GPT-6.1 Sol had the highest measured recall in our benchmark, at about a third of GPT-5.6 Sol's cost
In the 2026-10-17 run, GPT-6.1 Sol had the highest measured recall of 17 models, with few false alarms, at about a third of GPT-5.6 Sol's cost per review.
In CodePanel's benchmark run of 2026-10-17, GPT-6.1 Sol had the highest measured recall of the 17 ranked models: 90% of known bugs from open-source history. It shares first place with six others, because no gap among those seven is measurable. A GPT-6.1 Sol review cost $0.041, against $0.1151 for GPT-5.6 Sol, the previous Sol model in the run.
The numbers
Each model reviewed the same fixtures as CodePanel's Bugs & Security reviewer, on its own. Every model in the table ran at medium reasoning effort. The methodology page covers the fixtures, the three judges and the intervals.
| GPT-6.1 Sol | GPT-5.6 Sol | |
|---|---|---|
| Rank | 1 | 1 |
| Recall [95% CI] | 90% [81%–98%] | 71% [56%–85%] |
| False alarms per clean PR | 0.7 [0.2–1.2] | 1.0 [0.5–1.5] |
| Catches on the defect's lines | 95% | 97% |
| Cost per review | $0.041 | $0.1151 |
| Mean time per review | 26s | 54s |
GPT-6.1 Sol and Claude Opus 5.5 are the only two models measurably above every model outside first place. GPT-6.1 Sol's 19-point lead over GPT-5.6 Sol is not measurable once the test corrects for the number of pairs compared, so both stay at rank 1.
GPT-6.1 Sol is also one of the three judges. A judge's vote is left out when it scores a model from its own maker, and for GPT-6.1 Sol that changed nothing: its recall is 90% with or without the OpenAI vote. The same rule cost Claude Opus 5.5 four points; with every judge counted, Opus would be a fraction of a point ahead, and neither gap is measurable.
Price and cost per PR
At list price on 2026-10-16, GPT-6.1 Sol is $2.00 / $10.00 per million input / output tokens, the same as GPT-6 Sol ($2.00 / $10.00), which it followed a week later, on 2026-09-29. GPT-6 Sol was not in the run. GPT-5.6 Sol lists at $4.00 / $20.00, 2.0× as much per token.
The measured gap per review is wider than the list-price gap: 2.8× rather than 2.0×. Both read about the same input, 15,336 and 15,447 tokens a review on average, but GPT-5.6 Sol wrote 2.6× as many output tokens (2,664 against 1,036), and both models charge five times as much for output as for input. GPT-5.6 Sol also took twice as long per review.
$0.041 is one reviewer seat on one round. A typical call of 30,000 input and 3,000 output tokens is $0.09 at list price. A CodePanel review of a pull request runs several reviewer seats and an orchestrator, and each reviewed push is a new round; what AI code review costs works through a whole pull request.
Where it fits
GPT-6.1 Sol is the model to pick when one reviewer has to catch as much as possible on its own, and of the seven in first place only GPT-6 Luna raised fewer false alarms. In a panel, where the orchestrator checks each finding against the code, the case is weaker: GPT-6 Luna and Claude Haiku 5.5 tie it for first, and a GPT-6.1 Sol review cost 14× what a GPT-6 Luna review did. Best LLM for code review sets out the choice per seat and the benchmark leaderboard has the live ranking. These figures are from the 2026-10-17 run and will not change with later ones.
To try it as a reviewer, add an OpenAI key in CodePanel's settings (providers and keys) and ask your coding agent to add GPT-6.1 Sol to your panel through the MCP server.