Pi Coding Agent swap cuts gpt-6-sol limit burn to 10%

The switch did not start as a search for a better tool. It started as a quota problem. Codex with gpt-5.6-sol at high thinking once used about 10% of the 5-hour limit for a single implementation pass; that figure later rose to 15–20% even at medium, and gpt-6-sol occasionally spiked to roughly 40% during a single task. Review, once 5–10%, settled above 10% as well.

The developer suspected the harness itself was part of the cause — that the large system prompts Claude Code and Codex add to advertise their many features were feeding the model irrelevant information and burning tokens. Pi Coding Agent was adopted with subagents deliberately left out, since they were one of the suspects. A short appended system prompt and a minimal web extension replaced the heavier defaults.

The result was less about raw performance than about predictability: gpt-6-sol at high thinking now holds near 10% for implementation, and the developer reports no felt drop in quality. Review still costs about 10%, so the review step was handed to Opus 5.5, whose low limit consumption made that possible. Whether this holds once the workflow grows more complex — more subagents, longer tasks — is the part still unproven.