Grok 4.7 arrives in Devin with a scope-control tradeoff

Cognition has added xAI’s Grok 4.7 to the Devin Desktop app and Devin CLI, according to the Devin announcement. During the week of September 21, 2026, the model also reached GitHub Copilot and Cursor.

Cognition’s internal evaluation shows why developers may choose the model selectively. Grok 4.7 performed well on complex backend work that required broad codebase exploration and integration testing, but its tendency to modify adjacent code lowered its aggregate score.

Deep investigation, bigger diffs

Cognition evaluated Grok 4.7 on FrontierCode 1.1, its proprietary benchmark based on pull requests from open-source repositories. The benchmark grades correctness and mergeability, meaning whether a reviewer could reasonably accept the resulting change.

Grok 4.7 scored 59.4% on the Extended split. Cognition found its strongest results in difficult, multi-module Java, Go, and Ruby tasks. The model often inspected dependency source code, traced behavior across modules, and validated changes with integration tests.

Those broader investigations also produced unnecessary edits. Cognition reports that Grok 4.7 finished slightly behind Grok 4.6 overall because it expanded task scope and generated larger diffs. A targeted bug fix, for example, could include unrequested refactoring in neighboring modules, increasing review time and regression risk.

A benchmark shaped like code review

FrontierCode grades submissions against reviewer-defined rubrics. Some criteria are designated as blockers, and failing one caps the score for that run, similar to a required change during code review.

For version 1.1, Cognition audited more than 1,000 blocker criteria and reclassified 75 that it considered too strict. The revision reduces cases where a minor issue imposes the same penalty as a defect that would prevent a pull request from merging.

The benchmark permits internet access for documentation and other legitimate research. A prompt-based classifier and programmatic verifier check whether an agent retrieved the source pull request containing the answer. Runs that fetch the original solution receive a score of zero.

Because FrontierCode is proprietary, its 59.4% result should be treated as Cognition’s measurement rather than a directly reproducible industry score. Its review-oriented design still provides useful evidence about diff quality, task scope, and merge readiness.

A larger base model at lower token rates

Grok 4.7 uses a new base model with 2.1 trillion parameters, up 40% from Grok 4.6’s 1.5 trillion. Parameters are learned weights within a model, and their count alone does not predict coding performance.

xAI says it used a longer reinforcement-learning schedule and a harder set of multi-hour training problems. The company also incorporated internal SpaceX engineering records intended to strengthen reasoning about hardware and physical systems.

Grok 4.7 costs $2 per million input tokens and $6 per million output tokens. A run consuming one million input tokens and 100,000 output tokens would cost $2.60 at those rates. Devin charges the provider’s published rates under usage-based billing.

Mixed results beyond Devin

Results from broader leaderboards suggest that Grok 4.7’s performance depends heavily on the task format. Each benchmark uses a different dataset and scoring system, so their raw numbers are not directly comparable.

The combined results favor Grok 4.7 for structured backend work and long-running investigations. Its weaker Terminal-Bench score and FrontierCode scope penalties indicate less consistent performance on command-line tasks and narrowly constrained changes.

Choose it for deep backend work

Cognition’s findings support using Grok 4.7 for tasks with the following characteristics:

• Large Java, Go, or Ruby backends with behavior spread across several modules.
• Problems that require inspecting framework or dependency source code.
• Changes that need integration tests against working services or components.
• Long-running investigations where reviewers can accommodate a larger diff.

Grok 4.6 may remain a better fit for small, tightly specified patches or parallel agent runs where extra file changes increase token use and review effort. Teams adopting Grok 4.7 should define explicit scope constraints and inspect the final diff for unrelated edits.

Developers can access the model by downloading Devin Desktop or installing the Devin CLI, then selecting Grok 4.7 from the model picker.