Codex CLI — 8 OpenAI models head-to-head · V1
Codex 8-Model Observation
- fastest bug-fix
- 13s
- spread (fastest→slowest)
- 17.8×
- models perfect on T3
- 7/8
- mypy / ruff errors anywhere
- 0
Speed king · Thoroughness king · Dead model
Speed king
gpt-5.3-codex-spark
Bug fix in 13 seconds (1 bug/sec). Code generation in 29 s. Test writing in 62 s with 33 tests at 100% coverage. Architecture in 27 s at 170 words/sec. The 3.6× faster runner-up gpt-5.4-mini still trails Spark on every task.
Thoroughness king
gpt-5.2
T1 code-gen 520 s producing 514 LOC + 2 test files. T6 architecture 773 s producing 8,659 words across 14 files. Use when "complete and exhaustive" matters more than speed.
Dead model
gpt-5.1-codex-mini
Failed every task. 0 files, 0 LOC, 2-3 second timeouts, 7 of 19 on bug fix. Effectively non-functional for agentic coding tasks. Do not use.
The speed-task observed score table
- gpt-5.1-codex-mini
- 2
- gpt-5.3-codex-spark
- 13
- gpt-5.4-mini
- 47
- gpt-5.2-codex
- 83
- gpt-5.3-codex
- 93
- gpt-5.4
- 129
- gpt-5.1-codex-max
- 146
- gpt-5.2
- 232
8 entries · ordered by Duration (lower is better) as observed in this run · not a ranking
Reasoning level has ZERO quality impact on test-driven tasks. Every Codex model at every reasoning level (low / medium / high / xhigh) achieved 19/19 perfect on T3. Reasoning level only affects speed — low is consistently 1.5–2× the speed of xhigh for identical output. Reserve xhigh for open-ended architecture work where the additional reasoning depth matters.
What every model does well
T5 test write
All 7 working models hit 100% coverage
Tied for most tests written: gpt-5.3-codex-spark and gpt-5.2 at 33 tests each. Spark did it in 62 s; gpt-5.2 took 329 s.
T6 architecture
gpt-5.2 produced 14 files / 8,659 words
Spark produced 4,601 words in 27 s. The "complete the system" task is where reasoning depth and persistence matter most.
--full-auto trap
--full-auto still uses workspace-write sandboxing
True agent mode requires --dangerously-bypass-approvals-and-sandbox. This was the most-asked clarification across the observation; it now appears in every Codex deployment doc.
Reasoning level paradox
Same quality, different speed
For test-driven tasks (T1 code gen, T3 bug fix, T5 test write), the model produces the same final artifact at low / med / high / xhigh — but takes 1.5–2× longer at xhigh. Use low unless the task is open-ended.