Elon Musk’s xAI finally stopped rewriting the calendar. Grok 4.7 launched Monday afternoon as the lab’s best model to date—“a notable improvement over Grok 4.6 at the same price and speed”—after at least five public timeline slips since late July.
Decrypt’s Jose Antonio Lanz frames the release as capability without coronation: bigger weights, familiar pricing, immediate availability, and a leaderboard habit that still reads “silver medal” on several economically meaningful tests.
From “four weeks out” to “it’s live”
Musk’s public countdown since July bounced from four weeks to a few weeks to 3–4 weeks to “10 days” on September 1 to “needs a few more days to cook” on September 11. Monday’s drop skipped the waitlist theater. Grok 4.7 is live in the Grok app, Cursor, Grok Build, and the xAI API. GitHub’s changelog separately notes Copilot rollout for Pro through Enterprise SKUs under usage-based provider pricing.
xAI’s product pitch leans on longer thinking on hard problems, more self-checking than 4.6, and what it calls its strongest safety guardrails yet. Musk’s own one-liner on X: a strong combo of intelligence, speed, and low cost.
2.1 trillion knobs and a SpaceX textbook
Parameter count jumps to 2.1 trillion—about 40% above Grok 4.6’s 1.5 trillion. API pricing stays put at $2 per million input tokens and $6 per million output tokens. The training story adds a physical-world twist: supplemental SpaceX data spanning Starlink telemetry, manufacturing records, and engineering failure logs, aimed at hardware and systems reasoning beyond pure web text.
That industrial diet is marketing catnip for robotics, aerospace-adjacent coding, and “why did this part fail” vibes. It does not automatically mint first place on office-work benches.
Second place, different fonts
On GDPval—professionally vetted knowledge-work tasks scored like chess Elo—Grok 4.7 landed at 1695 behind Claude Fable 5.1’s 1735. Artificial Analysis’s AA-Briefcase multi-hour office suite showed the same order: 1657 versus Fable’s 1678. CursorBench 4.0 put Grok in the expensive-middle for coding tasks inside the editor—costlier per task than some OpenAI/Anthropic options and still short of Fable 5.1’s sweep across price points.
The pattern is not new. Grok 4.5 debuted with industry-scale training compute and still trailed Claude and OpenAI on key scores; 4.6 lagged on coding autonomy; earlier Grok 4.20 traded reliability for speed and personality. Musk himself tempered expectations pre-launch, saying 4.7 should land roughly on par with Claude Opus 5.0 rather than newer Opus 5.1, with multimodal still needing work—and sketching 4.8, 4.9, and Grok 5 as the ladder toward frontier leadership without dates attached.
Why “almost first” still ships product
Most people meeting Grok are not running GDPval in a lab coat. They meet it on X, in a phone app, in a Tesla dashboard, or as a Cursor/API option that undercuts Anthropic and OpenAI on token economics. xAI’s recurring bet is blunt: good enough, cheap, and ubiquitous beats best-but-pricier for everyday volume.
For developers, the Monday checklist is practical. Same price as 4.6, more capacity, more self-verification rhetoric, SpaceX-flavored extras, and distribution into the IDEs and agents people already open. For leaderboard watchers, the story is continuity: xAI keeps closing gaps without owning the podium—yet.
Geeknewz read: Grok 4.7 is a shipping culture win after a summer of date roulette. Whether “notable improvement” becomes “new default” depends less on another Elo screenshot and more on whether second-place quality at first-place availability keeps winning the soft war for daily clicks.
Distribution is the silent benchmark
Benchmarks decide bragging rights; distribution decides defaults. Landing in Cursor, Build, the standalone app, social surfaces, car voice paths, and now GitHub Copilot means the new model does not need a clean sweep of Elo charts to matter. It needs to be the option already selected when someone opens a tab.
That is also why price discipline matters. Holding \$2/\$6 while adding capacity and aerospace-flavored training is a product strategy dressed as a rate card: keep the switching cost emotional, not financial.
Source: Decrypt