xAI's 'same price' pitch on Grok 4.7 skips the part where the bill doubles

Here's what xAI's launch page for Grok 4.7 says about price: it didn't change. $2 per million input tokens, $6 per million output, same as Grok 4.6 five weeks ago.

Here's what the launch page doesn't say: how many tokens Grok 4.7 needs to actually finish a job. That number comes from Artificial Analysis, not xAI, and it's the one that should have been the headline.

Artificial Analysis tested Grok 4.7 at its highest reasoning setting and found it used roughly 81,000 output tokens to complete an average Intelligence Index task. Grok 4.6, at a comparable setting, used about 36,000. That's not a rounding difference — it's more than double, on a model xAI is selling as unchanged in price.

Run the arithmetic and Artificial Analysis's own cost-per-task figures do the rest of the talking: roughly $3.74 to finish an average task at Grok 4.7's top setting, about $2.73 at a lower one. Compare that with GPT-5.6 Sol Max — the model xAI's own launch chart uses as its expensive-rival comparison, at four times Grok 4.7's list price — which finishes the same class of task for about $1.99. The cheaper-per-token model is not the cheaper-per-task model. xAI's chart shows you the rate card. It doesn't show you the bill.

None of this makes Grok 4.7 a bad model. It makes 'same price' a half-true headline, and half-true headlines are exactly the thing worth reading past.

xAI's own seven-row comparison table puts Grok 4.7 at 37.6% on Terminal-Bench 4.0, up from Grok 4.6's 20.3%. That's the number in the launch post. VentureBeat reported that Artificial Analysis, running its own independent harness, measured Grok 4.7's Terminal-Bench 4.0 score at roughly 26% at its xHigh setting — well below xAI's figure, and well below GPT-6 Astra's independently measured 59.6% on the same test. Both numbers can be technically accurate under different testing conditions. Neither appears alongside the other on xAI's launch page, and only one of them is the one xAI chose to publish.

Coverage of this launch — Android Headlines among it — has settled on 2.1 trillion parameters for Grok 4.7's base model, up 40% from a reported 1.5 trillion for Grok 4.6, plus a claim that xAI folded in supplementary SpaceX engineering data: Starlink telemetry, manufacturing logs, failure-analysis records. Search xAI's actual launch page for either figure and you won't find it. Both trace back to Elon Musk's own posts on X, filtered through outlet reporting, not to anything xAI has put its name to in documentation. That doesn't make the claims false. It does mean anyone repeating '2.1 trillion parameters' as an xAI-confirmed fact is repeating Musk, not the company.

One more thing the coverage undersells: Grok 4.7 launching inside Cursor on day one isn't really a partnership story any more. SpaceX — xAI's parent since a 2026 merger — closed its $60bn acquisition of Cursor's maker, Anysphere, in mid-August. Cursor is now in-house distribution, not a third-party integration, which is worth knowing before reading too much into how fast Grok 4.7 reached a coding tool that used to be neutral ground between the major labs.

Grok 4.7 also arrives roughly eight weeks late by its own founder's early timeline — 'in 4 weeks' on 24 July became 'a few more days' by 11 September, with Musk conceding the model stopped too early on hard tasks and didn't check its own work rigorously enough. It landed a day before Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Sol and Luna both cut price. xAI is the only one of the three that didn't. Read that as confidence in the model, or read it as a company with less margin to give away — the token-efficiency numbers above suggest the second reading deserves at least as much weight as the first.