Grok 4.7 Now Available in Devin
WindsurfWindsurf added xAI's Grok 4.7 to Devin Desktop and Devin CLI, the same week the model also reached GitHub Copilot and Cursor. Built on a larger base model with a longer reinforcement-learning run, Grok 4.7 is strongest on hard, multi-module backend tasks in Java, Go, and Ruby, where it investigates deeply and verifies its work with real integration tests. It trails Grok 4.6 on Cognition's aggregate FrontierCode 1.1 score, however, because it tends to over-scope and produce larger diffs than requested. Pricing stays at $2 per million input tokens and $6 per million output tokens.
Key Takeaways
- Grok 4.7 is now live inside both Devin Desktop and Devin CLI, rolling out the same week it also reached GitHub Copilot and Cursor.
- The model runs on a larger, roughly 2.1-trillion-parameter base (about 40% bigger than Grok 4.6) trained with a longer reinforcement-learning run weighted toward multi-hour tasks.
- On Cognition's FrontierCode 1.1 benchmark, Grok 4.7 is strongest on hard, multi-module backend work in Java, Go, and Ruby, where it reads dependency source and verifies changes with real integration tests.
- Despite that strength, it scored 59.4% on FrontierCode 1.1 Extended, trailing Grok 4.6 on the aggregate because it tends to over-scope and touch code beyond what was requested.
- AlphaSignal noted the wider diffs can raise review time and regression risk, recommending Grok 4.6 instead for small, precisely scoped changes.
- Pricing holds at $2 per million input tokens and $6 per million output tokens, keeping Grok 4.7 competitively priced against other frontier models available in Devin.
Sources & Mentions
4 external resources covering this update
Grok 4.7 Joins Devin's Model Lineup
Windsurf's Devin Desktop and Devin CLI added support for Grok 4.7, xAI's newest coding-focused model, the same week it also became available inside GitHub Copilot and Cursor. Grok 4.7 replaces Grok 4.6 as xAI's most capable model for coding and knowledge work, built on a larger base model and trained with a longer reinforcement-learning run deliberately weighted toward harder, longer-running engineering tasks.
Built for Hard, Multi-Module Backend Work
According to Cognition, Grok 4.7's biggest strength inside Devin shows up on hard, multi-module backend tasks written in Java, Go, and Ruby. On those problems the model investigates dependencies deeply, reads through source code across modules, and verifies its own changes end to end with real integration tests rather than stopping once the code compiles. The underlying model reportedly grew to roughly 2.1 trillion parameters, about 40% larger than Grok 4.6, with training data that pulled in SpaceX engineering material to sharpen its reasoning about hardware-adjacent problems.
The Trade-off: Scope Creep
The gains come with a catch Cognition was upfront about: Grok 4.7 scored 59.4% on Cognition's FrontierCode 1.1 Extended benchmark, trailing Grok 4.6 on the aggregate score. The model tends to over-scope its changes, producing larger diffs that touch code beyond what a task actually asked for. AlphaSignal noted this wider blast radius can increase review time and regression risk, making Grok 4.7 a weaker fit for small, precisely scoped patches even as it outperforms on deep, investigative backend work.
Guidance and Pricing
Cognition's recommendation is straightforward: reach for Grok 4.7 on large, multi-module backends that need deep investigation, and stick with Grok 4.6 for narrow, well-defined changes. Pricing holds steady at $2 per million input tokens and $6 per million output tokens, keeping Grok 4.7 competitively priced next to the other frontier models already available inside Devin, including Claude Fable 5.1, GPT-6 Astra, and Cognition's own SWE-2.