Cognition reports 60.5% on FrontierCode 1.1 in Devin. Chart version labels differ, and reasoning effort and the full harness are undisclosed. This is a platform-reported result.
Tasks suitable for judgment: Performance on real-world software-engineering tasks, code quality, and mergeability in Cognition's custom FrontierCode test; availability and the then-current promotional terms in Devin Desktop and Devin CLI.
Tasks not suitable for extrapolation: General success rates across all codebases, results in other Agent clients, production costs, long-term stability, or cross-model rankings under undisclosed conditions.
Applicable model version: The page only says Gemini 4 Argon; it provides no API model ID, snapshot name, or build version.
Test environment or client: Devin Desktop and Devin CLI; the evaluation is Cognition's FrontierCode 1.1 (the chart is titled FrontierCode 1.1 Extended).
Reasoning effort and parameters: Not specified. The cost-chart legend says each point corresponds to a reasoning effort, but the page does not identify the effort used for 60.5%, nor disclose temperature, token limits, tool configuration, or repeated-run settings.
The page describes FrontierCode as Cognition's proprietary benchmark for evaluating code quality and mergeability on real-world software-engineering tasks. The chart explains:
The score is a weighted aggregate of rubric items.
Solutions that fail the blocking criteria receive a score of 0.
Cost ($) in the cost chart is the mean USD spend per rollout.
Each point in the cost chart represents a reasoning effort.
The body calls the benchmark FrontierCode 1.1, while the chart title calls it FrontierCode 1.1 Extended. The source does not explain whether the names refer to the same task set or different versions, so this article does not merge or infer them.
The article title and body say that Gemini 4 Argon is live in Devin Desktop and Devin CLI. The page does not provide the client version, underlying API request, tool list, sandbox configuration, timeout, concurrency, retry, or human-intervention rules, so the complete runtime harness cannot be reproduced from this article alone.
Cognition's observation from the evaluation is that Argon prepares thoroughly before writing code: it reads widely, checks git history for repository conventions, and reviews its own diff before finishing. The page says it is strongest on bug fixes and refactors in backend and systems code, especially Python, C/C++, and PHP; this is a qualitative observation from the page, not a separately measured subscore.
| Page item | Gemini 4 Argon result | Comparison or qualification given directly by the page |
|---|---|---|
| FrontierCode 1.1 (body wording) | 60.5% | Cognition says it exceeds GPT-6.1 Sol (60.4%) and says it matches GPT-6 Sol (60.7%) |
| FrontierCode 1.1 Extended (chart wording) | 60.5% | The chart also shows Gemini 4 Argon at 60.5%; its score is a weighted aggregate of rubric items, and a solution that fails the blocking criteria receives 0 |
The page also says this is the highest-scoring Gemini model on the leaderboard, four points ahead of Gemini 3.7 Flash. This is Cognition's page claim, not an independent verification, and it cannot be extrapolated to undisclosed tasks or configurations.
Gemini 4 Argon is live in Devin Desktop and Devin CLI; the page says it is 75% off list price through 2026-10-27.
This 75% discount consists of two parts: Google's 50% introductory pricing through 2026-12-01, plus an additional 50% discount in Devin through 2026-10-27.
The page does not give Gemini 4 Argon's actual dollar amount per rollout; the cost chart only describes mean USD spend per rollout and reasoning-effort points, with a dashed line for the current price.
The body directly states that Gemini 4 Argon scores 60.5% on FrontierCode 1.1; the comparison values are GPT-6.1 Sol at 60.4% and GPT-6 Sol at 60.7%.
Adjacent scores visible in the FrontierCode chart include GPT-6.1 Sol 60.4%, Gemini 4 Argon 60.5%, GPT-5.6 Sol 60.6%, and GPT-6 Sol 60.7%. These are chart readings and cannot be compared outside their benchmark version, task set, and harness.
The page defines the score as a weighted aggregate of rubric items and says that a solution failing the blocking criteria receives 0 points.
The page defines cost as mean USD spend per rollout and says each point represents a reasoning effort; the Gemini 4 Argon promotional points reflect the additional 50% discount in Devin through 2026-10-27.
The model behavior observations on the page are: reading widely first, checking repository conventions in git history, and reviewing the diff before finishing; the page says it is strongest on backend and systems-code bug fixes and refactors, especially Python, C/C++, and PHP.
Confirmed result: The Devin page directly shows Gemini 4 Argon's 60.5% score and says scoring uses a weighted rubric aggregate, with 0 for failure of the blocking criteria.
Environment boundary: The result comes from the Devin Desktop and Devin CLI product environments; it should not be treated as a result for the Gemini API, Google AI Studio, or another Agent harness.
Incomplete model configuration: The page provides no API model ID, snapshot, reasoning effort, sampling parameters, tool configuration, task count, run count, scoring implementation, or failed-run retry rules.
Source naming boundary: The body says FrontierCode 1.1, while the chart says FrontierCode 1.1 Extended; without more method details, they should not be interpreted as either the same or different version.
Promotion is not model capability: The 75% off offer, Google's 50% introductory pricing, and Devin's additional 50% discount are price terms through specified dates and should not be conflated with the 60.5% capability score.
Vendor-reported result: The page was published by Cognition/Devin and does not provide enough material for full independent reproduction; statements such as “highest-scoring Gemini model” and “four points ahead” should remain page claims.
To reproduce or run a comparable test, at minimum fix:
The exact Devin Desktop or Devin CLI version and the Gemini 4 Argon model build actually used.
The FrontierCode 1.1 and FrontierCode 1.1 Extended task sets, versions, input repositories, and acceptance criteria; resolve the body/chart naming difference first.
The rubric items, blocking criteria, weighting, pass/fail handling, timeout, retry, and human-intervention rules.
Reasoning effort, sampling parameters, tool and sandbox permissions, and the number of rollouts per task.
Input/output tokens, elapsed time, mean USD cost per rollout, and whether the promotion applied.
The page does not disclose all of this information. Therefore, only its public page readings and method descriptions can be reproduced; one cannot claim to have reproduced the full conditions behind 60.5%.
Cognition / Devin: https://devin.ai/blog/gemini-4-argon
Gemini 4 Argon
