PR review bake-off · redesign-plugin-page · Nov 6

Which model should review your PRs?

“Review the shit out of this PR or branch. Is the PR implemented fundamentally the right way??”

Four Codex runs of the same prompt on the same branch, all on the fast (priority) tier. Each claim was checked against the working tree.

★ Use this for PR review

2× GPT-6.1 Terra · high, in parallel

Reviews 1 + 3Codexfast mode

✓Half the cost: $0.18 for both runs, against $0.37 for one Sol high run.
✓About 5× faster: both finish in 17s, against 1m 33s.
✓More coverage: together they found both main bugs, plus the two issues no Sol run caught (refresh/pin mismatch and lossy profile check).

2× Terra high

$0.18

17s · 418K tokens

1× Sol high

$0.37

1m 33s · 528K tokens

Trade-off: Terra reviews are static. They don't reproduce bugs or run checks, and Review 3 made one wrong claim, so treat their combined output as a list to triage.

Use Sol high when you need proof. It was the only run that reproduced the bugs and ran pnpm tc:web, the tests and the fixtures. Every claim it made was correct.

All four reviews, ranked by review quality

#	Model	Accurate?	Unique finding	Evidence	Tokens	Cost	Wall time	Tools
1	G GPT-6.1 Sol · Review 4 highfastmost rigorous	All correct	None beyond Review 2	Verified Reproduced both bugs, `pnpm tc:web`, tests, fixtures in 4 layouts	528,12490.3% cached 51K fresh473K cached3.5K out	$0.369	1m 33s	9
2	G GPT-6.1 Sol · Review 2 mediumfast	All correct	None	Read tests and fixtures; no reproduction	402,27688.1% cached 47K fresh353K cached1.9K out	$0.299	45s	8
3	G GPT-6.1 Terra · Review 1 highfast★ pick	All correct	Refresh/pin mismatch	Static only	210,55383.1% cached 35K fresh174K cached1.4K out	$0.103	17s	5
4	G GPT-6.1 Terra · Review 3 medium*fast★ pick	2 of 3 correct	Lossy profile check	Static only	207,24687.9% cached 25K fresh181K cached1.2K out	$0.080	16s	6

Fresh input Cached input Output Bars are relative to the largest run.

Earlier runs on other branches Nov 3 and Oct 31 · usage only, not quality-ranked

Started	Model	Tokens	Cost	Wall time	Tools
Nov 3, 00:26:52 01a1a084-ad1a	GPT-6.1 Sol · medium	1,104,576	≈$0.26†	3m 06s	18
Nov 3, 00:26:39 01a1a084-7c7d	GPT-6.1 Sol · medium incomplete	—	—	—	0
Nov 3, 00:12:50 01a1a077-d468	GPT-6.1 Sol · xhigh	2,646,428	≈$0.71†	6m 04s	23
Oct 31, 17:27:01 7828a1ad-e099	Claude Opus 5.5 (+ Haiku 5.5)	1,157,704	$1.0412	2m 40s	28

Ranking, accuracy, unique findings and evidence come from your cross-check of the four reviews. Tokens, wall time, tool calls, model and tier come from each Codex session log: Review 1 01a1b258-ee22, 2 01a1b25b-ef44, 3 01a1b255-2656, 4 01a1b25b-b817.
*You listed Review 3 as Terra high. Its session log records reasoning effort medium.
Costs are estimates: each session's logged tokens × OpenAI Fast-mode rates, which are 2× Standard. Sol: $4 input, $0.20 cached, $20 output per 1M tokens. Terra: $2 input, $0.10 cached, $10 output. All four runs logged service_tier: priority, and no request went over the 272K long-context limit. Codex logs don't record dollar cost.
The 2× Terra time assumes both run at once, so it is the slower of the two (17s).
†The Nov 3 logs don't record a service tier, so those costs use Standard rates. If they ran in Fast mode, double them.
Each setup ran once. Treat the ranking as a single data point, not a benchmark.
“Fresh” input is input minus cached input. Tool calls count top-level calls.