xAI Launches a Model That Is Better Than the Last One, Except on the Tests

xAI released Grok 4.7 last week as, in the company’s own framing, its most capable model yet for coding and professional knowledge work — a larger model, trained longer, with more reinforcement learning aimed at multi-hour tasks, priced identically to its predecessor at $2 per million input tokens and $6 per million output. This is, on its face, a normal AI launch: bigger number, same price, confident adjectives. What makes it a story is that the independent scoreboard everyone actually trusts told a different story than the one xAI told, and a coding-tool CEO and an AI commentator then spent several days publicly arguing about whether xAI’s own numbers were the wrong numbers to be arguing about.

On the Artificial Analysis Intelligence Index — a composite of ten benchmarks maintained by a lab with no horse in this particular race — Grok 4.7 scored 46. Claude Fable 5.1 and GPT-6 both scored 53. That is not a rounding error; that is Grok 4.7 landing in the middle of the pack it was launched to lead. xAI’s own release materials, meanwhile, showed 4.7 beating 4.6 on all seven of its selected benchmark rows. Both of these things are true at once, which is possible only because they are measuring different things, using different yardsticks, selected by different parties with different interests in the outcome.

The more interesting fight was about money, not intelligence. AI commentator Theo Browne ran his own tests and found Grok 4.7 burning roughly 81,000 tokens per task — more than double Grok 4.6’s 38,000 — pushing real-world cost to about $3.74 per task, higher than GPT-6 Astra’s $3.26. His single most expensive test run cost roughly $20; no other model in his comparison broke $10, save one Astra run at $11.44. His conclusion, delivered with the specific weariness of someone who has watched this exact genre of launch before, was that Grok 4.7 offers “last-gen performance at a current-generation price” — a headline number holding steady while the actual bill, measured in what a task costs a real user to complete, quietly doubles.

Cursor CEO Michael Truell disputed this, citing production data showing only a 5-to-30% increase in tokens per request. Browne’s response was the kind of distinction that sounds pedantic until you sit with it: a request isn’t a task. In an agentic workflow, one task can spawn dozens of tool-call requests, and if the new model simply makes more of those smaller requests, the per-request average can look flat while the per-task total — the number that actually appears on an invoice — climbs. It is the token-economy equivalent of a restaurant claiming prices haven’t gone up because each individual breadstick still costs the same, while quietly bringing you twice as many breadstick-shaped bills.

Here the story acquires its one unmissable wrinkle: Grok 4.7 is available inside Cursor at launch, xAI and Cursor share a corporate parent in SpaceXAI, and the executive publicly rebutting the “it’s more expensive than it looks” critique runs the coding tool that ships the model in question. Nobody involved has to be acting in bad faith for this to be an awkward position — Truell may sincerely believe his production data tells the fuller story — but it is nonetheless the case that the loudest defender of Grok 4.7’s cost profile has a seat at the table where that cost profile turns into revenue.

xAI’s own benchmark table has its own inconvenient footnote: Grok 4.7 scored lower than Grok 4.6 on Cognition’s Frontier Code test, an actual regression on an actual predecessor, which the company attributed to the new model tending to “overcope on others” — AI-industry phrasing for a model that talks itself into overthinking a problem it used to just solve. It is a strange thing to build into a press cycle organized entirely around the premise that the new model is better than the old one: an admission, buried in the fine print, that on at least one measure, it demonstrably is not.

None of this makes Grok 4.7 a bad model, exactly — it posted real gains on several benchmarks, ships at a genuinely low headline price, and will likely be perfectly serviceable for a large share of the people who use it. What it demonstrates, again, is that the AI industry’s benchmark tables have become a genre of fiction with real numbers in it: technically accurate, selectively framed, and reliably followed within seventy-two hours by someone with a spreadsheet and no launch to protect, pointing out that the number on the invoice tells a different story than the number on the slide.

Sources: