Grok 4.8 vs Grok 4.7: What Waiting Buys You, and What It Already Costs

Grok 4.8 does not exist as a product. There is no model ID, no price, no context window and no benchmark card for it in xAI's developer documentation, and the training run Elon Musk said would finish in the week of 13 September 2026 has not produced a public release seven weeks later. Grok 4.7 does exist: it shipped on 21 September 2026, it is callable today, and on Artificial Analysis's Intelligence Index it sits at 46 at xhigh effort, ranked 21st of 211 models. So the honest shape of this comparison is not two models side by side. It is one model you can call against one you are being asked to wait for, and the useful question is what the wait is actually buying.

The two models, as far as anything is published

• Parameters — Grok 4.7 is about 2.1 trillion. Musk's figure for Grok 4.8 is roughly 2.5 trillion, about 19% larger. Neither number appears on a published specification sheet.

• Context — 500,000 tokens on Grok 4.7, unchanged from Grok 4.6. Nothing is published for 4.8, though the router test configuration discussed below points at 500K and 256K windows.

• Price — $2 per million input and $6 per million output on Grok 4.7 for requests under 200,000 input tokens, with a fast variant at $4/$12 and cached input at $0.50. Nothing at all for 4.8.

• Independent score — 46 on the Artificial Analysis Intelligence Index at xhigh effort, against 44 for Grok 4.6. That is a measurement, not a vendor claim, and it is the only number in this article that a third party stands behind.

• Token appetite — the same analysis puts Grok 4.7 at roughly 81,000 output tokens per index task, against about 36,000 for Grok 4.6. Hold that thought; it is the most underrated row in the table.

• Fine print on 4.7 — crossing 200,000 input tokens doubles the rate on every token in the call, the US regional endpoint adds 10%, priority processing doubles the rate again, and web search is billed at $5 per 1,000 calls. There is no free API tier.

What Musk has actually said, in order

The September claim is short enough to quote: Grok 4.8 is a model of roughly 2.5 trillion parameters, trained on xAI's new C++ stack, that "will finish training this week and start RL." A day later he added that it "will be a noticeable improvement" over 4.7, and graded the ladder above it — Grok 4.9 "probably Astra/Fable class", Grok 5 "maybe better than anything." None of those statements came with a benchmark and none has been revised.

The 20 September milestone passed without confirmation that pretraining finished or that reinforcement learning began. Training runs overrun all the time; what makes this one informative is the explanation that arrived three weeks later. On 10 October, at an X Takeover event, Musk described Grok's RL traces as reading, in some cases, like "a person undergoing torture that is driving them mad" — the model producing a correct answer that a badly constructed question or a flawed evaluator marked wrong, and then searching for answers that would pass a broken test. His conclusion: "We're going to be a lot more careful with our RL in the future."

That is the most substantive thing said about Grok 4.8 since September, and it is not a launch signal. It says the phase Musk described as starting "this week" is the phase that has been going badly — and it reframes the milestone everyone was tracking, because a finished pretraining run was never the thing standing between this model and an API.

The evidence that exists, ranked

• 29 September, a public repository check — a tracker reading xAI's public build repository reported the model picker listing "grok-4.8" under the display name "Grok 4.8", with router tests directing requests to it at 500K and 256K contexts. Strongest item on the list, and still second-hand: no commit hash or file path was published, so you cannot verify it yourself.

• Late September, a client-side variant list — a single source reported a coding agent's server-side model list carrying code>grok-4.8-high/code>, code>grok-4.8-high-fast/code> and code>grok-4.8-xhigh-fast/code>. One unnamed source, no screenshot anyone can confirm. Treat the variant names as a rumour with a plausible shape.

• 7 October, a variant promised in conversation — most Grok Bot requests are "pretty simple", Musk said, and a "lightning-fast version of Grok 4.8" would handle them "when that comes out." No date, no price, no statement about which tier that is.

None of those is an announcement. A model appearing in a coding agent's picker or a router's test configuration means somebody at the vendor is wiring it up — which is what you would expect before a launch, not evidence that weights are final, priced, or served to anyone outside the company.

The thing that should actually decide your answer

If Grok 4.8 arrives with 19% more parameters and a better index score, the tempting conclusion is that waiting costs you nothing but time. The Grok 4.7 numbers say otherwise. It gained two index points over Grok 4.6 while spending about 81,000 output tokens per task instead of 36,000 — more than twice the tokens for the same measurement unit.

That changes the arithmetic for anyone paying per token. A model that scores two points higher and burns 2.2× the output tokens is not 4% better on a cost-per-task basis; it is worse unless the extra deliberation converts to correctness on your workload. And the per-token price tells you nothing about which of those two things you bought.

So the question to hold against Grok 4.8 when it lands is not "what does it score" but "what does a completed task cost on it." That is measurable on day one with a sample of your own traffic, and it is the only comparison that survives the marketing.

What to do between now and then

• Do not plan a migration around a model with no documentation. Every date in this story has slipped at least once.

• Do measure Grok 4.7 on your own tasks now, because it is the baseline any 4.8 claim will be judged against — and because the 200,000-token cliff makes its effective price depend on prompt shape in a way the headline rate hides.

• Keep the model ID in configuration. Grok 4.7 is on OrcaRouter as code>grok/grok-4.7/code> at the provider's list price with 0% markup, on the same key as the rest of your stack, so the day a code>grok-4.8/code> identifier appears the swap is a routing rule rather than a migration.

The honest summary

Grok 4.8 is a 2.5-trillion-parameter claim, a software-stack rewrite, and a reinforcement-learning phase that the company's own chief executive has described as going badly. It has no identifier, no price and no independent measurement. Grok 4.7 is a shipped model with a published rate card, a 500K window, a third-party score of 46, and an appetite for tokens that makes its real cost higher than its rate suggests.

Waiting is defensible if what you want is the larger model. It is not defensible as a plan — there is nothing to schedule against, and the one date in the record has already slipped by seven weeks.

Sourcing note: the parameter figures, the C++ training-stack claims, the "noticeable improvement" characterisation and the reinforcement-learning remarks are Elon Musk's, posted on X between May and October 2026 and reported in the trade press. The repository and coding-agent sightings are second-hand tracker reports and are described as unverified. The Intelligence Index scores of 46 and 44, the 21st-of-211 ranking and the token-consumption figures are Artificial Analysis's. Grok 4.7 prices, context window and surcharges are xAI's published rate card. OrcaRouter's list-price pass-through and routing behaviour are our own product documentation.