We ran Corti S1 and OpenAI’s GPT-6.1 Sol on the same six real bugs, one attempt each, with each project’s own tests as the judge. Both fixed all six. S1 won five of the six rounds, finished the set in 10:16 against 18:45 (1.8× faster), and cost $0.68 against $1.13 at list prices.
Typing was never the slow part. Writing code took 1:13 of S1’s 10:16 and 2:02 of GPT-6.1 Sol’s 18:45. The difference is reading and planning: 7:12 of the 8:29 gap, or 85%, is time GPT-6.1 Sol spent there.
How we ran it
S1 is the model family in Corti Models, which runs open-weight models on sovereign infrastructure through an OpenAI-compatible API (see how Corti Models works).
Each bug is a real issue from an open source project, taken from SWE-bench Verified: Matplotlib, Pylint, scikit-learn, Django (twice), and Flask. Both models worked as coding agents in the OpenCode harness, the program that gives a model tools to read files, run commands, and edit code. Effort was set to max for both, and each worked on a prepared checkout of the project. Every number here comes from their first attempt at each task, run on November 11, 2026.
The project’s own tests decide what counts, and the first fix that passes them wins the round. All 12 attempts passed, so no round turned on a failure. The clock starts when a model picks up the task and stops at its final message, so it includes everything in between: reading, thinking, running commands, and waiting for them to finish.
Each race below replays the real run, sped up so a round takes about 15 seconds. The bar under each clock fills as time passes. We sorted every step by what it did: reading and planning, editing (green), testing, or finding the right Python install (hatched).
Round 1, Matplotlib
Make histograms keep their set range in density mode.
View the issue and fix on GitHub.
Watch the green on each bar, which marks the moments a model edits code. S1 makes its first change at 0:17 and finishes in 0:49. GPT-6.1 Sol doesn’t edit anything until 1:24, and it writes 5,137 output tokens where S1 writes 1,095.
Winner: Corti S1
Round 2, Pylint
Stop the code checker removing the wrong import path.
View the issue and fix on GitHub.
This is the widest gap of the six, 0:53 against 4:15. S1’s first edit comes at 0:30. GPT-6.1 Sol’s doesn’t come until 2:43. GPT-6.1 Sol also keeps a to-do list, updating it four times in this round. S1 doesn’t use one in any round.
The hatched stretches in the bars are both models hunting for the right Python install, up to 17 seconds in a round. That comes from the test setup, not the models, so the bars mark it separately.
Winner: Corti S1
Round 3, scikit-learn
Make a clustering model's two ways of predicting agree.
View the issue and fix on GitHub.
S1 reaches its first edit 40 seconds ahead and still finishes 15 seconds behind. It also takes more steps here than in any other round, 31 against 14, and about 14 seconds of its run go on a changelog entry. The tests don’t read changelogs.
Winner: GPT-6.1 Sol
Round 4, Django
Fix a crash showing 'time since' with time zones on.
View the issue and fix on GitHub.
Winner: Corti S1
Round 5, Flask
Refuse a blueprint that has an empty name.
View the issue and fix on GitHub.
Flask is the only round where both models are editing inside a minute, at 0:34 and 0:47. S1 still finishes 1:15 sooner.
Winner: Corti S1
Round 6, Django
Reject the wrong kind of value in a decimal field cleanly.
View the issue and fix on GitHub.
This is the closest of S1’s wins, 2:53 against 3:36. S1 spends more time testing than GPT-6.1 Sol does here, 1:12 against 0:51, and that is its longest stretch of testing in any round.
Winner: Corti S1
All six rounds
| Round | Bug | Corti S1 | GPT-6.1 Sol |
|---|---|---|---|
| 1 | Matplotlib: Make histograms keep their set range in density mode | 0:49 | 2:23 |
| 2 | Pylint: Stop the code checker removing the wrong import path | 0:53 | 4:15 |
| 3 | scikit-learn: Make a clustering model's two ways of predicting agree | 2:42 | 2:27 |
| 4 | Django: Fix a crash showing 'time since' with time zones on | 1:30 | 3:20 |
| 5 | Flask: Refuse a blueprint that has an empty name | 1:27 | 2:42 |
| 6 | Django: Reject the wrong kind of value in a decimal field cleanly | 2:53 | 3:36 |
| Total | 10:16 | 18:45 |
S1 took more steps overall, 129 against 110, and still finished first. On average a step took it about 5 seconds, against about 10 for GPT-6.1 Sol. It also wrote about half the output tokens, 18,621 against 37,236. Cost is each attempt’s tokens at list prices.
What this doesn’t show
Six tasks with one attempt each is a small sample, and one attempt is one draw from a model’s range. Run it again and the times will move. It also tests one kind of work, fixing a bug in a project the model can read and test, with one harness at one effort setting.
The split between reading, editing, and testing is our own sorting of each step by what it did, so read the 85% as approximate.
To try S1 on your own bugs, start in the Corti Console with free credits and keep the coding agent you already use.






