GPT-6.1 Sol in Devin Scores 60.4% on FrontierCode at $0.31 per Task

OpenAI has released GPT-6.1 Sol, and Devin added it on release day: the model is available in Devin Desktop and Devin CLI. On FrontierCode 1.1, the Windsurf team's own benchmark, it scores 60.4%. GPT-6 Sol gets 60.7% and GPT-5.6 Sol gets 60.6%. The quality gap is within statistical noise, but the price per task has dropped sharply: at medium reasoning effort, GPT-6.1 Sol costs $0.31, while GPT-6 Sol spends $1.66 at maximum effort to reach its best result. That is a saving of 81%.

Same result, a different bill

FrontierCode 1.1 evaluates models on real engineering tasks using two criteria: solution quality and mergeability, meaning whether the changes can be accepted into a codebase without being rewritten. This is closer to a developer's day-to-day work than classic pass/fail tests that only ask whether a task was solved.

According to Windsurf, the picture across all modes looks like this:

• GPT-6.1 Sol, overall score: 60.4%, versus 60.7% for GPT-6 Sol and 60.6% for GPT-5.6 Sol.
• Medium effort level: $0.31 per task.
• GPT-6 Sol at maximum: $1.66 per task for its best result.
• Savings at every effort level: 44% to 57% relative to GPT-6 Sol, with a score within one point of it or higher.
• Low effort level: 58.1% for $0.21, versus 50.5% for GPT-6 Sol at the same setting.

The real shift is at the bottom of the scale

The new model did not raise the quality ceiling, and the numbers show it. The lower part of the curve is more interesting. At low effort, GPT-6.1 Sol gains 7.6 points over GPT-6 Sol, going from 50.5% to 58.1%. The price stays at $0.21 per task. On the FrontierCode leaderboard, this is the best result among all models costing less than $0.30 per task.

This matters for agentic workflows. When an agent runs hundreds of times a day on small tasks such as fixing tests, refactoring, or updating dependencies, the difference between $0.21 and $1.66 becomes a noticeable line in the budget. For that price, you can now get a result that previously required the maximum setting.

Where it sits on the overall map

GPT-6.1 Sol is not the absolute leader. Ahead of it on the leaderboard are Claude Opus 5.5 (65.3%), Claude Fable 5 (64.9%), GPT-6 Astra (64.5%), Claude Sonnet 5.5 (64.4%), Claude Fable 5.1 (63.6%), SWE-2 (62.5%), and Grok 4.6 (61.3%). Grok 4.7 ranks below it with 59.4%. The model is 4.9 points short of the top spot. But on the price-versus-quality chart, it sits in the lower-left corner: prices on the axis run from $0.25 to $5 per rollout, and in the cheap zone GPT-6.1 Sol has almost no competitors.

The caveat is obvious: FrontierCode is Windsurf's own proprietary benchmark, and the figures were published on the blog of a company that sells access to models. The post includes no independent verification. Still, a methodology that scores how mergeable the code is looks more sensible than many public tests.

How to use it in practice

Teams already working in Devin should run their typical tasks on GPT-6.1 Sol at medium effort and compare the results with their current configuration. By Windsurf's numbers, the quality loss is minimal while the cost drops several times over. For high-volume small tasks, the low setting is worth testing: 58.1% at $0.21 covers a significant share of routine work. Maximum effort levels and top models like Claude Opus 5.5 are better reserved for complex tasks, where the last points of quality matter and their price is justified. Before switching your entire workload, it helps to assemble your own sample of real tickets: a result on a closed benchmark does not guarantee the same behavior on your codebase.