Change defaults and order · Compare models · All models and source details
Compare models below before setting up U-M GPT. Once your
umgpt or umgpt-low identity exists, use the steps here to change its saved
default and picker order.
Run these commands from the repository directory, after creating the identity you want to update. The helper skips either standard identity that does not exist.
1. Find your settings file.
./bin/codex-home model-settingsOpen the path printed after Edit your settings:, normally
~/.codex-homes/umgpt-models.toml. Your choices survive repository updates.
2. Choose your defaults and favorites.
umgpt = "gpt-6-luna"
umgpt_low = "gpt-6-luna"
umgpt_order = ["gpt-6-sol", "gpt-6-astra", "o3"]
umgpt_low_order = ["claude-sonnet-5", "gemini-3.8-flash", "claude-haiku-4-5"]The first two lines select defaults; the arrays put your other favorites in
order. Keep each array on one line. Use [] for automatic ordering.
Find exact IDs in the available model list.
The picker shows your default → your favorites → remaining models by family, newer versions first. Sorting never changes your default. The example above starts the low picker with Luna → Sonnet → Gemini → Haiku.
3. Apply, reload, and check.
./bin/codex-home apply-model-settingsRestart the CLI, or run Developer: Reload Window in VS Code and start a
new Codex chat. Check your choices in the model picker. Run
model-settings again to compare saved and active defaults.
Changes not showing? · Reset or use a model for one session
24 September 2026 · USD per million tokens. Scores use Artificial Analysis Intelligence Index v4.3.2; higher is better. Prices use provider rates, with only Opus 5.5's verified 10% U-M adjustment. Sources and pricing conditions.
| Model | Index | Input | Output | Context¹ | Tokens/s² |
|---|---|---|---|---|---|
| GPT-6 Luna | 29 | $0.10 | $0.50 | 1.05M | 142.8 |
| GPT-6 Sol | 40 | $2.00 | $10.00 | 1.05M | 114.3 |
| GPT-6 Astra | 50 | $10.00 | $50.00 | 1.05M | 45.3 |
| Claude Haiku 4.5 | 17 | $1.00 | $5.00 | 200K | 96.1 |
| Claude Sonnet 5 | 28 | $2.00 | $10.00 | 1M | 67.1 |
| Claude Opus 5.5 | 51 | $4.40 | $22.00 | 1M | 77.9 |
| Claude Fable 5.1 | 49 | $10.00 | $50.00 | 1M | 54.1 |
| Gemini 3.8 Flash | 40 | $0.75 | $3.75 | 1M input | — |
| DeepSeek V4 Flash 0731 | 34 | $0.14 | $0.28 | 1M | 221.1 |
| Kimi K3 | 44 | $3.00 | $15.00 | 1M | 35.3 |
| Kimi K2.5 | 23* | $0.60 | $3.00 | 256K | — |
| Llama 4 Maverick | 10* | $0.20 | $0.80 | 1M | 114.9 |
| Llama 4 Scout | 8* | $0.10 | $0.30 | 328K | 100.1 |
* Estimated by the evaluator; independent evaluation on this index is pending. ¹ Rounded provider limits, not verified U-M limits. ² Output speed after the initial wait, not measured through U-M. “—” means not reported. Test settings, cached-input prices, exact limits, and qualifications.
In this shortlist, Sonnet 5 at medium effort is the closest Claude option to Luna on this index. Haiku is the cheaper Claude alternative. Luna is the repository default for both identities.
Higher and further left means a higher score at a lower token price; token price does not measure cost per completed task. The chart also shows Sonnet at high effort (index 32; 65.4 tokens/s). Small score gaps may not be meaningful.
