Cognition Integrates xAI's Grok 4.7 Into Devin

Cognition has integrated xAI's new Grok 4.7 model into the Devin Desktop application and Devin CLI, while the model also launched on GitHub Copilot and Cursor. Built on a massive 2.1 trillion parameter base model—representing a 40 percent size increase over the 1.5 trillion parameters of Grok 4.6—the new model was trained using SpaceX engineering records and a longer reinforcement-learning schedule. Developers can access Grok 4.7 via the xAI API for $2 per million input tokens and $6 per million output tokens. Under this pricing, a query using one million input tokens and 100,000 output tokens costs $2.60.

In Cognition's internal evaluations using its proprietary FrontierCode 1.1 Extended benchmark, Grok 4.7 achieved a score of 59.4%. This benchmark, which assesses the mergeability of pull requests, was recently updated after Cognition audited over 1,000 blocker criteria and reclassified 75 of them. On other benchmarks, such as the Artificial Analysis Intelligence Index v4.3.2, Grok 4.7 improved by 111 points over Grok 4.6, placing it behind Claude Fable 5.1, which scored 55%, and GPT-6 Astra, which led with 60%.

Despite its raw power, Grok 4.7 scored slightly below Grok 4.6 on the FrontierCode benchmark due to a habit of expanding its task scope. While the model excelled at complex, multi-module backend tasks in Java, Go, and Ruby, it frequently made unrequested edits to adjacent files. For example, a simple bug fix might trigger unnecessary refactoring in neighboring modules, expanding the final diff and increasing regression risks.

For software practitioners, these characteristics mean Grok 4.7 is highly effective for deep, long-running investigations that require tracing dependencies and writing integration tests. However, for minor, tightly constrained patches, the older Grok 4.6 remains a more efficient choice. Teams deploying Grok 4.7 should establish strict scope boundaries and carefully review pull requests to avoid bloated codebases and excess token costs.