Devin adds Gemini 4 Hydrogen: 38.2% on its coding test at $2.40 a task

Cognition has added Gemini 4 Hydrogen to Devin Desktop and Devin CLI. According to the company’s announcement, the model scores 38.2% on its FrontierCode 2.0 benchmark at $2.40 per task, about 40% less than Gemini 4 Carbon.

A near-top score at a lower cost per task

FrontierCode 2.0 is Cognition’s benchmark for real-world engineering work. It rates solutions on quality and mergeability, and, according to the company, a solution that fails any blocking criterion gets a zero. The final score is a weighted aggregate of rubric items.

On that test Gemini 4 Hydrogen reaches 38.2%. Cognition puts that within a point of GPT-6.1 Sol at 39.0% and close to Gemini 4 Carbon at 40.0%. Cost is measured as mean USD spend per rollout at list API pricing, public discounts included. Hydrogen comes in at $2.40 per task against $4.04 for Carbon, a gap of roughly 40%.

Cognition frames cost as the main news: “The bigger change is cost.” It says the result brings Gemini closer to the cost efficiency of GPT-6.1 Sol, though the announcement does not print Sol’s per-task price.

The company adds a second claim about reasoning effort. At medium effort, Hydrogen already beats the best result Gemini 4 Argon achieved, at less than a seventh of Argon’s cost per task. The page does not say what that medium-effort score is.

Who can use it today, and how

The model is live in Devin Desktop and Devin CLI. Cognition links a download for the desktop app and install instructions for the CLI. The post does not say which Devin plans include Hydrogen, whether it carries a usage multiplier, or whether it is the default model. Those details are not established by the announcement.

The benchmark is Cognition’s own, and the scores come from its own runs. Neither the 38.2% figure nor the cost numbers have been independently reproduced, and the post does not detail how many rollouts were run per task. The company also publishes a public FrontierCode leaderboard where the score-versus-cost chart can be inspected.

Why price may matter more than the score gap

The pattern follows the previous day’s release. Cognition reported that Gemini 4 Carbon hit 40.0% on its coding test, up from 26.3% for Gemini 4 Argon. Hydrogen gives up less than two points of that score for a cheaper run, so a team choosing a model inside Devin now has a trade-off to weigh rather than a single best option.

Outside Devin, the pricing gap between the Gemini 4 tiers is large. According to our earlier report on Hydrogen’s launch, Google lists it at $1.50 per million input tokens and $7.50 per million output tokens, with the price flat across the full 1 million-token context window. Artificial Analysis lists Argon at $4 and $20, so Hydrogen costs 37.5% as much on both sides. That is consistent in direction with Cognition’s finding that Hydrogen at medium effort undercuts Argon by a wide margin, though per-task cost also depends on how many tokens an agent burns.

Quality is not simply traded away for price either. In the same report, Artificial Analysis gave Hydrogen an Intelligence Index score of 51 at its default setting against 50 for Argon. That is a different test from FrontierCode, and Google’s own higher-effort figure of 54 is self-reported, but both point to a cheaper model that is not weaker than the one it undercuts.

For agentic coding, where a single task can involve many model calls, a per-task price matters more than a per-token one. A drop from $4.04 to $2.40 adds up across a team running dozens of tasks a day. Whether the saving holds depends on a team’s own codebase, since FrontierCode measures a fixed set of tasks that may not resemble its work.

What the announcement leaves open

Cognition does not publish Sol’s cost per task in the post, so the claim that Hydrogen narrows the efficiency gap to OpenAI’s model cannot be checked from the text alone. The post also does not cover latency, context limits inside Devin, or how Hydrogen performs on tasks outside FrontierCode. Readers who need those answers will have to look at the leaderboard or try the model in Devin themselves.