Model Recommendations

info

This advice is current on 2026-10-09. It is subject to change. The MAG Dashboard shows the models that you can use now.

Use this page to select a model for a task. For the name format, see Model Selection & Options.

Task	Model names	Providers
Hard math and hard coding	claude-opus, gpt-6.1-sol	claude-opus: amazon, google. gpt-6.1-sol: openai, microsoft, amazon
Routine coding	gpt-6-luna:high, claude-sonnet, gemini-flash	gpt-6-luna: openai, microsoft, amazon. claude-sonnet: amazon, google. gemini-flash: google
Short questions, summaries, and data extraction	gpt-6-luna:low, claude-haiku, gemini-flash-lite	gpt-6-luna: openai, microsoft, amazon. claude-haiku: amazon, google. gemini-flash-lite: google
Safety checks on text	gpt-oss-safeguard-120b	amazon
Embeddings	text-embedding-3-large, cohere-embed-v4	text-embedding-3-large: openai, microsoft. cohere-embed-v4: amazon

Notes

• A reasoning model needs enough output tokens. Set max_completion_tokens to 2000 or more.
• A higher effort gives better answers for hard tasks. It also uses more tokens and more time.
• Use the :flex modifier for work that can wait. The price is lower.
• Pin the provider when you must use one allocation bucket. See Model Selection & Options.