Claude Sonnet 5.5 was the fastest reviewer in our benchmark, and tied for first
Claude Sonnet 5.5 tied for first in the 2026-10-17 run at 10 seconds a review, the fastest of 17 models, and cost more per review than its list price suggests.
In CodePanel's benchmark run of 2026-10-17, Claude Sonnet 5.5 tied for first place on known bugs from open-source history, catching 77% of them, and was the fastest of the 17 models ranked in that run, at 10 seconds a review. It also raised the most false alarms of the seven tied for first, though not clearly more than Claude Opus 5.5, and it is measurably above fewer of the other models than Opus is.
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.
| Claude Sonnet 5.5 | Claude Opus 5.5 | Claude Haiku 5.5 | |
|---|---|---|---|
| Rank | 1 | 1 | 1 |
| Recall [95% CI] | 77% [63%–89%] | 87% [76%–96%] | 60% [44%–75%] |
| False alarms per clean PR | 1.9 [1.4–2.4] | 1.7 [1.1–2.4] | 1.4 [0.9–1.9] |
| Catches on the defect's lines | 85% | 92% | 82% |
| Cost per review | $0.0588 | $0.1589 | $0.004 |
| Mean time per review | 10s | 33s | 16s |
No gap among the three Claude models is measurable, so all three share rank 1 with four models from other makers. The difference between them is in which of the other models each one beats. Opus 5.5 is measurably above every model outside first place; Sonnet 5.5 is measurably above 5 of those 10. The other five are Gemini 3.8 Flash and the two GLM-5.3 and two DeepSeek V4 models on Workers AI.
Sonnet 5.5 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 Sonnet 5.5 that lowered its recall from 81% to 77%.
Price and cost per PR
At list price on 2026-10-16, Sonnet 5.5 is $2.00 / $10.00 per million input / output tokens, the same as Claude Sonnet 5 ($2.00 / $10.00), which it followed on 2026-09-28. Sonnet 5 was not in the run. Opus 5.5 lists at $4.00 / $20.00, and in the run an Opus 5.5 review cost 2.7× a Sonnet 5.5 one, partly because Opus wrote 2.6× as many output tokens.
Sonnet 5.5 has the same list price as GPT-6.1 Sol ($2.00 / $10.00), yet a Sonnet 5.5 review cost 1.4× as much. On the same fixtures, Anthropic's API counted an average of 22,909 input tokens a review and OpenAI's 15,336 for GPT-6.1 Sol, about 1.5× as many. Two makers' per-token prices do not compare directly; the measured cost per review does.
$0.0588 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
Reviewers run in parallel, so the slowest seat sets the wait on every push, and in this run no model was quicker than Sonnet 5.5. Its false alarms matter less in a panel, where the orchestrator checks each finding against the code before the review is posted. As a lone reviewer, Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon had higher measured recall, though not measurably higher, and GPT-6.1 Sol raised fewer false alarms at a lower cost per review. Haiku 5.5 tied it for first, and a Sonnet 5.5 review cost 15× a Haiku 5.5 one. 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 Anthropic key in CodePanel's settings (providers and keys) and ask your coding agent to add Sonnet 5.5 to your panel through the MCP server.