Large Language Models
Starburst can use several different large language models to generate responses. By default, the following models are available in Starburst:
| Model | Avg cost per 100 queries | Average response time | Notes |
|---|---|---|---|
| Google Gemini 3.6 Flash | US$0.50* | 7 seconds | (Recommended) The best Google model here in terms of speed and quality. [*promotional pricing until January, 2027] |
| Google Gemini 3 Flash Preview | US$0.50 | 8 seconds | Only slightly slower than 3.6 Flash, and its responses are likely to be just as high quality in nearly all realistic student uses. |
| OpenAI GPT-5.6 Luna | US$0.14 | 6 seconds | (Recommended) The best OpenAI model for most cases. High quality responses at a high speed. |
| OpenAI GPT-4.1 nano | US$0.05 | 2 seconds | The lowest-cost and fastest of these options. It can provide accurate, helpful responses in a wide variety of simple applications, but it will be noticeably less accurate in more complex or less common topics and may exhibit less fluency in languages other than English. A good choice only if cost is the most important factor. |
Not recommended
The larger frontier models from OpenAI and Google produce very high-quality responses, but their cost and/or speed make them poor choices. This includes the Gemini “Pro” models from Google and GPT-5.6 Sol from OpenAI. Their costs will be two to five times the cost of the most expensive models above, and they could take twice as long (or more) to respond.
These models will produce some of the highest quality responses available, but the improvement over the recommended models above will be relatively small, especially relative to their substantially slower generation and higher costs. They’re probably not worth the cost and latency except for graduate-level classes, niche topics, or uncommon languages. Most classes and students are unlikely to see a useful difference in responses between these and the above models.
They are not provided as options by default, but you can always configure one as a custom model for your own classes if you want to try it.