Gemini 4 Hydrogen lands in Devin, scoring 38.2% on FrontierCode 2.0

Cognition, the company behind the Devin coding agent, has added Google's Gemini 4 Hydrogen to Devin and reports that it scores 38.2% on FrontierCode 2.0, just behind GPT-6.1 Sol (announcement). The number looks low until you know the test. FrontierCode is Cognition's own benchmark, and instead of asking whether code passes a specification, it asks whether a project's maintainer would actually merge the pull request. According to the benchmark page, each task was built by the open-source maintainers of the repository it comes from, grading covers correctness, tests, scope discipline and style, and a solution that fails any blocking criterion scores zero. Version 2.0, released in late October, has 132 tasks and deliberately vague prompts that leave engineering decisions to the model. At its launch the top model scored 45%, and Claude Fable 5.2 has since taken first place at 49.1.

That context changes what 38.2% says about Hydrogen. Google does not position Hydrogen as its strongest model. Its launch post describes a fast, cheaper model for everyday coding, priced at $1.50 per million input tokens and $7.50 per million output tokens, meant to work on its own or as the execution engine beside Google's frontier models, Gemini 4 Carbon and Argon. Landing just behind GPT-6.1 Sol on a strict merge test, as Cognition describes it, would be a reasonable result for that role. Cognition's post does not give GPT-6.1 Sol's score, the reasoning level Hydrogen ran at, or the cost per task, and its linked material was not read here, so the size of the gap and the price trade-off cannot be judged from this announcement. Fable 5.2's 49.1 was also posted without a reasoning level, so the roughly 11-point gap to the leader compares headline figures, not matched settings. The score is also Cognition's measurement on its own benchmark rather than an independent one.