Pydantic's code mode cuts agent tokens 30% and speeds long tasks 51% across 8 models
samuelcolvin · x · 2026-10-29
Pydantic released a report on its code mode: instead of issuing 200 tool calls to update 200 records, the model writes a single Python loop. Across 8 models and 5,276 runs, long tasks ran 51% faster and succeeded 16 points more often.
User @dlewis reports a 30% token reduction after implementing code mode with Monty, plus the ability to scale to tools with very long outputs by letting the agent write filtering code.
Related event: Pydantic's Code Mode Cuts Agent Tokens 30%, Speeds Tasks 51%(2 posts)→
More from coding & agent
- AI writes, tests and reviews the code — and engineers are losing understanding of their own systems — julianweisser · 2026-10-29
- Coding agents burn 95-99% cached tokens per benchmark task, token usage chart shows — AccBalanced · 2026-10-29
- Open-source screen recorder Cap v0.6.2 ships 89% faster share links, 2.8x exports — JosephJacks_ · 2026-10-29
- Professor builds videos with Codex and Blender, where creative goals resist alignment — CSProfKGD · 2026-10-29
- Grok voice agent breaks down, sparking rant against renting other people's agent harnesses — alexcovo_eth · 2026-10-29
- A silent skip counted as success: rebuilding a bot lane that can't go quiet unnoticed — petrusenko_max · 2026-10-29