Coding agent cuts token cost 80% by batching actions into one turn
typically loop through inspect, wait, patch, wait, build, wait, test, wait — calling the LLM again at every step, which burns time and money. An project called Tura batches actions that don't need fresh reasoning into a single , cutting the number of LLM round-trips (turns) by 80%.
On the Deep SWE benchmark (60 coding tasks), Tura in Balanced/High mode passed 48/60 tasks (80%) using about 230 million tokens, 2,017 rounds, and roughly $221 in estimated cost. By comparison, Codex CLI on Medium passed only 38/60 (63.3%) while using over 330 million tokens, 3,140 rounds, and about $257, and Codex CLI on High did even worse: 36/60 (60%) passed at over 450 million tokens, 6,074 rounds, and about $327.
A lighter Tura (Direct/High) hit 65% (39/60) using just 75 million tokens, 969 rounds, and about $100. Full code and benchmark details are published on GitHub (Tura-AI/tura).
Key points
- Batches non-reasoning steps to cut LLM round-trips (turns) by 80% versus typical inspect-wait-patch loops
- On the 60-task Deep SWE benchmark, Tura's Balanced/High mode hit an 80% pass rate versus 60-63% for Codex CLI
- Cost comparison: roughly $221 for Tura versus up to $327 for Codex CLI
- Code and full benchmark results are open on GitHub (Tura-AI/tura)