Write your own code review checks in plain English, then run each one on the model you choose. Macroscope supports open source models like GLM, Kimi, DeepSeek, Qwen, and MiniMax alongside frontier models like Claude and GPT, so you can tune every check for cost and quality.
Usage-based pricing. Pay only for what you use. No seat fees.
MacroscopeBench results
On MacroscopeBench, the best open-weight model for code review is GLM 5.3 from Z.ai, with a Score of 77.4 at max effort, followed by Kimi K3 (72.9) and DeepSeek V4.1 Flash (72.0). The top frontier configuration, Claude Opus 5.5 at max effort, scores 83.9, so frontier models still lead, but only 9 of the 52 configurations on the board score higher than GLM 5.3. All of them run as Macroscope Check Run Agents: set the model in a check's front matter and it runs on every pull request.
| Model | Model ID | Effort | Score | Recall | Precision | Cost per task |
|---|---|---|---|---|---|---|
| GLM 5.3 | glm-5-3 | max | 77.4 | 76.4% | 78.4% | $3.86 |
| Kimi K3 | kimi-k3 | max | 72.9 | 66.7% | 80.3% | $2.82 |
| DeepSeek V4.1 Flash | deepseek-v4-1-flash | max | 72.0 | 83.1% | 63.5% | $0.75 |
| Claude Opus 5.5Top overall | claude-opus-5-5 | max | 83.9 | 83.8% | 84.0% | $11.58 |
Score is the F1 of recall and precision. The board runs every configuration through the same review harness on 195 review tasks with 144 catalogued bugs, three trials each. Cost is the billed review cost per task. Full results for all 52 configurations are on the MacroscopeBench board, and the method is in the MacroscopeBench write-up.
Connect GitHub. Install the Macroscope GitHub app on the repositories you want checked.
Write the check. Add a markdown file to .macroscope/check-run-agents/ and set model: glm-5-3 in its front matter.
Commit it. On your default branch, it runs as a GitHub check run on the next pull request. Billed at $0.01 per Agent credit (raw LLM cost plus 5%), with 1,000 free credits a month.
Check Run Agents now support open source models. Each has its own tradeoffs around cost, latency, context window, and performance.
glm-5-3Z.ai GLM 5.3. Thinking is always on, with low, high, and xhigh levels.
kimi-k3Moonshot AI Kimi K3. Thinking is always on, with low, high, and xhigh levels.
deepseek-v4-1-flashDeepSeek V4.1 Flash. A low-cost option for checks that run on every PR.
glm-5-3-flashThe lowest per-token price of the open-weight models on Macroscope.
minimax-m3Thinking is on by default and can be turned off.
qwen-3-8-maxA graded thinking ladder from off to xhigh, with every level distinct.
Frontier models, alongside
Top OSS models match frontier models on many coding and reasoning tasks. Route the checks that do not need a frontier model to a cheaper one, and keep the rest on Claude or GPT.
Teams run suites of custom agents on every PR. At that volume, model choice drives the bill. Pick the cheapest model that does the job for each check.
Pay only for what you use. No per-seat fees, no surprises. Start with $100 in free usage.
Set the model in each agent's markdown frontmatter. Tune cost, latency, and performance check by check, not account-wide.
Every agent runs as a native check run on your pull requests. Define it in plain English markdown and commit it to your repo.
Open source models sit alongside Claude and GPT. Keep frontier models on security audits and the checks that demand them.
Define checks in plain English markdown. Set the model per agent. Macroscope runs them as real GitHub check runs.
Usage-based pricing. Pay only for what you use. No seat fees.
Change one line to run a check on an open source model. See the models and pricing docs for the full list.
On MacroscopeBench, the best open-weight model for code review is GLM 5.3 from Z.ai, with a Score of 77.4 at max effort, followed by Kimi K3 (72.9) and DeepSeek V4.1 Flash (72.0). The top frontier configuration, Claude Opus 5.5 at max effort, scores 83.9, so frontier models still lead, but only 9 of the 52 configurations on the board score higher than GLM 5.3. All of them run as Macroscope Check Run Agents: set the model in a check's front matter and it runs on every pull request.
Join teams building with Macroscope to catch more bugs and merge PRs faster than ever. Simple usage-based pricing. No seat fees, no surprises.