Models and reasoning effort
The AI models available on Codexe, which plan each needs, defaults, reasoning effort levels, relative costs and output limits.
Updated
On this page
Codexe isn't tied to one AI provider. You choose a model per project (on Create or in Project settings), and your plan decides which models you can pick.
#Available models
| Model | Provider | Minimum plan | Relative cost |
|---|---|---|---|
| DeepSeek V4 Flash | DeepSeek | Free | 1× |
| GPT-6 Luna (default on Free) | OpenAI | Free | 1× |
| Claude Haiku 4.5 | Anthropic | Free | 1× |
| Grok 4.3 | xAI | Free | 1× |
| DeepSeek V4 Pro | DeepSeek | Pro | 1× |
| Claude Sonnet 5 (default on paid plans) | Anthropic | Pro | 2× |
| Grok 4.6 | xAI | Pro | 2× |
| GPT-6 Sol | OpenAI | Pro | 2× |
| Claude Opus 5.5 | Anthropic | Business | 3× |
| Claude Fable 5.1 | Anthropic | Business | 8× |
| GPT-6 Astra | OpenAI | Business | 12× |
| GPT-5.5 Pro | OpenAI | Enterprise | 41× |
"Relative cost" is how many tokens a model uses for the same work compared with the cheapest ones. The pricing page always shows the current list and multipliers, and the model picker only offers what your plan includes.
#Choosing a model
- Small plugins, scripts, edits: the defaults are fast and cheap.
- Large multi-file projects, mods, tricky APIs: a stronger model (Sonnet 5, GPT-6 Sol, or Opus 5.5 and up on Business) usually gets more right the first time.
- Web search only works with the GPT-6 and GPT-5.5 models. See Web search.
#Reasoning effort
Reasoning effort is how much the model thinks before answering. Choosing it is a Pro feature; Free accounts run every model at its lowest setting.
| Model | Levels you can choose |
|---|---|
| Claude Sonnet 5, Opus 5.5, Fable 5.1 | Low, Medium, High, xHigh, Max |
| GPT-6 Luna, Sol, Astra | Low, Medium, High, Extra High, Max |
| GPT-5.5 Pro | Medium, High, Extra High |
| Grok 4.3, 4.6 | Low to Extra High |
| DeepSeek V4 | Low, High, Max |
| Claude Haiku 4.5 | No effort control |
Low and Medium cost the same. High, Extra High and Max cost progressively more tokens per request. They're worth it for complex, multi-file work and wasted on a one-line change.
#Output limits
On the Free plan a single reply can be at most 4,000 output tokens. A project's first generation is allowed up to 16,000 so the whole project fits, and Codexe pays for the part above 4,000 itself. Paid plans have no per-reply cap beyond the model's own limit.
#Retired and replaced models
When a provider retires a model, projects pinned to it ask you to pick a current model in settings before generating again (chat messages fall back to the default model automatically). Older models that were replaced by newer versions disappear from the picker but keep working for projects that already use them.