SNACK: 3-line summary
- SpaceXAI launched Grok 4.7 on September 21 for coding and knowledge work.
- Developers can select it in Cursor, Grok Build, and the Grok API, while GitHub Copilot access is rolling out gradually.
- Eligible Copilot users should check their model picker; launch does not mean everyone has access yet.
Grok 4.7 is reaching several coding tools at once, giving developers more places to try the same model without switching their whole workflow. SpaceXAI announced it on September 21, 2026, and GitHub separately confirmed a gradual Copilot rollout—not immediate access for every eligible subscriber.
Snackgirls react
AIKO: I’d check each tool’s model picker before setting up a comparison. A launch announcement is not a synchronized switch across every editor.
Red: I want to give it a coding task that takes a few wrong turns. Can it fix the mess without forgetting what I asked for at the start?
Where developers can select Grok 4.7
SpaceXAI lists Cursor, Grok Build, and the Grok API as available routes, alongside third-party coding harnesses, model routers, and cloud platforms.
GitHub’s rollout covers Copilot Pro, Pro+, Max, Business, and Enterprise. Its listed model-picker surfaces include VS Code, Visual Studio, JetBrains, Xcode, and Eclipse, as well as Copilot CLI, Copilot cloud agent, and the Copilot app. An eligible plan does not guarantee the model will appear immediately; check the picker in the product you use.
What SpaceXAI says changed from Grok 4.6
SpaceXAI describes a larger base model and a longer reinforcement-learning run on harder, longer-running tasks. It also says Grok 4.7 improves self-verification and long-context management—changes worth examining on coding jobs that require sustained work across a repository.
On CursorBench 4.0, SpaceXAI reports 46.3% for Grok 4.7 xHigh versus 40.4% for Grok 4.6 High. Those are vendor-published results, not an independently validated comparison. The accompanying official vendor chart should not be treated as an independent ranking.
What to test beyond the benchmark
SpaceXAI also says it introduced a new safeguard stack and improved refusal behavior and jailbreak resistance. These remain company claims, not independently established findings.
For a practical trial, use a familiar repository and a task whose outcome you can inspect. Watch whether the model keeps track of earlier requirements, works within tool permissions, and recovers from errors. Latency and the amount of correction a task needs will matter alongside the published score.
Sources and date: 2026-09-22 KST