Grok 4.8 Is in RL: What the Promised Bot Rollout Actually Means

Grok 4.8 has finished pre-training and is now in reinforcement learning, with a limited rollout to Grok Bot users promised "later this week" — that is what SpaceXAI's owner said on X on 10 November 2026, and it is the first concrete schedule anyone has offered for a model that has been behind its own deadline since September. It is still not a model you can call. There is no grok-4.8 identifier on SpaceXAI's developer documentation, no rate card, no published context window and no independent score of any kind; the newest entry on that page is Grok 4.7, which shipped on 21 September 2026 with a 500,000-token context window and a list price of $2.00 per million input tokens and $6.00 per million output tokens. Everything in this piece about Grok 4.8 is a public claim by the vendor or its owner, unreproduced and unaudited, and it is worth reading as a schedule rather than as a specification.

The distinction matters because "in RL" is a statement about a training run, not about a product. A model can sit in reinforcement learning for weeks, can fail its evals, can be rolled back to an earlier checkpoint, and can be handed to a small group of users on one surface while remaining invisible to everyone with an API key. The useful question for a developer is not "is Grok 4.8 coming" — it plainly is — but which of the things you would need in order to build on it exist today. As of 11 November 2026, none of them do.

What was actually said, in the order it was said

The substance came in a reply thread on X on the morning of 10 November, and the ordering is worth keeping because each line constrains the next. First, the schedule: pre-training is done, Grok 4.8 is in reinforcement learning, and the evals are starting to climb. Then the reason for the delay, which is a rewrite rather than a research problem:

"We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time. Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right. Now, 4.8 is in RL…"

Asked what the new stack is written in, the answer was C with a little C++, "but not too much", plus some assembly, built on work from Starlink engineers, with the team writing its own headers and calculating its own time including leap seconds. The framing offered was that superintelligence may one day write a better software stack but has not yet, and that this is tight code. Treat the engineering detail as vendor narrative — it is unverifiable from outside, and no part of it is a benchmark. What it does explain is the gap between a model that was announced for late September and a model that still has no identifier in November: the company rebuilt the machinery underneath the training run, and that work shows up as calendar time, not as capability.

Two claims in the same thread are the ones a reader should actually carry forward:

• "Grok 4.8 will bear no resemblance to 4.7 (thankfully)." A statement about the size of the change, made by the person with the most incentive to overstate it, and offered with no number attached.

• The new version is "not yet incredible", but "the foundations are built right" and it will "evolve to be something special". This is the rare case of a vendor pre-emptively lowering expectations for a model it is about to ship, and it is more informative than the enthusiasm around it.

What was not said matters just as much. There was no date for a wider release, no mention of the API, no context window, no parameter count, no price, and no benchmark of any kind. The rollout promised for this week is explicitly limited and explicitly to Bot users, with the stated intent being to "see how it goes".

Eight weeks of slipping, and what the picker traces were worth

The reason this week's promise is being read as news rather than as marketing is the history behind it. Grok 4.8 was first described in mid-September at roughly 2.5 trillion parameters on a new C++ stack, with a deadline around 20 September. That deadline passed. Grok 4.7 shipped instead, on 21 September, and became the model the company has sold ever since. Since then the only signals have been indirect: a grok-4.8 entry briefly visible in a first-party model picker with 500K and 256K context options at the end of September; a list of variant names — grok-4.8-high, grok-4.8-high-fast and grok-4.8-xhigh-fast — turning up in a third-party coding tool's model list; a remark in early October about a "lightning-fast version"; and mid-October comments about reinforcement learning on hard tasks being a form of torture that was "driving them mad".

Every one of those was a trace of work in progress, and none of them survived as a callable model. This is the pattern worth internalising: an identifier appearing in a picker or a client library means someone configured a deployment, not that a rate card exists. Prediction markets reflected the same uncertainty, pricing a Grok 4.8 release at roughly 13% by the end of October and around 65% by the end of November. The 10 November statement is the first thing since September that moves the second number on evidence rather than sentiment.

What "limited to Bot users" tells you — and what it does not

A gated rollout on a first-party consumer surface is a specific kind of event, and it is neither a launch nor a preview. It tells you the model can hold a conversation in the product's harness under supervision. It does not tell you:

• Whether an API identifier exists. Grok 4.7's Fast variant has been restricted to first-party surfaces since launch and is still not on the public API, so a working model in a product is no evidence of an endpoint.

• What the context window is. The 500K and 256K figures in the September picker trace are configuration strings from a UI, not documentation, and the vendor has published neither.

• What it costs. There is no rate card, and price is the one number that determines whether a model is usable for anything at volume.

• Whether it is better. A limited rollout is selected for by the vendor; the users in it are not a random sample and their reports, whatever they are, will not be a benchmark.

The historical base rate is that a Bot rollout precedes general API access by weeks, not days, and that the interval is set by safety review and by whether the inference stack holds up — and this is a model whose inference stack was rewritten from scratch in C and C++ a few weeks ago. If the rollout happens and holds, expect an identifier and a rate card to follow; if it slips again, the reason will be in the stack rather than in the model.

The baseline you can actually measure

Because Grok 4.8 has no published figures, the only honest comparison is against the model it replaces. The numbers below are Grok 4.7's, and they are not flattering, which is the point — they set the bar the new model has to clear rather than a bar it has already cleared.

On Artificial Analysis's Intelligence Index, Grok 4.7 currently sits at 35, which is rank 41 of the 74 models the evaluator scores and below the 40 that similar-priced models average. Do not compare that with the 46 that circulated when the model launched in September: the evaluator has re-based the scale, and 35 is the number to use for any comparison made today. The more revealing figures are the cost and time ones. Grok 4.7 burns $7.77 of tokens per Intelligence Index task, takes 25.9 minutes per task against a 12.8-minute median, and emits 148 million output tokens across the index against a 72.8-million median — it is, in the evaluator's phrasing, very talkative. Throughput is around 67 tokens per second, with a 53.6-second wait to the first answer token when reasoning is on.

There is a structural reason to expect Grok 4.8 to move those numbers rather than just the headline score. A model trained by reinforcement learning on long-horizon tasks tends to spend more tokens thinking, not fewer, so the token-efficiency and wall-clock figures are the ones to watch when the first independent measurements appear. If Grok 4.8 raises the index and also triples the cost per task, the practical answer for most workloads will still be a cheaper model.

The vendor's own comparison, which is the most useful thing said all week

Asked directly whether Grok 4.8 is better than Claude Opus 5.5, the answer on 10 November was "not as good", with Opus 5.5 called a "masterpiece". The follow-on was that Grok 5 should exceed Opus 5.5 and that this might happen in two or three months. That is a vendor conceding the current generation to a competitor and pointing at the next one, and it is the single most decision-relevant sentence in the thread: if you are choosing a frontier model today on capability, Grok 4.8 is not being offered as the answer, even by the people building it.

It also reframes the rollout. A limited Bot release is a way to gather signal on a model whose foundations are, in the vendor's own words, not yet incredible — not a way to take share from a stronger rival. Read the week's promise in that light and the realistic outcome is a small group of users getting a model that is being tested, ahead of a public release whose date nobody has stated.

What you can call today, and how to hedge

If you want Grok in production this week, the answer is Grok 4.7 through the API — grok/grok-4.7 is on OrcaRouter at SpaceXAI's own list price with no markup, the provider price passed through, so a vendor price change reaches you the same day rather than after a repricing pass. That matters on this model specifically because its rate card is tiered: $2.00 input and $6.00 output per million tokens below a 200K prompt, $4.00 and $12.00 at or above it, with cached input at $0.50 and a 75% cache discount on repeated prefixes. A long agent loop that re-sends the same system prompt every turn is paying the cache rate, not the headline rate, and the two-tier threshold means an unlucky prompt length can double the bill.

The hedge for an unproven model is the one this situation actually calls for. When Grok 4.8 does get an identifier, the sensible way to evaluate it is behind a failover rule rather than as a hard-coded model: route the traffic you care about to the model you have measured, put the new one in front on a fraction of it, and let the router fall back when it errors or times out. On OrcaRouter that is a routing configuration rather than a second integration — one key, one OpenAI-compatible endpoint, automatic failover across providers, and the same endpoint for every model in the Grok line, so moving a workload from Grok 4.7 to Grok 4.6 or Grok 4.3 to test the cost difference is a string change. It is also the reason not to hard-code anything to a grok-4.8 slug you saw in someone's client library: variant names in a third-party list are not a deployment, and a slug that 404s in production is a worse outcome than waiting a fortnight.

What would settle it

Four checkpoints, in the order they will probably arrive, each of which turns a claim into a fact:

• An identifier in the developer documentation — a grok-4.8 entry on SpaceXAI's model list with its own page. This is the gate for everything else and it is publicly checkable; the page currently stops at Grok 4.7.

• A rate card. Price is the number that decides whether a model is usable at volume, and a model with no published price has no production story.

• A context window and a knowledge cutoff. The 500K and 256K strings from the September picker trace are the only figures in circulation, and neither is documentation.

• An independent score. Artificial Analysis currently has no Grok 4.8 entry, and its Grok 4.7 baseline of 35 with $7.77 per task is the number any vendor benchmark will have to beat to be believed.

Until at least the first two of those exist, the correct reading of this week is narrow and worth stating plainly: a model that was supposed to ship in September is being tested with a small group of Bot users, its own maker says it is not yet better than the frontier, and the API that most readers of this page would actually use has not been told anything. The thing to watch is not the rollout announcement but the developer docs page — the moment a grok-4.8 row appears there with a price next to it, the wait is over, and until then every other signal is a schedule.