Codex Models — Which One to Use
Codex supports multiple OpenAI models. Default is gpt-5.4. You can switch per-session or set a permanent default.
Available Models
| Model | Best For | Trade-off |
|---|---|---|
gpt-5.4 | General coding, reasoning, tool use | Balanced speed and capability |
gpt-5.3-codex | Complex refactors, large codebases | Slower, higher quality |
gpt-5.4-mini | Quick tasks, subagents | Faster, less thorough |
gpt-5.3-codex-spark | Instant iteration, prototyping | Research preview, Pro only |
gpt-5.4
The default. Handles most professional coding work well. Strong at reasoning through multi-step problems. Good tool use for file edits, searches, and commands.
Use for: daily coding, debugging, code review, feature implementation.
gpt-5.3-codex
Industry-leading for pure coding tasks. Thinks longer, produces higher quality code. Better at complex architectural changes.
Use for: major refactors, migrating frameworks, untangling legacy code.
Downside: slower response times.
gpt-5.4-mini
Fast and cheap. Good enough for straightforward tasks. Codex uses this internally for subagents when speed matters more than depth.
Use for: simple renames, formatting, quick lookups, cost-sensitive automation.
gpt-5.3-codex-spark
Research preview for ChatGPT Pro subscribers. Optimized for near-instant responses. Good for rapid iteration when you’re exploring approaches.
Use for: brainstorming, quick prototypes, when you want immediate feedback.
Setting Your Model
In config.toml
Set your default:
model = "gpt-5.4"
Via CLI Flag
Override for a single session:
codex -m gpt-5.3-codex
Mid-Session
Switch without restarting:
/model gpt-5.4-mini
The TUI shows current model in the status bar.
Reasoning Effort
Some tasks need deeper thinking. Control this with model_reasoning_effort:
model_reasoning_effort = "medium"
Levels: minimal, low, medium, high, xhigh
Higher effort = more reasoning tokens = slower responses but better solutions.
When to increase:
- Debugging race conditions
- Architectural decisions
- Complex algorithm implementation
When to keep low:
- Simple edits
- File organization
- Routine refactors
Cost Perspective
Two ways to pay:
ChatGPT Subscription — Fixed monthly cost (Plus, Pro, etc.). Good for individual developers doing interactive work. Usage limits apply.
API Key — Pay per token. Better for:
- Heavy usage exceeding subscription limits
- CI/CD automation
- Predictable per-project billing
Use /status in the TUI to check token usage during a session.
Model Selection Strategy
My recommendation:
- Start with gpt-5.4 — It handles 90% of tasks well
- Switch to gpt-5.3-codex for major refactors or when gpt-5.4 struggles
- Use gpt-5.4-mini for quick automation tasks
- Try Spark if you’re on Pro and want faster iteration
Don’t overthink it. The default is good. Switch when you notice quality or speed issues.
Subagent Models
When using multi-agent features, Codex spawns subagents for parallel tasks. Control their model:
[agents]
max_threads = 6
max_depth = 3
Subagent model selection is configured per-agent via [agents.<name>] sub-tables with a config_file pointing to a separate TOML. Using a smaller model for subagents saves tokens.
Related
- Installation — Get started with Codex
- config.toml — Full configuration reference
- Plan Mode — Model selection affects planning quality