Warmstart went from hundreds/day on OpenAI & Anthropic to $299/mo with Lunaroute for current traffic. Open-weight models, no per-token fees. Try it: app.lunaroute.com/start?ref=… We each get $200 credit after your first eligible monthly payment. I know the founder; I’m a customer.

11/ How do you get AI code review that finds real bugs without wasting time and money? I ran 2,416 reviews: 8 models, prompts vs harnesses, effort levels, open vs closed weights. The result: a step-by-step setup that matched Opus 5 at 1/13 the cost: dsifry.github.io/harnesseval…

10/ Don't take my word for it. The page has an explorer: pit any model, harness and effort level against each other across all 66 configurations, by bugs, noise, dollars and seconds. dsifry.github.io/harnesseval… What are you running for AI code review, and has anyone measured it?

9/ Disclosure: I wrote github.com/dsifry/metareview, one of the two harnesses tested. It is MIT licensed and free. Every harness was judged by the same frozen instruments, Compound Engineering holds the single top score, and the data and code are public so anyone can re-run it.

8/ The caveats, because they matter: - six high-severity PRs, two codebases, one run per configuration - the GLM-5.3 runs used a 4-bit quantized, hosted build - score intervals overlap: read "comparable", not "better" Evidence for a pilot, not a verdict. Run your own evals.

7/ "Verified" means proven. 42 bugs were labelled by the benchmark's human reviewers. The models kept reporting more, so I audited them: 105 additional bugs, each with a test that fails on the PR and passes on the fix. 147 verified bugs. No vibes in the answer key.

6/ A harness is not free. Median cost vs the same model as a single prompt: about 10× the tokens, 5× the dollars, 3.5× the wall-clock, plus more unsupported findings to triage. That is exactly why the model you run it on matters so much.

5/ The picks: Best value: GLM-5.3 + metareview, low effort. 72 of 147 bugs, $0.22, 95 seconds. Top score: Opus 5 + Compound Engineering, medium. 88 of 147, $6.12. Tightest budget: GLM-5.3-Flash + metareview, low. 71 of 147, $0.024, more noise to triage.

4/ Step 4: vet the host before you send it your code. Open weights mean many hosts can serve the same model, on very different terms. Ask about jurisdiction, data retention, training on your code, exact weights and precision, and usage rights. Checklist: dsifry.github.io/harnesseval…

3/ Step 3: stop paying for "high" by default. High reasoning effort showed no measurable gain over medium in 17 of 22 head-to-heads, and it cost more in 20 of them. It measurably helped in 4 and hurt in 1. Default to low or medium and make high earn its place.