Which AI model should you use for coding?
There is no single best model for coding — there is a right model for each step. How Ava Supernova routes work across two fleets, and how to pick between them.
Ask which model is best for coding and you will get a different answer every month, usually from whoever shipped most recently. It is the wrong question. A real coding task is not one request — it is reading the codebase, planning a change, writing it, then checking it did what you meant. Those four jobs do not want the same model, and paying frontier prices for all four is how people end up with a bill they cannot explain.
The better question is how the work gets routed. That is what Ava Supernova is built around, and it is why an account gives you two complete fleets rather than one model behind a logo.
Maestro — one conductor, end to end
Maestro runs the whole pipeline on Qwen: the same model serves Scout, Architect, Builder and Verifier. Nothing is handed between providers, so behaviour stays consistent from the first read to the final check, and the cost is predictable because there is only one rate card involved.
It is the default, and for most people it is the right default. Pick it for daily work, for predictable spend, and for wanting Ava to simply work without you thinking about routing at all.
Supernova — a specialist for each step
Supernova is the polyglot. A DeepSeek Flash coordinator reads the task and picks the right specialist for each part of it, with Qwen 3.7 Plus taking Builder and vision work. The coordinator runs a 1M token context, so it can hold a large codebase in view while it decides who does what.
Pick it for heavy multi-step work — a refactor across many files, a migration, anything where the planning and the writing genuinely want different strengths.
Why Qwen and DeepSeek
Because they earn it on the two things that matter for agentic work: they are strong at code, and they are cheap enough to run across every step of a task rather than only the glamorous one. An agent makes a lot of calls. A model that is excellent and unaffordable gets used once and then rationed, which defeats the point of having an agent at all.
Both fleets are live on every plan, including Free. You switch between them whenever you like — it is a setting, not a tier.
When you want something else entirely
Bring your own key. Ava speaks to DeepSeek, Qwen, Kimi, Mistral, Zhipu, MiniMax, Xiaomi, Tencent Hunyuan and NVIDIA directly, plus anything OpenAI-compatible — including models running on your own machine, where nothing leaves it at all. Those requests go straight to the provider and never touch your credit allowance.
Kimi is the usual reason people reach for this, and it is a perfectly good one. We do not serve it on an account; your own key works fine.
The honest summary
There is no best model for coding, and anyone selling you one is selling you one model. There is a right model for each step of the work, and the thing worth having is the routing that puts the right one on each step without you managing it. That is what the fleets are, it is what an account buys, and Free gives you 300 credits a month to test the claim before you pay anything.