The AI models behind Lumi
Lumi works with leading AI models from Anthropic, OpenAI and Google. Instead of sending every request to one model, smart routing picks the model that fits your request and the current state of your project, while reusing model cache wherever possible. Each task goes to the model that does it best, and you save Chat credits.
How smart routing works
Understands your request
Before each reply, Lumi works out what you are asking for: a quick text or style change, a new feature, a large refactor, or a plan for something new.
Matches the best model
Based on the task and the current context, such as attached images, Plan mode, or a problem that is still unsolved, Lumi picks the model best suited to handle it.
Reuses model cache
When you keep working on the same project, Lumi tries to stay on the model that already has your project in its cache, so it does not have to read everything again. Replies come faster and use fewer Chat credits. If a model is unavailable, Lumi switches to a backup automatically.
Models Lumi works with
Updated November 2026.
Anthropic
Claude Fable 5.1
Anthropic's most capable model, built for the hardest reasoning and long, multi-step coding work.
Claude Opus 5.5
Excels at complex front-end work and multi-step changes across large codebases.
Claude Sonnet 5.5
Balanced and dependable with tools, a good fit for complex features and larger refactors.
OpenAI
GPT-6 Astra
OpenAI's most capable model, aimed at deep reasoning on difficult problems.
GPT-6.1 Sol
Strong at reasoning, planning and hard debugging, which makes it a good fit for Plan mode.
GPT-6.1 Terra
Fast reasoning with a long context window and image understanding.
Gemini 4 Argon
Google's most capable model, with a long context window and strong multimodal understanding.
How the models compare
What each model is best at. All of them can work with about 1 million tokens of context at once. Kimi K3.1 Pro and Gemini 4 Carbon are listed for comparison only.
| Model | Maker | Best at | In Lumi |
|---|---|---|---|
| Claude Fable 5.1 | Anthropic | The hardest reasoning and long, multi-step coding | Yes |
| Claude Opus 5.5 | Anthropic | Complex front-end work and large codebases | Yes |
| Claude Sonnet 5.5 | Anthropic | Reliable tool use for complex features and refactors | Yes |
| GPT-6 Astra | OpenAI | Deep reasoning on difficult problems | Yes |
| GPT-6.1 Sol | OpenAI | Planning and hard debugging | Yes |
| GPT-6.1 Terra | OpenAI | Fast reasoning and image understanding | Yes |
| Gemini 4 Argon | Long context and multimodal understanding | Yes | |
| Kimi K3.1 Pro | Moonshot AI | Open-weight model, strong at coding and 3D modeling | For comparison |
| Gemini 4 Carbon | Long-running tasks and cybersecurity, in staged release | For comparison |
Frequently asked questions
Can I choose the model myself?
You do not need to. Lumi automatically picks the most suitable model for every request, so you can focus on describing what you want to build.
Which AI model is best for coding?
It depends on the task. Claude Opus 5.5 and Claude Fable 5.1 are strong at front-end work and large codebases, GPT-6.1 Sol at planning and debugging, and Gemini 4 Argon at long context and images. On Lumi you do not have to choose: smart routing picks the model for each request.
How does Lumi save my Chat credits?
Credit use depends on the model and on how much work a request takes. Lumi reuses model cache wherever possible, so project content a model has already read is not processed again, and follow-up requests in the same project cost fewer credits.
Why doesn't Lumi use one model for everything?
Each model has different strengths in coding, reasoning, image understanding and speed. Matching the model to the task gives better results than relying on any single model.
Start building with Lumi