Intel / article

Measuring Time Horizon using Claude Code and Codex

Source: metr.org
Date: Published Feb 13, 2026

Context

A coding assistant designed for interaction need not score better in an unattended . In METR’s comparison of agent software, GPT-5 in Codex and Opus 4.5 in Claude Code did not show a statistically significant time-horizon advantage over the evaluator’s default setups. Each model was compared across the surrounding software rather than against a different model.

The conditions limit what follows from that result. The tasks provided no human feedback, the integration wrappers had rough edges, and the GPT-5 comparison did not use the specialized GPT-5-Codex model. The study helps check whether the evaluator’s simpler setups miss a large advantage in these tests; it does not establish equivalence across everyday interactive coding work.

Terms in this piece · Glossary

benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source metr.org