Five days ago, this blog covered Grok 4.7 as vaporware — a model Elon Musk promised “in about 10 days” on September 2, missed, and then pre-graded against a two-month-old Anthropic release while still not shipping it. That watch item is now closed. Grok 4.7 went live on xAI’s own model docs on September 21, and within the same news cycle it showed up as a selectable model in three separate, unrelated coding tools: GitHub Copilot, Cursor, and Devin.
That’s the headline. The more interesting story is buried in Cognition’s own announcement post, which does something vendors rarely do: publish a number that makes their own model choice look worse.
The numbers xAI wants you to see#
xAI’s launch post frames Grok 4.7 as a straightforward step up from Grok 4.6, with the same $2/$6 per-million-token pricing for prompts under 200K tokens (doubling to $4/$12 above that) and a 500K context window. Self-reported benchmark gains, per xAI:
- CursorBench 4.0: 46.3% (up from 40.4%)
- DeepSWE v1.1: 71.0% (up from 65.2%)
- Terminal-Bench 4.0: 38.0% (up from 20.3%)
- AA Briefcase v1.1: 1,657 (up from 1,546)
Standard caveat applies here, the same one this blog has applied to every self-reported benchmark table since the Terminal-Bench 4.0 statistical-tie piece: these are xAI’s own numbers, not yet cross-checked against an independent leaderboard. The Terminal-Bench jump in particular — nearly doubling from 20.3% to 38.0% — is the kind of gain that’s worth watching for independent confirmation before treating as settled. CursorBench is a joint xAI/Cursor benchmark, which is worth knowing given Cursor’s parent company (SpaceX, since its $60B acquisition of Cursor closed in June) is also the entity training the model being scored on it.
The number Cognition didn’t have to publish#
Devin’s own launch post for Grok 4.7 is where this gets genuinely useful. Cognition runs its own evaluation suite, FrontierCode 1.1, and its write-up states plainly that Grok 4.7 “scores below Grok 4.6 on the aggregate because it tends to over-scope, producing larger diffs with changes not requested by the evaluation task.”
Read that twice. This isn’t a company burying a bad number in a footnote — it’s a company that just added the model to its own product, in the same post announcing that addition, saying the new version of the model regresses on their own aggregate score because it can’t stop touching code outside the task it was given. Cognition does credit Grok 4.7 with real strengths: it’s apparently strong on “hard, multi-module backend tasks in Java, Go, and Ruby, where it investigates deeply, reads dependency source, and verifies its work end to end with real integration tests.” That’s not nothing. But the failure mode they chose to name — unrequested edits inflating diff size — is precisely the one that matters most for anyone deciding how much to trust an agent’s output without reading every line of it.
Why this is a Claude Code story, not just a Grok story#
Three vendors integrated the same third-party model on the same day: GitHub (Copilot, across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, and the CLI), Cursor (via its joint training arrangement with xAI), and Cognition (Devin Desktop and CLI). None of them built Grok 4.7. All three are, in this specific sense, distribution and harness companies wrapping someone else’s frontier model — the same commoditization-of-the-model-layer pattern this blog flagged with GitHub’s HydraFusion routing layer and again with OpenAI’s own Agents API beta exposing its harness as a rentable product. When the model layer is interchangeable, the harness — how tool calls get scoped, how diffs get reviewed, how much an agent is allowed to touch without asking — is where the actual differentiation lives.
That’s exactly where an over-scoping model becomes a stress test. In an IDE-anchored tool, the default interaction loop is “accept” or “reject” on a suggestion a human is skimming, often under time pressure, often on a diff that’s already larger than expected because the model decided to “helpfully” touch three adjacent files. A larger, more capable model that also produces larger, less-scoped diffs makes that skim-and-accept loop measurably riskier — not because Copilot or Cursor did anything wrong, but because the underlying model’s stated failure mode is structurally the hardest one for a glance-and-approve workflow to catch. Claude Code’s permission-gated, terminal-native model — where every file touch is a discrete, auditable action rather than a suggestion buried inside a larger diff — isn’t immune to a model that wants to over-scope, but it’s a better-matched harness for catching it before it lands, one tool call at a time rather than one diff-review pass at a time.
What to actually take from this#
Grok 4.7 is a real model with real gains on multiple benchmarks, and Cognition deserves credit for disclosing a regression against its own competitor model choice instead of quietly picking the flattering number. But the launch also hands the industry a clean, vendor-sourced case study in why “the model got smarter” and “the model got safer to trust unsupervised” are different claims — and why the harness wrapped around a model, not just the model’s benchmark table, is what actually determines whether that gap matters in production.
Sources: xAI Grok 4.7 announcement · xAI model documentation · GitHub Changelog, Sept 21, 2026 · Cursor model documentation · Devin blog: Grok 4.7