Grok 4.8 is a 2.5-trillion-parameter model trained with a new C++ software stack, according to Elon Musk, who wrote on X that it “will finish training this week and start RL”. The reply went out at 01:24 UTC on 14 September 2026, still the evening of 13 September in the US, and it is the first time the model has been named in public. xAI, which now publishes under the SpaceXAI name, has released no announcement, model card, API identifier or price for Grok 4.8.
Musk added more detail about ten hours later. He said Grok 4.7, which is still unreleased, “should be roughly on par with Opus 5.0, not 5.1”, that Grok 4.8 “will be a noticeable improvement”, that Grok 4.9 “is probably Astra/Fable class”, and that Grok 5 “maybe better than anything”. When we read it on 15 September, the original Grok 4.8 post had passed 3 million views. For the wider model landscape these claims sit in, see our AI models and tools hub.
This article works through both posts line by line. It sets the 2.5T figure against every parameter count xAI and Musk have published, traces the C and C++ stack back to its first mention in May, explains what moving into RL involves, and checks the claims against the release record of Grok 4.5, 4.6 and 4.7. The same weekend, Musk also backed Dario Amodei’s call to pace frontier AI, which makes the speed of this roadmap worth reading closely.
Table of contents
- What Musk Actually Said About Grok 4.8
- What 2.5 Trillion Parameters Does and Does Not Tell You
- The New C++ Software Stack Behind Grok 4.8
- What “Finish Training and Start RL” Means for Grok 4.8
- Grok 4.7 Is Still Unreleased, and That Shapes Grok 4.8
- Where Grok 4.8 Would Land Against Opus 5, Fable and Astra
- Serving a 2.5T Model: Memory, Speed and Price
- What Is Still Unknown About Grok 4.8
- What Grok 4.8 Means for Teams Building on Grok
- Frequently Asked Questions
- References
What Musk Actually Said About Grok 4.8
Everything public about Grok 4.8 comes from three short replies on X. None came from xAI’s corporate account and none links to a technical document, so the exact wording carries all of the weight.
The 22-word reply that named Grok 4.8
At 13:14 UTC on 13 September, the account @techdevnotes asked: “Apart from Grok 4.7, what are we even expecting from SpaceXAI in september”. Musk answered about twelve hours later: “Grok 4.8, which is a 2.5T model trained with our new C++ software stack, will finish training this week and start RL”.
That one sentence makes four separate claims. Grok 4.8 exists as a training run. It has 2.5 trillion parameters. It was trained on a new software stack written in C++. And its current training phase ends this week, after which the RL stage begins. When we read the post through the fxtwitter API on 15 September it showed 3,037,955 views, 17,800 likes, 1,186 replies and 1,148 bookmarks.
The follow-up that ranked the whole ladder
At 11:19 UTC on 14 September, replying to another user, Musk placed Grok 4.8 on a ladder of versions: “Grok 4.7 should be roughly on par with Opus 5.0, not 5.1. Better in some ways, worse in others. We need to fix multimodal performance. Grok 4.8 will be a noticeable improvement. Grok 4.9 is probably Astra/Fable class. Grok 5 maybe better than anything. We shall see.”
That is a lower bar than the one Musk set a month earlier. On 12 August he wrote that “Grok 4.7 will exceed all current models”. Grok 4.7 is now pitched at parity with a single Anthropic model, Grok 4.8 is described only as an improvement on it, and the claim of beating the field has moved two versions further out.
“That will be Grok 5”
Ten minutes before that post, replying to a user who had asked how close Grok 4.8 would come to artificial general intelligence, Musk wrote: “That will be Grok 5”. The reply had 356,723 views when we read it. It is the only statement Musk has made about what Grok 4.8 will not be, and it offers no definition of the term.
| Claim | Posted (UTC) | Can it be checked today? |
|---|---|---|
| Grok 4.8 is a 2.5T model | 14 Sep, 01:24 | No. xAI has published no parameter count for any Grok 4 model |
| Trained with a new C++ software stack | 14 Sep, 01:24 | No. There is no technical report, and earlier posts describe the stack as C and C/C++ |
| Training finishes this week, then RL starts | 14 Sep, 01:24 | Only by waiting for a later update |
| A noticeable improvement over Grok 4.7 | 14 Sep, 11:19 | No. Grok 4.7 itself has not been released |
| Grok 4.7 roughly on par with Opus 5.0 | 14 Sep, 11:19 | No. Anthropic’s model list has no Opus 5.1 to compare against |
| Grok 4.9 probably Astra/Fable class | 14 Sep, 11:19 | No. It is a forecast about a model with no announced training run |
| General intelligence arrives with Grok 5 | 14 Sep, 11:09 | No. No definition or test was given |
What 2.5 Trillion Parameters Does and Does Not Tell You
A parameter is one of the learned numbers inside a large language model, adjusted during training until the model’s predictions improve. The headline figure is the easiest part of the post to repeat and the hardest to interpret, because Musk did not say what kind of parameter count 2.5T is.
Total parameters versus active parameters
xAI’s only published architecture is Grok-1. When the company released those weights in March 2024, it described a “314 billion parameter Mixture-of-Experts model with 25% of the weights active on a given token”. In a mixture-of-experts design, a router sends each token to a small set of expert sub-networks, so the model stores far more parameters than it uses for any single token.
That distinction changes what 2.5T means for Grok 4.8. DeepSeek-V3, for example, has 671 billion total parameters but activates 37 billion for each token. If Grok 4.8 follows the Grok-1 pattern, 2.5T is most likely a total count, and the compute spent per token depends on an active count Musk has not given. A dense 2.5T model, where every parameter works on every token, would be a very different and far costlier system to serve.
From Grok-1’s 314 billion to Grok 4.8’s 2.5 trillion
Musk has put numbers on most recent Grok runs, always in posts rather than documentation. On 25 May he wrote that the “V9-Medium (1.5T)” foundation model had finished training and would be “a major improvement over the 0.5T v8-small that currently serves all Grok production traffic”. On 28 July he called Grok 4.6 “the 1.5T model” and Grok 4.7 “the 2.1T model”. Grok 4.8 at 2.5T is the next step on that line.
| Model | Parameters | Source | Status on 15 Sep 2026 |
|---|---|---|---|
| Grok-0 | 33 billion | xAI, November 2023 | Prototype that preceded Grok-1 |
| Grok-1 | 314 billion, 25% active | xAI, March 2024 | Open weights, Apache 2.0 |
| v8-small | 0.5 trillion | Musk on X, 25 May 2026 | Served production traffic in May |
| V9-Medium | 1.5 trillion | Musk on X, 25 May 2026 | Foundation model, training finished in May |
| Grok 4.6 | 1.5 trillion | Musk on X, 28 Jul 2026 | Released 12 August, current flagship |
| Grok 4.7 | 2.1 trillion | Musk on X, 28 Jul 2026 | Unreleased, in RL rework |
| Grok 4.8 | 2.5 trillion | Musk on X, 14 Sep 2026 | Finishing its main training run |
The arithmetic from those posts is simple. At 2.5T, Grok 4.8 is about 19% larger than Grok 4.7’s 2.1T, 67% larger than the 1.5T behind Grok 4.6, five times the 0.5T model that served production traffic in May, and almost eight times Grok-1.
Why size alone will not settle it
Scaling research has shown for years that parameter count is only one input. DeepMind’s 2022 Chinchilla study found that many large models were undertrained, and that model size and training tokens should be scaled in equal proportion for a fixed compute budget. A bigger model fed too little data can lose to a smaller one trained properly.
xAI’s own launch notes point the same way. The Grok 4.6 page credits a longer supplemental training run, curated model-generated data, high-quality engineering data and “an improved optimizer and training recipe” for its gains, and it does not mention parameters at all. The 2.5T figure tells you Grok 4.8 is large. Whether it is better will depend on its data, its recipe and the RL stage that starts after this week.
The New C++ Software Stack Behind Grok 4.8
The second half of Musk’s sentence is the more unusual claim. Most frontier labs train on shared frameworks with a Python front end and compiled kernels underneath. Musk has spent almost four months saying xAI is replacing that arrangement with code written for its own hardware.
From JAX and Rust to an in-house stack
xAI has never hidden its original tooling. The Grok-1 release notes say the model was “trained from scratch by xAI using a custom training stack on top of JAX and Rust”. JAX is an open-source Python library, developed at Google, that compiles array programs for accelerators through the XLA compiler. It is widely used for very large training runs, which makes it the natural baseline for Musk’s comparison.
On 28 May, Musk wrote that “SpaceX has almost finished writing V1.0 of an in-house AI training stack in C” and that “the potential speed improvement vs JAX for large training runs is over an order of magnitude”. That post has been viewed more than 30.7 million times, about ten times as often as the Grok 4.8 reply.
Exact-mapping software to 220,000 GB300s
The same post said the stack “exact-maps to 220k GB300s with 800G NICs, making heavy use of pipeline parallelism and getting as close to bare metal as possible”. Every phrase in that sentence is a design decision rather than a slogan, and each one explains something about how Grok 4.8 was built.
GB300 is Nvidia’s Grace Blackwell Ultra platform. Nvidia’s GB300 NVL72 rack holds 72 Blackwell Ultra GPUs and 36 Grace CPUs with 20 TB of GPU memory, so 220,000 GPUs is the equivalent of about 3,056 of those racks. The 800G NICs are 800-gigabit network cards, the links that carry gradients and activations between racks. For context on what running hardware at that scale involves, see our page on data center operations.
Pipeline parallelism splits a model’s layers into stages that sit on different GPUs, so a model too large for any single device can train like an assembly line. Nvidia’s Megatron-LM research showed how pipeline, tensor and data parallelism can be combined to train models with a trillion parameters across thousands of GPUs. “Exact-map” and “bare metal” mean writing that scheduling for one known hardware layout instead of a portable framework that has to run anywhere.
C, C++ or both? The wording has drifted
Musk has described the stack in several different ways over four months. The table lists every post we could verify through the fxtwitter API, with times in UTC.
| Date | What Musk wrote | Language named |
|---|---|---|
| 28 May, 06:27 | “almost finished writing V1.0 of an in-house AI training stack”, over an order of magnitude faster than JAX | C |
| 28 May, 16:40 | “Next will be writing the inference stack in C for simultaneous high-speed RL across a large block of GB300s. (We do use a little C++ tbh, but not much)” | C, a little C++ |
| 29 Jun, 00:10 | Gains in “~3 months when the entire training and inference stack is written in C/C++ and massively simplified (most software layers will be deleted completely)” | C/C++ |
| 8 Jul, 21:30 | “Grok 4.5 is not yet using our internally developed C/C++ inference software”. “Doubling or more of the current speed is probably achievable.” | C/C++ |
| 14 Sep, 01:24 | Grok 4.8 is “trained with our new C++ software stack” | C++ |
The move from “C” to “C/C++” to “C++” may simply be shorthand, and nothing in the posts says the design changed. What the timeline does show is a schedule kept. The 29 June post promised the full stack in about three months, and the Grok 4.8 reply arrived 77 days later describing a model trained on it.
What remains unclear is the inference side. Musk said in July that Grok 4.5 was not yet using the C/C++ inference software. He has not said whether Grok 4.6 runs on it today, or whether Grok 4.8 will be served with it when it launches.
Deleting software layers is the real bet
The most important phrase in the June post is “most software layers will be deleted completely”. Every layer between a training script and a GPU exists for a reason: portability, debugging tools, safety checks, or letting researchers try new ideas without rewriting kernels. Removing layers trades that flexibility for speed on one fixed cluster.
That trade can pay off for a company that owns its hardware and trains one family of models on it. It also concentrates risk. A bug in a custom communication path can be harder to find than one in a framework used by thousands of teams, and the order-of-magnitude speed claim has no public benchmark behind it. If Grok 4.8 ships on time and performs as described, that is the strongest evidence the stack works.
What "Finish Training and Start RL" Means for Grok 4.8
Musk’s timing claim is specific: the current training phase ends this week and reinforcement learning begins. Understanding those stages explains why a model that is “finishing training” can still be weeks or months from release.
Pre-training, supplemental training, SFT and RL
Large models are built in stages. Pre-training teaches the base model to predict text from a very large corpus. Supplemental training adds targeted data, such as the SpaceX engineering corpus Musk said on 21 July would go into “the 2T run”. Supervised fine-tuning, or SFT, then trains the model on worked examples of the behaviour wanted.
The last stage is reinforcement learning, where the model attempts tasks, receives a score from automated or model-based graders, and is updated to make high-scoring behaviour more likely. For an agentic model this is where it learns to persist through long tasks, use tools and check its work. It is also where the most tuning happens, because every choice about the reward shapes what the model learns to do.
How xAI described the pipeline for Grok 4.5 and 4.6
xAI’s release notes give the clearest public picture of what Grok 4.8 will go through next. The Grok 4.5 page says the model “was trained across tens of thousands of NVIDIA GB300 GPUs” and that its RL training “covers hundreds of thousands of tasks, centered on multi-step software engineering and other technical work”. It adds that the stack is “built for highly asynchronous training, so agentic rollouts can run for many hours while learning continues”.
The Grok 4.6 page describes the SFT data as trajectories regenerated with Grok 4.5 and filtered “with model-based checks”, followed by RL on knowledge work, general coding and specialist environments for kernel optimisation, web development and computer-aided design. If Grok 4.8 follows that recipe, the RL stage is not a short finishing step. It is a full training programme of its own.
Why the RL stage is where Grok 4.7 stumbled
Grok 4.7 shows what can go wrong. On 11 September Musk wrote that it “needs a few more days to cook”, adding: “We might have penalized response length too much (or something) in RL, as it still gives up on hard tasks (that it can do!) too early and isn’t yet sufficiently rigorous in checking its work.”
That is a classic reward-design problem. Penalising long answers keeps serving costs down and makes a model feel quick, but push the penalty too far and the model learns that stopping early is safer than working a hard problem through. Grok 4.8 enters the same stage with the same risk, and with Grok 4.7’s fix still unfinished.
Grok 4.7 Is Still Unreleased, and That Shapes Grok 4.8
Grok 4.8’s announcement cannot be separated from the model it follows. Grok 4.7 has been promised on several dates, and none of them has held.
Seven weeks of moving dates
On 24 July Musk wrote “Grok 4.6 in 2 weeks and Grok 4.7 in 4 weeks”. Grok 4.6 arrived close to plan. Grok 4.7 did not, and the promised date has moved several times since. Here is the full record, with dates implied by each post.
| Date | What Musk said about Grok 4.7 | Implied release |
|---|---|---|
| 24 Jul | “Grok 4.7 in 4 weeks” | About 21 Aug |
| 28 Jul | “the 2.1T model released a few weeks later”, after Grok 4.6 around 7 August | Late August |
| 12 Aug | “should be ready in 3 to 4 weeks”, initial training complete, SpaceX data in supplemental training | 2 to 9 Sep |
| 15 Aug | “Fable 5 is smarter overall”, but Grok 4.7 “has a good chance of exceeding all current models” | None given |
| 2 Sep | “Grok 4.7 comes out in 10 days” | 12 Sep |
| 11 Sep | “needs a few more days to cook”, RL penalised response length too much | Mid-September |
| 14 Sep | “roughly on par with Opus 5.0, not 5.1”, multimodal needs fixing | None given |
Measured against each date Musk gave, the slippage is easy to count. Grok 4.6 was promised “around August 7” and shipped on 12 August, five days late. The 1.5T V9-Medium base was said on 25 May to be “2 to 3 weeks to public release”, and the next Grok flagship, Grok 4.5, arrived on 16 July, 31 days after the later end of that window. Grok 4.7 is 25 days past the 24 July estimate and still counting.
Correction: Grok 4.7 has not been shelved
Startup Fortune ran the headline “Elon Musk Shelves Grok 4.7 and Unveils Grok 4.8 Instead”. Nothing Musk wrote says that. His 11:19 UTC post on 14 September still describes Grok 4.7 as a model that is coming, compares it with Opus 5.0, and presents Grok 4.8 as its successor rather than its replacement. Delayed is accurate. Shelved is not supported by any post we found.
Three numbers in the coverage that do not match the posts
Crypto Briefing’s headline said Musk “hints at 3 trillion-parameter Grok model”, and its text said Grok 3 and Grok 4 “were built on a 3 trillion-parameter architecture”. We found no Musk post that says this, and his own figures for the Grok 4 series top out at 2.5T for Grok 4.8. A July explainer on wan27.org put Grok 4.6 at “approximately 2 trillion” parameters, while Musk’s 28 July post calls it “the 1.5T model”.
The third is about timing. Several reports dated Musk’s post 13 September, which is correct in US time zones, while the post itself is stamped 01:24 UTC on 14 September. Neither date is wrong, but it explains why reports of the same post disagree by a day.
What the delay pattern means for Grok 4.8
Musk said on 12 August that Grok 4.7’s initial training was complete. It was still unreleased 34 days later. If Grok 4.8 finishes its main run around 20 September and needs the same 34 days, it would not appear before about 24 October. That is arithmetic on Musk’s own posts, not a forecast, and a faster RL stage on the new stack could shorten it.
Where Grok 4.8 Would Land Against Opus 5, Fable and Astra
Musk’s ladder names three reference points from rival labs: Anthropic’s Opus 5.0, Anthropic’s Fable line and OpenAI’s Astra. Checking them against published results shows how much room those words leave.
Grok 4.6 is the only measured baseline
No benchmark exists for Grok 4.7 or Grok 4.8, so the latest numbers xAI has published are for Grok 4.6. The table uses xAI’s own evals from its 12 August launch page, including its third-party comparisons.
| Evaluation | Grok 4.6 High | Grok 4.5 High | GPT-5.6 Sol Max | Fable 5 Max |
|---|---|---|---|---|
| AA Intelligence Index | 61 | 56 | 61 | 62 |
| CursorBench v3.2 | 69.9% | 66.7% | 67.2% | 70.5% |
| DeepSWE v1.1 | 65.9% | 54% | 73% | 70% |
| FrontierCode v1.1 (Extended) | 61.3% | 56.6% | 60.6% | 63.6% |
| Terminal-Bench v3.0 | 26% | 15.7% | 34.6% | 34.1% |
| APEX-Agents | 57.5% | 47.1% | 56.7% | 59.2% |
Grok 4.6 matches OpenAI’s GPT-5.6 Sol on the composite index and leads it on three of the five other rows, but trails Anthropic’s Fable 5 on all six and trails both rivals clearly on DeepSWE and Terminal-Bench. That is the gap Grok 4.7 and Grok 4.8 are meant to close.
There is no Opus 5.1 to compare against
Musk said Grok 4.7 should be on par with “Opus 5.0, not 5.1”. Anthropic’s model overview lists four current models, Fable 5.1, Opus 5, Sonnet 5 and Haiku 4.5, and none of the model identifiers on that page is an Opus 5.1. Anthropic launched Opus 5 on 24 July, describing it as coming “close to the frontier intelligence of Claude Fable 5 at half the price”.
So the “not 5.1” half of the comparison refers to a model that has not been released, or perhaps to Fable 5.1, which is Anthropic’s newest model. The posts do not say which. Either way, the one checkable reading is that Grok 4.7 is aimed at Opus 5, Anthropic’s mid-priced model, rather than its top one.
| Model | Released | API ID | Input / output per million tokens | Context |
|---|---|---|---|---|
| Grok 4.6 | 12 Aug 2026 | grok-4.6 | $2 / $6 (under 200k tokens) | 500k tokens |
| Claude Opus 5 | 24 Jul 2026 | claude-opus-5 | $5 / $25 | 1M tokens |
| Fable 5.1 | 1 Sep 2026 | claude-fable-5-1 | $10 / $50 | 1M tokens |
| GPT-6 Astra | 3 Sep 2026 | gpt-6-astra | $10 / $50 (short context) | 1,050,000 tokens |
| Grok 4.7 | Not released | None | Not published | Not published |
| Grok 4.8 | Not released | None | Not published | Not published |
“Astra/Fable class” covers a wide band
Musk said Grok 4.9, not Grok 4.8, is “probably Astra/Fable class”. The best public yardstick for that phrase is the Artificial Analysis Intelligence Index table OpenAI published with GPT-6 Astra. On it, GPT-6 Astra scores 61.2, below Opus 5 at 63.1, while Fable 5.1 leads at 65.7. Our look at Claude Fable 5.1 covers that model in more detail.
That creates an odd ladder. On this one composite index, the Opus 5 level Musk set for Grok 4.7 sits above the Astra level he set for Grok 4.9. The two vendor tables also use different index versions and settings, so xAI’s figure of 61 for Grok 4.6 is shown below for context only, not as a like-for-like result.
The whole field sits within seven points. “Astra/Fable class” could mean anything from 61 to 66 on this measure, so Grok 4.8 does not need a dramatic jump to be close to the rival frontier on paper, but it needs published results before any placement means much.
Serving a 2.5T Model: Memory, Speed and Price
A model has to be served as well as trained, and size shows up directly in the hardware bill. Musk’s own posts already hint at the trade-off: on 28 July he called Grok 4.7 “slightly slower to serve, albeit with even better token efficiency”.
The weight-memory arithmetic
Each parameter stored at 16-bit precision takes two bytes. The weights of a 2.5T model therefore need about 5 TB at 16-bit, 2.5 TB at 8-bit and 1.25 TB at 4-bit, before any memory for the conversation context. Training needs far more. The ZeRO paper from Microsoft Research puts mixed-precision training with the Adam optimizer at 16 bytes per parameter for weights, gradients and optimizer states, which for 2.5T parameters is 40 TB.
The weights alone fit comfortably inside a single rack’s GPU memory, even at 16-bit. The training states do not: they need at least two racks’ worth before any activations, which is why pipeline parallelism across many racks is central to the stack Musk described.
Musk’s own speed promise
The strongest concrete promise about serving is the 8 July post: the C/C++ inference software is not yet in use, and “doubling or more of the current speed is probably achievable”. If that software is live by the time Grok 4.8 launches, a larger model could be served at similar speed to today’s smaller one. If it is not, the extra size is likely to show up as slower responses or a higher price.
What Grok 4.8 might cost is not known
xAI has published no price for Grok 4.8. For reference, Grok 4.6 costs $2 per million input tokens and $6 per million output tokens under 200,000 tokens of context, rising to $4 and $12 above it, and xAI says a fast variant costs twice as much. That is well below Opus 5 at $5 and $25, and far below Fable 5.1 and GPT-6 Astra at $10 and $50. Our LLM API pricing guide compares the wider market.
Price has been xAI’s clearest advantage. On 15 August Musk himself wrote that “Fable 5 is smarter overall, but Grok is much faster and lower cost”. Whether Grok 4.8 keeps that position at 2.5T depends on the inference work the stack was built for.
What Is Still Unknown About Grok 4.8
Strip away the replies and very little about Grok 4.8 is on the record. The table separates the claims from the facts that would let anyone test them.
| Question | What exists now | What would settle it |
|---|---|---|
| Is 2.5T total or active? | One word, “2.5T”, in a reply | A model card listing total and active parameters |
| Is it better than Grok 4.6? | “A noticeable improvement” over an unreleased model | Published evals with harness details |
| When will it ship? | Training ends “this week” | An xAI release note and a docs.x.ai model entry |
| What will it cost? | Nothing | A pricing table, including the long-context tier |
| What does the C++ stack change? | Speed claims in posts from May to July | Throughput or latency figures against Grok 4.6 |
| How was it tested for safety? | Nothing specific to this model | A system card or safety section, as Grok 4.6 had |
No model card, no ID, no price
As of 15 September, xAI’s news page lists no Grok 4.7 or Grok 4.8 post, and its model documentation still names grok-4.6 as the newest model. The API pages list grok-4.6, grok-4.5, grok-4.3 and older models. Anyone planning around Grok 4.8 is planning around a statement, not a product.
How to read the next announcement
When xAI does publish Grok 4.8, a few details will tell you more than the headline. Check whether the parameter figure is total or active. Look for the context window and the long-context price, since Grok 4.6 doubles its price above 200,000 tokens. Read which harness each benchmark used. And look for any statement on whether the C/C++ inference software serves the model, because that is what the speed claims depend on.
What Grok 4.8 Means for Teams Building on Grok
For developers and businesses, the practical question is what to do before Grok 4.8 exists as a product. The answer is mostly to build on what is released and keep switching costs low.
Build on Grok 4.6 today, not on a post
Grok 4.6 is the model you can actually use. It is available through the xAI API, in Cursor and Grok Build, through OpenRouter, Vercel and Cloudflare, and, according to xAI’s news page, in GitHub Copilot, Amazon Bedrock, Microsoft Foundry and Google’s Gemini Enterprise Agent Platform. Our guide to using Grok Build covers setup, limits and costs.
The Cursor connection matters for Grok 4.8 too. Musk said in May that “a lot of Cursor data was added” in supplementary training, and SpaceX has since completed its acquisition of Cursor. Coding tasks are where xAI has put most of its evaluation effort, so they are the likeliest place for Grok 4.8 to show a difference first.
Keep model choice swappable
Grok 4.7 has shown how far a release can move. Teams should avoid hard-coding a model name deep in their applications. Route requests through a thin internal layer, keep an evaluation set of your own real tasks, and run it against Grok 4.6, Opus 5 and any new release on the day it arrives. Then Grok 4.8 becomes a configuration change you can test, not a rebuild.
Watch the release notes, not the replies
The reliable signal will be a docs.x.ai release note and a new model entry with an ID and a price. Replies on X have given useful direction for a year, but they have also given at least four dates for Grok 4.7 that did not hold. Treat the Grok 4.8 reply as a statement of intent, and wait for the documentation before you commit budgets or roadmaps to it.
Frequently Asked Questions
These are the questions readers are most likely to have about the announcement, answered from the posts and documentation above.
When will Grok 4.8 be released?
No date has been given. Musk said its main training run finishes this week and RL starts next. Grok 4.7 was still unreleased 34 days after Musk said its initial training was complete, so a release before late October would be faster than the recent pattern.
Is Grok 4.8 bigger than GPT-6 Astra or Fable 5.1?
Nobody outside those companies can say. OpenAI and Anthropic do not publish parameter counts for their current models, and xAI has not published one for Grok 4.8 either. The 2.5T figure comes only from Musk’s post.
What does the new C++ stack change for users?
Directly, nothing yet. It is training and serving infrastructure. If it delivers the speed Musk has claimed, users should see faster responses or lower prices for a model of this size. Neither has been measured in public.
Will Grok 4.8 be available through the API?
Grok 4.6 launched with API access on day one, alongside Cursor and Grok Build, so it is likely. xAI has not confirmed it, and there is no grok-4.8 model entry in its documentation yet.
References
Elon Musk on X: Grok 4.8, a 2.5T model trained with our new C++ software stack
Elon Musk on X: Grok 4.7 roughly on par with Opus 5.0
Elon Musk on X: In-house AI training stack in C
Elon Musk on X: Grok 4.7 needs a few more days to cook
Introducing Grok 4.6 (SpaceXAI)
Introducing Grok 4.5 (SpaceXAI)
Open Release of Grok-1 (SpaceXAI)
SpaceXAI API Models and Pricing
Introducing Claude Opus 5 (Anthropic)
Claude Models Overview (Anthropic)
Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM
Training Compute-Optimal Large Language Models
ZeRO: Memory Optimizations Toward Training Trillion Parameter Models
Elon Maps the Grok Roadmap Through Grok 5 After 4.8 Training (TeslaNorth)
Elon Musk Shelves Grok 4.7 and Unveils Grok 4.8 Instead (Startup Fortune)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.