Team splits AI coding work across four models and logs what each catches

A developer team stopped relying on a single model after its subscriptions kept hitting usage limits and the model was checking its own work. Claude now plans and reviews, Grok attacks Claude's drafts, Codex builds larger components, and a router called Tīrtha handles routine building. A ledger from Sep 25 to Oct 17 shows Grok raised 519 findings, of which 309 held up on manual re-check and 207 were absent from Claude's draft, while Codex raised 50 findings with 39 holding up. The team reports Grok reviews cost a median of about two cents each, though Codex runs are token-heavy and hit usage limits seven times in the same period. The authors acknowledge that Tīrtha's track record is only a few weeks old and that the ledger was partly graded by Claude itself, so the benefits remain unproven.