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-6.1-sol", "gpt-6-sol", "gpt-6-astra"]
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
16 November 2026 · USD per million tokens. Prices are U-M Toolkit standard estimates. Scores use Artificial Analysis Intelligence Index v5.0.1; higher is better. Price conditions, evaluator settings, and sources.
This shortlist covers 18 currently advertised text models. Eleven have a verified score for the displayed effort; missing scores remain blank rather than borrowing a result from another index version or model revision.
| Model | Evaluated effort | Index | U-M input | U-M output | Tokens/s¹ |
|---|---|---|---|---|---|
| GPT-6 Luna | Medium | 18 | $0.10 | $0.50 | 95 |
| GPT-6 Sol | Medium | — | $2.00 | $10.00 | — |
| GPT-6 Astra | Medium | 52 | $10.00 | $50.00 | 44 |
| GPT-6.1 Sol | Medium | 46 | $2.00 | $10.00 | 52 |
| Claude Haiku 4.5 | Reasoning | 7 | $1.00 | $5.00 | 116 |
| Claude Haiku 5.5 | Medium | 25 | $0.10 | $0.50 | 221 |
| Claude Sonnet 5 | Medium | — | $2.20 | $11.00 | 102 |
| Claude Sonnet 5.5 | Medium† | 31 | $2.00 | $10.00 | 128 |
| Claude Opus 5.5 | Medium† | 48 | $4.40 | $22.00 | 107 |
| Claude Fable 5.1 | Medium† | 43 | $11.00 | $55.00 | 87 |
| Claude Fable 5.2 | Medium† | 56 | $8.80 | $44.00 | 104 |
| Gemini 3.8 Flash | Medium | 30 | $0.75 | $3.75 | 118 |
| Grok 4.6 | Xhigh | 31 | $1.25 | $6.00 | 71 |
| DeepSeek V4 Flash 0731 | Max | — | $0.19 | $0.51 | 141 |
| DeepSeek V4.1 Flash | Revision unverified | — | $0.30 | $1.20 | — |
| DeepSeek V4 Pro | Revision unverified | — | $1.74 | $3.48 | — |
| Kimi K3 | Max | — | $3.00 | $15.00 | 48 |
| Kimi K2.5 | Reasoning | — | $0.60 | $3.00 | 43 |
† Evaluator's default fallback configuration. ¹ External end-to-end output speed, including time to first token; not measured through U-M. “—” means no verified comparable value. Efforts differ where shown; these are not settings the helper applies. Detailed qualifications.
Rates depend on context/cache and can change. Haiku 5.5's provider has a higher-price tier above 100K input tokens; GPT-6.1 Sol's provider has a tier above 272K. Provider conditions versus U-M estimates. Token prices do not determine cost per completed task.
Luna remains the repository default for both identities. Compare models on your own tasks and verify the selected route before use; a listed ID and an external benchmark do not establish Codex tool-call compatibility.
Higher and further left indicates a higher benchmark score at a lower token price. No uncertainty intervals are supplied; small score gaps may not be meaningful. Full-size chart · Editable SVG · Current data and sources (CSV)
The September 2026 comparison, its provider prices, and v4.3.2 scores are archived separately. The two index versions are not directly comparable.
