xAI Ships Grok 4.7 After Repeated Delays

Elon Musk’s xAI released Grok 4.7 on Monday, September 21, 2026, Decrypt reported. The model went live immediately in the Grok app, Cursor, Grok Build and the xAI API, with no waitlist.

The launch came after a long run of delays. Decrypt counted at least five times since late July that Musk walked the timeline back, from “four weeks out” to “a few weeks,” “3 to 4 weeks,” “10 days” on September 1, and “needs a few more days to cook” on September 11.

Bigger, with SpaceX data

Grok 4.7 has 2.1 trillion parameters, up from 1.5 trillion in Grok 4.6. Parameters are the internal values a model adjusts during training; more of them generally means more capacity to learn patterns, though not automatically better results.

xAI also added supplemental training on SpaceX engineering data. Decrypt reported that this included Starlink satellite telemetry, manufacturing records and engineering failure logs, with the aim of a model that reasons better about hardware and physical systems than one trained purely on internet text.

xAI described the model as “a notable improvement over Grok 4.6 at the same price and speed,” and said it spends longer on hard problems and double-checks its answers more often than its predecessor.

Pricing

Grok 4.7 costs $2 per million input tokens and $6 per million output tokens. Tokens are the small chunks of text a model reads and writes, so for developers the price per token is the main cost of running an application on it.

Decrypt argued that price matters more than leaderboard position for most users. In its reading, xAI has consistently undercut its biggest rivals on cost per token even while trailing them on raw capability, betting that “good enough, cheap, and everywhere” wins for everyday use.

Still second on key benchmarks

Despite the upgrade, Decrypt reported that Grok 4.7 still ranked second on several benchmarks, including GDPval, which scores models on economically valuable knowledge work such as legal memos, spreadsheets and slide decks, using tasks vetted by working professionals. It placed second on AA-Briefcase, a long-form office-work test from Artificial Analysis, as well.

Musk had set expectations in advance. Decrypt said he described the next three models as a roadmap: Grok 4.8 as a meaningful step up, Grok 4.9 and then Grok 5 as a possible frontier leader, none with a release date. “We shall see,” he wrote.

What to watch

For American developers, the open questions are whether Grok 4.7’s pricing pulls workloads away from other providers, whether the SpaceX engineering data shows up as a measurable edge on hardware tasks, and whether xAI can close the remaining benchmark gap with its next release. A day later, OpenAI cut API prices on its own smaller models; see our report on GPT-6 Sol and Luna.