---
type: "post"
id: "gemini-4-carbon-is-live-in-devin-desktop-f5900a"
slug: "gemini-4-carbon-is-live-in-devin-desktop-f5900a"
title: "Gemini 4 Carbon is live in Devin Desktop and CLI, scoring 40.0% on FrontierCode 2.0."
description: "Cognition says Gemini 4 Carbon is now available in Devin Desktop and Devin CLI, and reports a 40.0% score on its FrontierCode 2.0 engineering benchmark."
retrieval_nugget: "Cognition says Gemini 4 Carbon is now available in Devin Desktop and Devin CLI, and reports a 40.0% score on its FrontierCode 2.0 engineering benchmark. The company also says Carbon improves on Gemini 4 Argon while costing less per task at its top score, making this a concrete model option for Devin users to evaluate."
published_at: "2026-11-10"
updated_at: "2026-11-10"
record_date: "2026-11-10"
date_kind: "published_at"
topics: ["gemini"]
entities: ["Cognition","Devin","Gemini 4 Carbon","Gemini 4 Argon","GPT-6.1 Sol","FrontierCode 2.0"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "Gemini 4 Carbon is now available in Devin Desktop and CLI, scoring 40.0% on FrontierCode 2.0 and beating Argon and GPT-6.1 Sol at lower cost."
prior_refs: ["https://newruntime.com/posts/gemini-4-carbon-is-now-available-for-developers-680123/"]
source_url: "https://devin.ai/blog/gemini-4-carbon"
source_urls: ["https://devin.ai/blog/gemini-4-carbon"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:gemini-4-carbon-is-live-in-devin-desktop-f5900a"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/gemini-4-carbon-is-live-in-devin-desktop-f5900a/","markdown":"https://newruntime.com/posts/gemini-4-carbon-is-live-in-devin-desktop-f5900a.md","json":"https://newruntime.com/posts/gemini-4-carbon-is-live-in-devin-desktop-f5900a.json"}
markdown_url: "https://newruntime.com/posts/gemini-4-carbon-is-live-in-devin-desktop-f5900a.md"
---

# Gemini 4 Carbon is live in Devin Desktop and CLI, scoring 40.0% on FrontierCode 2.0.

## Retrieval answer

Cognition says Gemini 4 Carbon is now available in Devin Desktop and Devin CLI, and reports a 40.0% score on its FrontierCode 2.0 engineering benchmark. The company also says Carbon improves on Gemini 4 Argon while costing less per task at its top score, making this a concrete model option for Devin users to evaluate.

## Summary
Cognition says Gemini 4 Carbon is now available in Devin Desktop and Devin CLI, and reports a 40.0% score on its FrontierCode 2.0 engineering benchmark. The company also says Carbon improves on Gemini 4 Argon while costing less per task at its top score, making this a concrete model option for Devin users to evaluate.

## What happened
Cognition announced on November 10, 2026 that Gemini 4 Carbon is live in Devin Desktop and Devin CLI. The post presents the release as a model availability update for developers using Devin, rather than as a separate general-purpose API launch. It does not state a separate Carbon price or describe changes to Devin subscriptions. Cognition provides links to download Devin Desktop and install Devin CLI, positioning those products as the stated access surfaces.

The announcement reports a 40.0% score for Carbon on FrontierCode 2.0, which Cognition describes as a benchmark evaluating models on real-world engineering tasks based on quality and mergeability. It compares that result with 26.3% for Gemini 4 Argon and 39.0% for GPT-6.1 Sol. Cognition says the score is a weighted aggregate of rubric items and that solutions failing blocking criteria receive zero. That makes the published percentage a composite result rather than a count of tasks completed.

Cognition also makes a cost comparison. It says that at low reasoning effort, Carbon already surpasses Argon's best result, and that Carbon's top score costs less than two-thirds as much per task as Argon's. The page defines benchmark cost as mean US-dollar spend per rollout at list API pricing, including public discounts. The comparison therefore concerns benchmark rollouts and published API-price assumptions, rather than a quoted Devin subscription cost. It does not give the exact dollar amounts in the announcement text.

The post links to a FrontierCode leaderboard and characterizes the tasks as real-world engineering work. It does not identify particular repositories, task categories, or the number of evaluated rollouts in the announcement. Nor does it state whether users must select Carbon manually, whether it is a default model, or how reasoning effort is exposed in the product interface.

Cognition does not describe a new configuration step, eligibility restriction, rollout schedule, or separate access tier for Carbon. The stated availability is limited to those two Devin surfaces; the announcement does not say that Carbon is available through Google's Gemini API or another provider's gateway. For current Devin users, the immediate change is the presence of Carbon in the desktop and command-line product environment.

## What the archive adds
New Runtime's earlier coverage on [Gemini 4 Carbon for developers](https://newruntime.com/posts/gemini-4-carbon-is-now-available-for-developers-680123/) said Google had made the model available to developers, highlighting cybersecurity performance and long-horizon capabilities. This is a new fact on a known story: Cognition is now specifying availability inside Devin Desktop and CLI and publishing a comparative FrontierCode 2.0 result. Earlier coverage of [Gemini 4 Argon on AI Gateway](https://newruntime.com/posts/gemini-4-argon-the-first-gemini-4-model-e3012c/) provides context for the model comparison, but this announcement concerns a separate product surface.

## Trend signal
This supports the pattern that [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/): a coding-agent environment can make a model choice operational by exposing it in the existing desktop and CLI workflows. The announcement is evidence of availability and benchmark claims, not proof that users can switch models freely or that the same result holds across harnesses.

## Why now
Cognition's announcement ties the timing to Carbon becoming live in Devin's desktop and CLI products. It supplies no further explanation for the release timing.

## Why it matters
For a team already using Devin, Carbon is a model option that can be compared against Argon on a coding workload without changing the surrounding product surface. Cognition's claimed score gap is substantial, and its cost comparison suggests that the strongest reported result may also require less spend per task than Argon's strongest result. That combination is more useful to procurement and engineering leads than a score alone.

But a benchmark score is not a deployment decision. Teams should test representative tasks, include review and rework in their cost accounting, and verify mergeability in their own repositories. The benchmark's zero for submissions that fail blocking criteria also makes pass rates and task mix important: a weighted aggregate can hide where a model succeeds or fails. Carbon's reported result is a reason to run a controlled evaluation, not to assume equivalent gains on a team's codebase.

## Limits and uncertainty
The figures and comparisons come from Cognition, the publisher of Devin and the FrontierCode benchmark. The announcement does not provide the exact per-task dollar totals, sample size, confidence intervals, task distribution, or independent replication. Its description says the score is weighted and that blocking failures receive zero, but does not provide enough detail here to reconstruct how each task and rubric item contributes. The availability statement covers Devin Desktop and CLI; broader API access, pricing, and rollout conditions are not specified.

## Skill to build now
Build a small evaluation set from recent coding tasks your team has already reviewed. Run a handful through Carbon and your current Devin model at comparable reasoning settings, then record accepted changes, blocking failures, human review time, and spend per completed task. That will test the specific quality-and-cost claim against your repositories rather than treating a vendor leaderboard as a forecast.

## Reader takeaway
Gemini 4 Carbon is now available inside Devin Desktop and CLI, with Cognition reporting a higher FrontierCode 2.0 score and lower cost at its top result than Gemini 4 Argon. Treat that as a prompt to evaluate Carbon on your own engineering work, not as a guarantee of better outcomes across every codebase.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Gemini](https://newruntime.com/topics/gemini.md): Explore the gemini topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)