok-4.8 entry. We found no xAI news page for Grok 4.8.

    xAI docs: Models

Short answer: Grok 4.8 has been described but not released

The question this page answers is simple: is there a verifiable Grok 4.8 timeline? There is a short one, and it is built entirely from statements by Elon Musk on X. Between September 14 and October 7 he described Grok 4.8 as a 2.5-trillion-parameter model trained with a new C++ software stack, said its pre-training would finish that week, said reinforcement learning would follow, and later referred to a fast version of it as something that will exist “when that comes out.” There is no xAI news post, no model card, no entry in the developer release notes and no model name in the API documentation.

A lot of coverage blurs the two. A founder’s post is a primary source for what he said; on its own it says nothing about what has shipped, and it is not a date. The September post says “this week” for the end of one training phase; it does not say when a model will be usable. The October post says “when that comes out,” which is the language of something that has not come out. On this page the Grok 4.8 entries are therefore labelled as statements, and the release entries are only the models xAI itself documents: Grok 4.5, 4.6 and 4.7.

Background: three releases in ten weeks

To read the Grok 4.8 statements you need the cadence they sit in. xAI, which now publishes under the SpaceXSI name on its news and docs pages, launched Grok 4.5 on July 8, 2026. The announcement framed it as the company’s first model trained specifically for coding and agents, trained alongside Cursor, and it was available on day one in Grok Build, in Cursor and in the console. The price, $2 per million input tokens and $6 per million output tokens, would not change for the next two releases.

Grok 4.6 followed on August 12, thirty-five days later. Its page describes a longer supplemental training run than Grok 4.5, supervised fine-tuning data regenerated with Grok 4.5, and agentic reinforcement learning across tasks such as kernel optimization, web development and computer-aided design. On August 28 the official Grok account put it into the grok.com, iOS and Android apps. The order was developer tools and the API first, then the consumer apps about two weeks later.

Grok 4.7 arrived on September 21, forty days after the Grok 4.6 announcement. Its announcement says it uses a new, larger base model, and it ran a longer reinforcement-learning phase on harder, longer tasks. The price and the list of reasoning effort levels in the docs stayed the same as for 4.6. The docs add that the model has a 500k-token context window, text and image inputs with text-only output, and a stated knowledge cutoff of May 2026.

How Grok 4.7 slipped, in Musk’s own words

The most useful context for the Grok 4.8 statements is the delay that preceded the 4.7 launch. On September 11 Musk replied that Grok 4.7 “needs a few more days to cook.” He gave a specific reason: the reinforcement-learning stage might have penalized response length too much, so the model still gave up on hard tasks it could do, and was not yet sufficiently rigorous in checking its work. Ten days later xAI’s launch post said the opposite had been achieved in the shipped model, that it works longer on difficult tasks and checks its own work more carefully. The launch page does not mention the delay or link the two, so we draw no conclusion beyond placing them in order.

We found no primary statement from xAI of a firm release date for Grok 4.7 before it shipped, so we do not describe the launch as late relative to a commitment. What the record supports is narrower: on September 11 the founder said more work was needed, and on September 21 the company released the model.

What the September 14 posts actually claim

Four posts on September 14, all by Musk, carry nearly everything verifiable about Grok 4.8. The first, at 01:24 UTC, is the headline: a 2.5T model, a new C++ software stack, training finishing that week, then RL. It is a reply to an account asking what xAI would deliver in September beyond Grok 4.7. It gives no benchmark, price, product or date.

The second, at 11:19 UTC, is a ladder. Grok 4.7 should be roughly on par with Opus 5.0 and not 5.1, “better in some ways, worse in others,” with multimodal performance needing a fix. Grok 4.8 “will be a noticeable improvement.” Grok 4.9 is “probably Astra/Fable class,” and Grok 5 “maybe better than anything. We shall see.” These are self-assessments and forecasts from the head of the company, hedged in his own wording, and nothing on xAI’s pages repeats them.

The third and fourth posts concern the software. At 14:07 UTC Musk says the pre-training software is now the internally developed C/C++ stack, written primarily by the Starlink software team, while the reinforcement-learning stack is still “somewhat of a legacy pile.” At 16:56 UTC he says the C/C++ training stack was written by humans and that AI is not yet good enough for extremely high-performance software, though it will be. Together they explain the claim in the first post: the “new C++ software stack” refers to pre-training, not to the whole pipeline. They also hint at why RL, not pre-training, may be the slower leg for 4.8, but that is our reading, not a statement from xAI.

The Grok Bot mention on October 7

The latest primary mention of Grok 4.8 is a side remark. Responding to a post about Grok Bot choosing among models for different tasks, Musk wrote that most Bot requests are simple and will be handled by a lightning-fast version of Grok 4.8 when that comes out, adding that the operating principle is to give Bot users the best combination of speed and intelligence. It tells us two things and no more: that a fast Grok 4.8 variant is planned inside xAI’s product thinking, and that as of October 7 the model had not shipped.

It also fits a pattern in the earlier launches. Grok 4.6’s post names Grok Bot among the places it is available at launch, and the Grok 4.7 announcement says the model was trained to understand the Grok Bot harness. The October 7 post is consistent with Bot being a primary target for a future model, though it gives no timeline.

What a release would look like, and what we are watching

Each of the three shipped models left the same trail: a news page on x.ai, an announcement thread from the official SpaceXSI account, a dated entry in the developer release notes and a model name on the docs models page. Grok 4.5, 4.6 and 4.7 each launched in Cursor and Grok Build first, with the API in the same announcement or shortly after. If Grok 4.8 ships, we will add it to the timeline only when at least the news page or the release notes entry exists, and we will say plainly which of those we saw.

Until then, the thing to watch is how long the gap runs between a statement and a launch. Grok 4.7 is the one example: Musk said on September 11 that it needed days, and the release came ten days later. For Grok 4.8 the only dated statement about training is from September 14, and we have no later primary update on whether pre-training finished or RL started.

What we could not verify

  • That Grok 4.8 will be released, when, or under what name. No xAI page announces it.
  • The 2.5-trillion-parameter figure, the new-stack claim and the training schedule, beyond the fact that Musk posted them. xAI has published no model card or technical detail for Grok 4.8.
  • Whether Grok 4.8 pre-training finished in the week of September 14 or whether RL has begun. We found no later primary post on either.
  • Any benchmark, price, context length or API name for Grok 4.8. None appears in xAI’s docs as of October 18.
  • Comparisons between Grok models and other labs’ models. The benchmark tables on xAI’s pages are xAI-reported, and we did not reproduce them, so we cite none as fact.
  • Whether Musk’s forecasts about Grok 4.9 and Grok 5 are plans or speculation. They are hedged in his own words, and nothing on xAI’s pages repeats them.

Where this stands on Oct 18

xAI has documented three models, Grok 4.5, 4.6 and 4.7, each with a date and a price. Grok 4.8 is, according to Musk, a model in training; he gave a timeframe for one training phase and none for a release. If you are deciding what to build on, the supported choice per xAI’s own models page is Grok 4.7. We will update this page, with a dated entry in the log above, when xAI publishes anything about Grok 4.8 on its own pages.

Related reading on this site: the Reflection Beam timeline and the Muse Spark Open timeline use the same format for two open-weight releases from the same week.

Keep reading on Front

Every post, newest first, is on the Front blog index.

FAQ

Is Grok 4.8 real?

It is real as a model Elon Musk says is in training. It has not been released, and xAI’s news pages, release notes and models page do not list it as of October 18, 2026.

When will Grok 4.8 be released?

No release date has been stated. Musk said on September 14 that training would finish that week and RL would start, and on October 7 referred to a fast version “when that comes out.”

What is the latest Grok model?

Grok 4.7, released September 21, 2026, available as grok-4.7 in the xAI API with a 500k-token context window.

Update log

  • — First published. Checked xAI docs and posts through Oct 18, 2026 (UTC+8).

Sources