LLM-Guide: Which model should I use when?

MAIA lets you choose between many language models. For most questions you don't need to: with “Auto”, MAIA picks a suitable model for each request. This guide helps when you want to choose a model yourself.

As of October 2026, release 1.24. For the full list of models and how to select them, see Model Selection Reference.

In short

“Auto”: the default

With “Auto”, MAIA analyzes each request and uses benchmark data to pick the model that fits this type of question best. Within a chat, the model can change from question to question. Each answer shows which model produced it, for example “Model: GPT-5.6 Luna”.

“Auto” only uses new models once we've tested them thoroughly. That's why you currently select Claude Opus 5 and 5.5, Claude Sonnet 5.5, GPT-6 Sol and Luna and the open-weight models manually.

The model families

Claude (Anthropic)

• Claude Opus 5.5 and Opus 5: For the most demanding tasks, such as in-depth research, complex analyses and tasks with many steps. Opus 5.5 is the newer model.
• Claude Sonnet 5.5: A strong everyday all-rounder, thorough and still quick.

All three are also available with “(Reasoning)”.

GPT (OpenAI)

At OpenAI, the names Sol, Terra and Luna stand for performance tiers: Sol is the stronger variant, Luna the faster one, and Terra sits in between. All GPT-5.6 and GPT-6 models always use reasoning.

• GPT-6 Sol: For demanding tasks such as research and analyses with many steps. According to OpenAI, it makes far fewer mistakes than GPT-5.6 Sol.
• GPT-6 Luna: Fast, for clearly defined tasks, for example summarizing documents, extracting information or answering quick questions.
• GPT-5.6 Sol: The strongest model in the GPT-5.6 series, for difficult, multi-step questions.
• GPT-5.6 Terra: Balances quality and speed, good for everyday tasks and document analysis.
• GPT-5.6 Luna: The fastest model in the GPT-5.6 series. For questions that need a lot of text at once, it performs worse than Sol.
• GPT-4.1, GPT-4.1 mini and GPT-4.1 nano: Work without reasoning and therefore answer quickly. The smaller mini and nano variants suit short, simple questions.

Gemini (Google)

• Gemini 2.5 Pro: Always uses reasoning. Google describes it as a model for complex questions and for analyzing large datasets and long documents.
• Gemini 3.5 Flash: Google's fast workhorse model, good for tasks with several steps. Available with and without reasoning. For harder questions, choose the “(Reasoning)” variant.

Open-weight models

Gemma 4, Apertus 70B, DeepSeek V4.1 Flash, GLM 5.3, Mistral 4 and Qwen 3.8 run entirely in the EU through a European provider. They're our first fully EU-sovereign models. Choose one of them if that matters to you. “Auto” doesn't use them.

With or without reasoning?

Models labeled “(Reasoning)” think about your request internally before answering. That pays off for complex questions, for example when MAIA should compare several documents, spot contradictions or work something out step by step. The answer takes a little longer. For simple questions you don't need reasoning.

Some reasoning models have a limit per person and model in a rolling 30-day window: 50 requests on the Professional plan, 1,000 on the Enterprise plan. See Model Selection Reference.

Tips

• Stick with “Auto” if you're unsure. It's the best choice for most questions.
• Compare for important questions. Ask the same question with two models and compare the answers.
• Your choice sticks. A manually selected model also applies to new chats until you change it or switch “Auto” back on.
• For especially precise answers, there's also the MAIA High Precision Mode.

Questions or feedback? Email us at support@getmaia.ai.