openai will sell you gpt-5.6 at $200/mo. anthropic will sell you opus 5 at $200/mo. 0 of the 10 most common ai complaints are actually model problems, and the $20 plan already handles the fix neither of them will link to 24 opus reviewers on every pull request, one opus writing the final call. raw pipeline ran 28,400 tokens per review at $2.14 and 42 seconds. drop a python filter between the reviewers and the approver and the same task lands at 3,900 tokens, $0.31, 9 seconds deepmind put a paper on it in march (liu et al. 2024). model precision collapses when the answer sits past the 60% mark of a long input. 24 unstructured comments in one call puts most of them exactly there the filter, four passes in code, zero model calls: reject empty reviews -> 4 of 24 reviewers return with no severity, no file path, or no line number -> approver sees a clean queue instead of guessing what to skip merge on file plus line -> 9 of 24 reviewers were flagging the same violation from different rules -> keeps the highest-severity report, notes the rule count next to it split the disagreements -> two reviewers arguing "block" vs "approve" on the same line becomes a routed decision, not a coin flip inside the approver -> across 80 pull requests, this caught 34 conflicting calls the concatenated dump silently averaged out rank by blast radius, ship -> the approver now reads 4k of prioritized findings instead of 28k of interleaved noise result on 80 pull requests: $2.14 -> $0.31 per review (84% drop) 42s -> 9s latency (79% drop) 11/80 required a human reviewer -> 2/80 after 34 conflicting calls surfaced that concatenation buried anthropic bills you for what opus reads. cutting what opus reads costs $0 in tokens. they will not send you an email about that drop your $200/mo ai sub to $20, full filter walkthrough in the article below