Microsoft MAI-Code-1-Flash: A $0.75/M Coding Model Now Default in Copilot

Microsoft Now Has Its Own Coding Model

Microsoft has rolled out MAI-Code-1-Flash to general availability as a default model option inside GitHub Copilot and VS Code. It's an inference-efficient agentic coding model — roughly 5 billion active parameters and around 51% on SWE-bench Pro — and crucially, it's Microsoft's own, not a rebadged OpenAI or Anthropic model.

The headline for cost-conscious developers is the price. Copilot's per-million-token rate for MAI-Code-1-Flash starts around $0.75, putting it in the same budget tier as DeepSeek V4 Pro ($0.435/$0.87) and Kimi K2.7-Code ($0.75/$3.50) — and far below frontier models like GPT-5.5 ($5/$30) or Claude Opus 4.8 ($5/$25).

Why does Microsoft want its own cheap model? Because under Copilot's new usage-based billing, every token you consume draws down credits — and if Microsoft is paying OpenAI or Anthropic wholesale for those tokens, its margins shrink. A capable house model that costs Microsoft far less to run changes the economics for both Microsoft and you.

What a Cheap Default Does to Your Credit Spend

GitHub Copilot moved to usage-based "GitHub AI Credits" earlier this year, metering consumption by token rather than by a fixed number of premium requests. That makes the default model a major lever on your monthly bill: whatever model handles your routine completions and chat sets the baseline burn rate.

Consider a developer doing 200 meaningful coding interactions a month, each averaging ~25,000 input and ~4,000 output tokens. On MAI-Code-1-Flash at roughly $0.75/M, that's about $4.40 in token cost. The same workload routed entirely through GPT-5.5 ($5 input / $30 output) would run closer to $50 — more than 10x.

That gap is exactly why a cheap default matters. If MAI-Code-1-Flash handles the bulk of everyday work — boilerplate, simple edits, code explanation — and you reserve a frontier model only for genuinely hard reasoning, your blended cost per interaction drops dramatically without a meaningful quality hit on routine tasks.

The Catch: Capability Has a Floor

A ~51% SWE-bench Pro score is solid for a 5B-active model, but it's well below the ~69% range that flagship models like Claude Opus 4.8 reach. On complex, multi-file refactors or subtle bug hunts, a cheaper model that's almost right can cost you more than an expensive one that's right the first time — because every failed attempt re-spends input tokens on the same context.

This is the recurring theme of usage-based billing: the cheapest model per token is not always the cheapest per completed task. A budget model that needs three tries to land a change can quietly out-cost a frontier model that lands it once. The win comes from routing — cheap model for routine work, expensive model for hard work — not from using the cheapest model for everything.

How to Use It Well

Make MAI-Code-1-Flash your default for completions and routine chat, and manually escalate to a frontier model when you hit a problem that needs real reasoning. Watch the ai_credits_used field that Copilot's usage API now exposes per user — it's the fastest way to see whether your model mix is actually saving money or whether retries are eating the savings.

The arrival of a cheap, capable house model is good news for Copilot users: it lowers the floor on everyday coding cost. Just don't let "cheapest per token" become "cheapest per task" in your head — they're different numbers. To compare blended cost across a realistic model mix, plug your workload into our AI coding cost calculator.