At 9:24 PM Eastern on September 13, 2026, someone on X asked what to expect from SpaceXAI in September apart from Grok 4.7. Elon Musk answered with a model nobody had named before: “Grok 4.8, which is a 2.5T model trained with our new C++ software stack, will finish training this week and start RL” (source). Twenty minutes later the post had 65,000 views. Two days earlier he had said Grok 4.7, the model before it, needs a few more days.
This page is the record for Grok 4.8: what has been said, by whom, when, and what xAI has published, which as of September 28 is still nothing. It sits next to our Grok 4.7 record, which carries the earlier model’s release on September 21 and the four missed dates before it. Every quotation below is from a post read at its own timestamp; dates are UTC unless marked Eastern.
On this page · 13 sectionsOpen
- What we actually know
- What Musk has said, word for word
- The stack that finally got a model
- The parameter ladder, all from one account
- The ladder, as he now grades it
- What finishing training has meant before
- Where this leaves Grok 4.7
- What would confirm it
- Where CellCog stands
- What we are watching for
- The tracker
- Update log
- Sources
Grok 4.8 exists in exactly two places as of September 14, 2026: a post by Elon Musk at 9:24 PM Eastern on September 13 naming it, and his post at 7:19 AM Eastern on September 14 grading it ‘a noticeable improvement’ over 4.7. xAI’s news page, model directory, API release notes and llms.txt contain no Grok 4.8, and no earlier Musk post uses the name.
His description, in full: a 2.5T model trained with xAI’s new C++ software stack, finishing training this week, then reinforcement learning. No date, no benchmark, no price, and no xAI document behind any of it.
It was posted while Grok 4.7 was still unreleased: the 4.7 target of September 11 had passed and Musk had said it needed a few more days. The successor was named two days later, in reply to a post asking what else to expect from xAI this month; 4.7 then shipped on September 21, eight days after the 4.8 post, at Grok 4.6’s price.
The C++ stack has its own timeline, all from Musk: a C training stack almost finished on May 28, the whole training and inference stack to be rewritten in C/C++ within about three months on June 29, and Grok 4.5 not yet on the new inference software on July 8. September 13 is the first time a numbered model is said to be trained on it.
Parameter counts, per Musk: a 0.5T model served all production traffic in May; the 1.5T V9-Medium became Grok 4.5 and 4.6; 4.7 is 2.1T; 4.8 is 2.5T. xAI has never published a parameter count for any Grok 4 model.
His own cadence puts a release weeks out: the 1.5T model finished training on May 25 and shipped as Grok 4.5 on July 16, 52 days later; the 2T run was to finish training the week of July 18 and Grok 4.6 shipped August 12, 25 days later. A model finishing training in mid-September lands, on those gaps, between early October and early November.
This is a living tracker. When xAI names Grok 4.8 on x.ai or docs.x.ai, a model id, a price or a launch post appears, the confirmed facts land here the same day next to a scorecard of every claim above.
§ 01What we actually know
| Claim | Status |
|---|---|
| Grok 4.8 exists and is in training | Musk on X, September 13 (9:24 PM Eastern); no xAI document |
| 2.5 trillion parameters | Musk on X, same post; unverified, no xAI spec for any Grok 4 model |
| Trained with a new C++ software stack | Musk on X, same post; the stack itself described by him since May 28 |
| Training finishes the week of September 14; RL follows | Musk on X, same post; no date for RL to end |
| Release date | Nothing from anyone |
| Benchmarks, price, context window, API id, model card | Nothing exists |
| Mentions of the name | Two, both Musk: September 13 (named) and September 14 (“a noticeable improvement” over 4.7); nothing earlier, nothing from xAI |
| Seen in Cursor’s model list | A commit to yetone/magpie, an open-source tool that reads coding agents’ model lists, 04:04 UTC September 28: Cursor’s server lists a grok-4.8 family and its CLI asks for grok-4.8-high-fast; no Cursor or xAI announcement |
| Where it sits on his ladder | Musk on X, September 14: 4.7 at Opus 5.0 parity, 4.8 “a noticeable improvement”, 4.9 “probably Astra/Fable class”, Grok 5 “maybe better than anything” |
| xAI surfaces (news, model directory, release notes, llms.txt) | No Grok 4.8; Grok 4.7 is the released flagship since September 21 (news index, launch post and model docs re-read that day) |
| Grok 4.7, the model before it | Released September 21, 2026, at Grok 4.6’s price; 40 days after 4.6 |
§ 02What Musk has said, word for word
The Grok 4.8 record is two posts, ten hours apart. The record for the stack it is said to run on is four months long, and every entry is his.
September 13, 9:24 PM Eastern (01:24 UTC September 14): “Grok 4.8, which is a 2.5T model trained with our new C++ software stack, will finish training this week and start RL” (source). The post is a reply; the question it answers reads “Apart from Grok 4.7, what are we even expecting from SpaceXAI in september”. Nothing about 4.7 in the answer.
September 14, 7:19 AM Eastern (11:19 UTC): “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.” (source) A reply to a post predicting that 4.8 would beat the unreleased models at Anthropic and OpenAI and that 4.7 would land with Fable 5.1 and Astra. He agreed with neither: 4.7 gets an Opus 5.0 grade, 4.8 an improvement, and the Fable and Astra class goes to 4.9.
May 28, 06:27 UTC: “SpaceX has almost finished writing V1.0 of an in-house AI training stack in C that exact-maps to 220k GB300s with 800G NICs, making heavy use of pipeline parallelism and getting as close to bare metal as possible.” (source)
May 28, 16:40 UTC: “Next will be writing the inference stack in C for simultaneous high-speed RL across a large block of GB300s.” and, in the same post, “(We do use a little C++ tbh, but not much)” (source)
June 29, 00:10 UTC: “Truly massive gains will come 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) and we exact-map Grok to work incredibly well on a GB300” (source)
July 8, 21:30 UTC: “Grok 4.5 is not yet using our internally developed C/C++ inference software that exact maps to the GB300 hardware. Doubling or more of the current speed is probably achievable.” (source)
Read together, the September 13 post is the first time the stack he has been describing since May is attached to a numbered model. Three months from June 29 is late September, so a model finishing training on the new stack in mid-September is inside his own window. Whether 4.7, whose initial training finished in August, was trained on the old stack or the new one, he has not said, and the Grok 4.7 launch post does not say either.
§ 03The stack that finally got a model
| Date (UTC) | What he said | What it implies |
|---|---|---|
| May 28 | A C training stack, V1.0 almost finished, mapped to 220,000 GB300s | Training side first |
| May 28 | Inference stack in C is next, for high-speed RL; a little C++ in use | RL is the target workload |
| Jun 29 | Entire training and inference stack in C/C++ in about three months | A late-September horizon |
| Jul 8 | Grok 4.5 is not yet on the new inference software | 4.5 shipped on the old stack |
| Sep 13 | Grok 4.8 is trained with the new C++ software stack | First model named on it |
The word changed along the way. May was a stack written in C, with a little C++ admitted; June was C/C++; September is a C++ software stack. That is a founder describing an engineering project in passing, not a spec sheet, and it should be read that way. What is consistent is the claim: fewer software layers, code mapped to one Nvidia system, and speed as the payoff.
§ 04The parameter ladder, all from one account
| Date (UTC) | Count | Attached to | Source |
|---|---|---|---|
| May 25 | 0.5T | The v8-small model serving all Grok production traffic at the time | Musk on X |
| May 25 | 1.5T | V9-Medium, finished training; became Grok 4.5 | Musk on X |
| Jul 18 | 2T | A model he tagged Grok 4.6 with a question mark | Musk on X |
| Jul 28 | 1.5T and 2.1T | Grok 4.6 is the 1.5T model; Grok 4.7 the 2.1T | Musk on X |
| Sep 13 | 2.5T | Grok 4.8 | Musk on X |
| Any date | Any count | Any Grok 4 model | xAI: none published |
Two cautions. The July 18 post attached a 2T model to Grok 4.6 with a question mark, and ten days later the 4.6 label went to the 1.5T model with the 2T-class run renamed 2.1T and called 4.7, so the founder’s own mapping of counts to names has moved once already. And total parameter counts for mixture-of-experts models describe memory, not speed; he said 4.7 would be “slightly slower to serve” than 4.6 for exactly that reason (xAI then shipped it at the same price and speed as 4.6), and 2.5T on a new stack could be faster or slower than 2.1T on the old one. Nobody outside xAI can say.
§ 05The ladder, as he now grades it
Ten hours after naming 4.8, Musk graded the whole line in one reply. It is the first time he has placed a Grok model beside a named competitor’s model instead of above the field.
| Model | September 14 grade | What he said before |
|---|---|---|
| Grok 4.7 | “roughly on par with Opus 5.0, not 5.1”; multimodal to fix | “will exceed all current models” (August 12) |
| Grok 4.8 | “a noticeable improvement” | 2.5T, new C++ stack, training ends this week (September 13) |
| Grok 4.9 | “probably Astra/Fable class” | Not on any earlier date in this record |
| Grok 5 | “maybe better than anything. We shall see.” | Not placed relative to 4.7 and 4.8 before |
Read against the parameter table above, the grades and the counts move together: 4.7 at 2.1T reaches Opus 5.0, 4.8 at 2.5T improves on it, and the Fable 5.1 and Astra class, the two models he was asked about, is two releases further out. For anyone waiting on 4.8 specifically, the useful line is the middle one: noticeable, not frontier.
§ 06What finishing training has meant before
The only calibration for “will finish training this week and start RL” is what happened the last two times he posted that shape of sentence.
| Model | The post | Posted (UTC) | xAI release | Gap |
|---|---|---|---|---|
| Grok 4.5 | “Grok foundation model V9-Medium (1.5T) has finished training.” then “2 to 3 weeks to public release.” | May 25 | July 16 | 52 days |
| Grok 4.6 | “2T Grok (Grok 4.6?) will finish training next week.” | Jul 18 | August 12 | 25 days |
| Grok 4.7 | “Initial training is complete” (from the August 12 post; SpaceX data added in supplemental training) | Aug 12 | September 21 | 40 days |
| Grok 4.8 | “will finish training this week and start RL” | Sep 13 | Not released | Open |
Apply those gaps to a training finish in the week of September 14 and Grok 4.8 lands between about October 9 and about November 6, if it follows either precedent. The 4.7 gap, closed on September 21 at 40 days, sits inside that range. The May 25 post also promised “2 to 3 weeks to public release” for a model that took seven, so the shorter end of any estimate he gives has not held yet. None of this is a prediction. It is the arithmetic on his own record, and the record says the release comes later than the first sentence implies.
§ 07Where this leaves Grok 4.7
Grok 4.7 shipped on September 21, 2026, eight days after 4.8 was named and 40 days after 4.6, at Grok 4.6’s exact price row; xAI’s own table places Fable 5.1 Max ahead on four of its seven rows. Grok 4.5 shipped on July 16 and was superseded by 4.6 on August 12, 27 days later, so the question for anyone on 4.7 is how long it stays the flagship: on the September 13 account, 4.8 was finishing training as 4.7 shipped. The Grok 4.7 record carries the spec, xAI’s benchmark table and every Musk claim graded. The September 14 grade held its shape: 4.7 arrived at 4.6’s price with an Opus 5.0-class self-assessment, and the frontier claim stays with 4.9.
§ 08What would confirm it
- A line on x.ai or docs.x.ai. A news post, a model page, a release note or a model id containing 4.8. Grok 4.6’s launch post arrived the same day as its API, and so did Grok 4.7’s.
- Cursor naming it. On September 28 a commit to yetone/magpie, an open-source tool that manages coding agents’ model settings, said Cursor’s server now lists a grok-4.8 family, with the CLI asking for grok-4.8-high-fast. Cursor put Grok 4.7 live the day xAI shipped it, so a Cursor announcement or a line on its models page would sit close to a release. Until then it is a string on a server, read by a third party.
- Musk saying RL has started. His own post makes “start RL” the next milestone; a post confirming it dates the clock.
- A 4.7 release: done, September 21. 4.8 was not folded into it; the open question is how long 4.7 stays the flagship.
- A price relative to 4.6. Grok 4.7 kept Grok 4.6’s row, $2 per million input tokens and $6 output on xAI’s API. A larger model on a faster stack could move either way; the Grok Bot pricing page changes with it.
- Anyone outside xAI naming a benchmark. No leak, no sheet and no anonymous checkpoint has been attributed to 4.8. The last stealth model guessed to be a Grok, Ox Alpha, was GLM-5.3-Flash.
§ 09Where CellCog stands
Our conflict, declared: we build CellCog, a platform where a business hires AI employees that keep memory and work as a team, and we choose the models underneath them. Since September 26, 2026 that is Claude Opus 5.5 at every tier, with the tier setting how hard it thinks. We do not route to Grok. We track xAI’s models the way an operator does, as candidate engines, and a Grok 4.8 that is fast, cheap and good at agentic work would get tested on our infrastructure the day it ships. Until then, this page records what has been said and refuses to round it up.
§ 10What we are watching for
- The xAI surfaces, several times a day: news feed, model directory, release notes, llms.txt.
- Cursor’s announcements forum and model list, since a grok-4.8 family appeared on its server on September 28.
- Musk’s account for a training-done or RL-started post, and for any date. The September 14 grade is the only new fact since the name; a date is what would move this page.
- How long Grok 4.7 stays the flagship: a 4.6-style four-week turnover would date 4.8 early.
- The stack claim spreading to 4.7: if he says 4.7 was also trained on the new stack, the first-model-on-it framing here changes.
§ 11The tracker
This page updates when facts change, not when posts get louder. As of September 28, 2026: Grok 4.7, the model before it, has shipped; a third-party tool’s commit shows a grok-4.8 family on Cursor’s server; Grok 4.8 has been named twice, by its founder, with a size, a stack, a stage and, on September 14, a grade (“a noticeable improvement” over a 4.7 he places at Opus 5.0 parity), and by nobody else. No date, no spec, no benchmark, no price exist in any xAI document. When xAI moves, the confirmed facts and a graded scorecard of every claim above land here the same day.
§ 12Update log
This is a living page; when the story moves, the update lands here.
As of September 14, 2026: page opened. Only source is the September 13 post; xAI surfaces read at 01:47 UTC, zero mentions.
September 14, 2026, 7:19 AM ET: second Musk post. “Grok 4.8 will be a noticeable improvement” over a 4.7 he now grades “roughly on par with Opus 5.0, not 5.1”; 4.9 “probably Astra/Fable class”, Grok 5 “maybe better than anything”. New: the ladder table; status, key point, word-for-word and tracker sections updated. x.ai/news read at 13:01 UTC, still zero mentions of 4.8 or 4.7.
September 21, 2026, 16:17 UTC: Grok 4.7 released, 40 days after 4.6 and eight days after 4.8 was named, at 4.6’s price. This page: status table, key point, glance and FAQ answers, the cadence table (4.7 gap closed at 40 days), the 4.7 section, the confirm and watch lists and the hero updated. xAI’s news index, launch post and model docs re-read the same day: zero mentions of Grok 4.8.
September 28, 2026, 15:23 UTC: a commit to yetone/magpie (04:04 UTC), an open-source tool for coding agents’ model settings, says Cursor’s server lists a grok-4.8 family and its CLI asks for grok-4.8-high-fast; the claim spread on X from 09:37 UTC. Cursor’s announcements forum, x.ai/news and OpenRouter carry no Grok 4.8 at 15:23 UTC. Status table, glance answer, confirm and watch lists updated; our routing line corrected to Opus 5.5.
§ 13Sources
Elon Musk on X, each read at the post’s own timestamp: September 13, 9:24 PM Eastern (Grok 4.8), replying to Tech Dev Notes; September 14, 7:19 AM Eastern (the ladder grades); May 25 (V9-Medium finished training); May 28, 06:27 UTC and 16:40 UTC (the C training and inference stack); June 29 (whole stack in C/C++); July 8 (4.5 not yet on the new inference software); July 18 (2T model); July 28 (1.5T and 2.1T); August 12 (4.7 initial training complete); September 11 (4.7 needs a few more days). xAI: news, model directory, Grok 4.6 launch post, read September 14, 2026, 01:47 UTC, and Grok 4.7 launch post, read September 21, 2026. Grok 4.5, 4.6 and 4.7 release dates as recorded on our Grok 4.7 record from x.ai. CellCog routing statements reflect our configuration on September 21, 2026.
The Cursor trace: a commit to yetone/magpie on GitHub, 112471fb, September 28, 2026, 04:04 UTC, read 15:20 UTC; the claim spread from a LuminaBench post on X at 09:37 UTC.
Q1Has xAI announced Grok 4.8?
No. We checked xAI’s news feed, its model directory, its API release notes and its llms.txt at 01:47 UTC on September 14, and re-read the news index, the Grok 4.7 launch post and the model docs on September 21 after 4.7 shipped: zero mentions of Grok 4.8. The only sources are two Musk posts, September 13 and September 14; a search of his own account for the name returns nothing earlier.
Q2What is the new C++ software stack?
Musk’s phrase for an in-house training and inference stack written in C and C++ to run as close to bare metal as possible on Nvidia GB300 systems. He first described a C training stack on May 28, said on June 29 that the entire training and inference stack would be in C/C++ within about three months, and said on July 8 that Grok 4.5 was not yet using the new inference software. Grok 4.8 is the first model he has said was trained on it.
Q3Is 2.5T a real specification?
It is Musk’s number, and his numbers are the only ones there are: 1.5T for the V9-Medium model behind Grok 4.5 and 4.6, 2.1T for Grok 4.7, 2.5T for Grok 4.8. xAI has not published a parameter count for any Grok 4 model, the Grok 4.7 launch post on September 21 included, and total counts for mixture-of-experts models say little about active parameters or speed.
Q4Will Grok 4.7 ship before Grok 4.8?
It did. Grok 4.7 shipped on September 21, 2026, while 4.8 was, on Musk’s account, finishing training. The question the September 13 post raised is now the live one: whether 4.7 becomes a short-lived release, the way 4.5 was replaced by 4.6 within four weeks.
Q5Should I wait for Grok 4.8?
For a model without a date, no. If you run on Grok today, Grok 4.7 is the released model, shipped September 21 at Grok 4.6’s price. If you are choosing where your agents run, the model underneath matters less than the layer around it: memory, task board, inbox and approvals a person can see. CellCog runs Claude Opus 5.5 at every tier since September 26; we do not route to Grok, and we would test 4.8 the day it ships.