Elon Musk announces Grok 4.8: 2.5-trillion-parameter model, new C++ stack, reinforcement-learning run begins

On 14 September 2026 at 01:24 UTC, replying to @techdevnotes's question "Apart from Grok 4.7, what are we even expecting from SpaceXAI in september," Elon Musk wrote:

"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 single 22-word post (x.com/elonmusk/status/2099308197802631191) is the entire primary record: it names Grok 4.8 in public for the first time and makes four separate claims — the model exists as a training run; it has 2.5 trillion parameters; it was trained with a new C++ software stack; its pretraining ends this week and reinforcement learning begins. By 15 September the post had passed ~3 million views (progressive robot measurement via fxtwitter API: 3,037,955 views, 17,800 likes).

Ten hours later Musk ranked the roadmap (11:19 UTC, 14 September):

"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."

Minute before that, asked how close Grok 4.8 would come to AGI, Musk answered: "That will be Grok 5." On 15 September (wccftech), asked whether the C/C++ stack was itself AI-written, Musk replied: "Our C/C++ training stack was written by humans."

The immediate context that makes the announcement legible:

• Grok 4.7 has been slipping for seven weeks. Promised "in 4 weeks" on 24 July, "ready in 3 to 4 weeks" on 12 August, "in 10 days" (→ ~12 Sep) on 2 September, Musk said on 11 September (also on X) that Grok 4.7 "needs a few more days to cook," explaining: "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." As of 17 September, Grok 4.7 is still unreleased and xAI's current flagship remains Grok 4.6 (released 12 August, 1.5T per Musk, $2/$6 per million tokens, 500k context — verified on the docs page).
• The C++ stack has been teased since May. On 28 May Musk wrote that SpaceX had "almost finished writing V1.0 of an in-house AI training stack in C" with "potential speed improvement vs JAX for large training runs... over an order of magnitude," "exact-maps to 220k GB300s with 800G NICs," and heavy pipeline parallelism; on 29 June he said the "entire training and inference stack written in C/C++" would arrive in ~3 months with "most software layers... deleted completely"; on 8 July he said Grok 4.5 was "not yet using" the C/C++ inference software and "doubling or more of the current speed is probably achievable." Grok-1 (March 2024) was trained on "a custom training stack on top of JAX and Rust." The 14 September post is the first numbered Grok claimed to be trained on the stack.
• Policy-and-safety backdrop inside the window: on 12 September Musk endorsed Anthropic CEO Dario Amodei's "We Must Pace the Frontier" essay with "Dario is right" (S15); on 14 September President Trump rejected new AI "guardrails"; and xAI's government push continued (Grok for Government live on the DoD's GenAI.mil at IL5 since 31 August; FedRAMP High pursuit since spring 2026). The discovery record claims a "Grok 4 acceptance of state AI-safety policies" preceded this — not corroborated in this pass (section 1).
• Competitive cadence: Anthropic shipped Fable 5.1 and Mythos 5.1 on 1 September, Google shipped Gemini 3.8 Flash on 2 September, OpenAI published GPT-6 Astra's safety overview on 3 September — three major labs shipped while xAI explained a delay.

• Before: The public xAI roadmap ended at Grok 4.7 (2.1T per Musk, unreleased after seven weeks of slippage, stuck in an RL rework). The only released Grok 4-series flagship was Grok 4.6 (1.5T, $2/$6). The C/C++ stack existed only as Musk's in-progress claim about future training ("almost finished writing V1.0," May) with no named model attached.
• Change (event): Musk publicly named Grok 4.8 — a 2.5T model that is the first numbered Grok claimed to be trained on the new C++ stack — and attached a concrete near-term milestone to it ("finish training this week and start RL"). He also publicly recalibrated expectations: Grok 4.7 is now pitched at parity with Opus 5.0 (not 5.1), 4.8 as "a noticeable improvement," with the "Astra/Fable class" leap pushed to 4.9 and "better than anything" / AGI pushed to Grok 5 — a visibly lower bar than his 12 August claim that Grok 4.7 "will exceed all current models."
• After: xAI's public story is now "one flagship released (4.6), one in RL rework (4.7), one finishing pretraining (4.8), two on the horizon (4.9, 5)" — a conveyor-belt framing that converts every release into a checkpoint in an ongoing campaign. For developers and enterprises, nothing changed operationally: grok-4.6 remains the only usable xAI frontier model, and Grok 4.8 remains untestable.

Everything technical about Grok 4.8 is company claim; the mechanism vocabulary is still interpretable:

1. What "2.5T" likely means. xAI's only published architecture is Grok-1's: a 314B-parameter Mixture-of-Experts model with 25% of weights active per token. If Grok 4.8 follows the MoE pattern, 2.5T is most plausibly a total parameter count; the per-token compute depends on an unpublished active count. A dense 2.5T model would be a very different, far costlier system to serve. Musk's arithmetic line (0.5T v8-small → 1.5T V9/Grok 4.6 → 2.1T Grok 4.7 → 2.5T Grok 4.8) all comes from posts, not documentation.
2. What the C++ stack is claimed to do. Per Musk's May–June posts: an in-house training stack written in C (later described as C/C++, then simply C++) that "exact-maps" to 220,000 NVIDIA GB300s with 800G NICs, uses heavy pipeline parallelism, strips out intermediate framework layers ("most software layers will be deleted completely"), and claims "over an order of magnitude" training speed against JAX plus "doubling or more" inference speed once the C/C++ inference software is live (which, as of 8 July, it was not for Grok 4.5). The 14 September post is the first attribution of an actual numbered training run to that stack. No benchmark, throughput figure or technical report supports any of it publicly.
3. What "finish training this week and start RL" means. Pipeline: pretraining → supplemental training (e.g., SpaceX engineering data) → SFT → RL (task attempts scored by graders; where agentic persistence, tool use and self-checking are learned). For Grok 4.5/4.6, xAI's own release notes describe RL over "hundreds of thousands of tasks" and "highly asynchronous training" — RL is a full training programme, not a finishing touch. And RL is exactly the stage where Grok 4.7 stumbled: over-penalizing response length taught it to give up early on hard tasks and skip self-verification. So "start RL this week" is the beginning of the risky phase, not the end of it.
4. Serving arithmetic. 2.5T parameters ≈ 5 TB of weights at 16-bit (fits one GB300 NVL72 rack's 20 TB) but ~40 TB of training states with Adam at 16 bytes/parameter — hence the multi-rack pipeline design. If the C/C++ inference software (not yet confirmed live for 4.5/4.6) is ready at launch, the size increase could be partly absorbed; if not, expect slower serving and/or a price step-up from 4.6's $2/$6.

• It positions xAI as the only lab running a "monthly checkpoint" frontier campaign at trillions-of-parameters scale. No competitor publishes a ladder of four future runs (4.7, 4.8, 4.9, 5). Whether or not the numbers hold, the frame — that xAI's model line is a continuously refilled pipeline rather than discrete products — shapes how the market prices xAI's (and SpaceX's) AI narrative, and it pressures rivals to pre-announce.
• It makes the C++/C/C++ stack the most consequential, least-documented infra bet in frontier AI. An in-house, human-written (per Musk, 15 September) C++ training stack claiming order-of-magnitude gains over JAX anywhere near the frontier, if real, is an infrastructure moat (hardware-specific, latency- and cost-advantaged); if overstated, it is a schedule risk that compounds with every 4.x release. Grok 4.8 is its first public test.
• It sharpens the "pace rhetoric vs. race roadmap" contradiction. Musk endorsed slowing the frontier on 12 September ("Dario is right," S15) and, two days later, announced the next two rungs of an aggressive scaling ladder. Analysts (TeslaNorth, progressive robot) immediately flagged the tension. It matters because xAI's safety posture is now a market-facing variable: the same week, Grok for Government went live at IL5 (31 Aug) and the Trump administration rejected "guardrails" (14 Sep).
• Its evidence status matters methodologically. This is a canonical "announcement without an artifact": xAI has published no model card, no API ID, no price, no benchmark. For the AI-news ecosystem, distinguishing "Musk said X" (CONFIRMED) from "xAI shipped X" (FALSE as of 17 Sep) is the core editorial discipline of this story.

• For xAI: a credible ("noticeable improvement"-level) Grok that could ship after the stack rewrite — making the order-of-magnitude infrastructure claims testable against a real product; and a recovery narrative that routes around the stalled 4.7 (Musk's reply was literally to a question about "what else... in September beyond Grok 4.7").
• For the frontier conversation: a concrete reference point for "how fast can a lab move if it owns its hardware and rewrites its stack" — the counterfactual to OpenAI/Anthropic/Google's quarterly-ish cadence; and, per the recalibrated ladder, an explicit benchmark target (Opus 5.0 parity → Astra/Fable class) that rivals' shipped models can now answer.
• For the public record: a falsifiable claim pair — "2.5T on a C++ stack, RL from ~week of 14 September" — that later xAI artifacts (model card, benchmarks, launch post) will confirm or refute.

1. Is 2.5T total, active, or dense? (Only a model card settles this.)
2. Did training actually finish in the week of 14 September, and has RL begun? (No public signal either way as of 17 September.)
3. Does the C++ stack deliver any of the claimed order-of-magnitude speedup or 2× inference gain? (No benchmark exists; 4.8's launch latency/price is the proxy.)
4. Is the C/C++ inference software live yet? (Musk said on 8 July it was not serving 4.5; no update since.)
5. Which RL design choices does 4.8 use to avoid 4.7's over-penalized-length failure? (Unknown — and the most technically interesting question.)
6. When does Grok 4.7 actually ship, and does 4.8 ship before or after it? (Progressive robot's arithmetic: if 4.8 finishes main training ~20 Sep and needs 4.7-scale RL time, release could slide toward late October.)
7. What price and context window will 4.8 carry — does the discount vs. Opus/Fable/Astra survive at 2.5T?
8. What exactly was the "state AI-safety policies" claim in discovery referring to? (Unresolved in this pass.)

• Immediate (this week): any xAI signal that pretraining concluded and RL is running — or, typically, silence; a grok-4.7 release may land first ("few more days" was 11 September).
• 1–3 weeks: watch for a Grok 4.7 launch (model ID, price, benchmarks) — the first test of whether the RL rework works; and any first mention of grok-4.8 in xAI docs/infrastructure (the pattern that preceded 4.7's date-stamped build reports).
• 4–8 weeks: earliest plausible Grok 4.8 launch window if RL is fast (progressive robot's arithmetic suggests later — possibly late October); launch artifacts to demand: parameter count (total vs. active), context window, long-context pricing, C/C++-inference confirmation, benchmark harnesses.
• Structural: whether the "ladder" (4.7→4.8→4.9→5) becomes xAI's standing release pattern, and how rivals' shipped models (Fable 5.1, Opus 5, GPT-6 Astra, Gemini 3.8 Flash) move the benchmark spread the ladder claims to climb.
• Verification latch: the story's claims become independently verifiable only when xAI publishes a model card, a launch post, or a docs entry — at which point the "2.5T / C++ stack / RL" claims swap from COMPANY CLAIM to CONFIRMED-or-refuted.

The news is not the model; the news is the announcement. Grok 4.8 does not exist as a product — no ID, no price, no card — and the entire public record is two founder posts (one of them in reply to a random account) that passed ~3 million views each. Report it as a company claim with a confirmed timestamp: Musk said a 2.5T Grok trained on the new C++ stack will start RL this week. That it was said is fact; that it is true is unverified; that it will ship is a wager against Grok 4.7's seven-week slip pattern and the RL failure mode that caused it. The two genuinely novel elements — a first numbered model on xAI's rewrite-the-stack bet, and a CEO endorsing frontier pacing while publicly racing it — are the story. Pre-announcement is a growing frontier habit; the discipline it demands is separating the tweet from the artifact, every time.

Key numbers for the report: 2.5T (Musk-stated; no model card), Grok 4.6 current flagship at $2/$6 and 500k context (docs-verified), 4.7 delayed 25+ days past its 24 July estimate, announcement 01:24 UTC 14 Sep 2026, no grok-4.8 artifact as of 17 Sep 2026.