Small fixes and scoped subagent leaves.
Software engineering · September 2026
Choose the model for the shape of the work.
I compared Claude Opus 5, Claude Fable 5, and every GPT-5.6 tier and effort level, without pretending vendor benchmarks form a complete league table.
Measured evidence
Published Coding Agent Index
These are the high-end configurations reported in OpenAI's shared Artificial Analysis table. I have no measured figure for Opus 5: it was not included.
Operating guide
Reasoning effort changes the job profile.
Routine features and test repair.
Ambiguous bugs and multi-file work.
Large migrations and deep verification.
Deepest single-agent reasoning.
Parallel research, implementation, and audit.
Sol and Terra support Ultra in the current Codex lineup. Luna, Opus 5, and Fable 5 stop at max.
Ultra under the microscope
More agents change the topology.
OpenAI reports Sol rising from 88.8% at max to 91.9% at Ultra on Terminal-Bench 2.1: a 3.1 percentage-point improvement.
88.8%Sol max+3.1 pp91.9%Sol Ultra
I use Ultra when branches can proceed independently and one integrator owns interfaces, conflicts, and final proof. I stay on max for tightly coupled hotspots.
Routing guide
Where I would start
Primary sources