# Microsoft Unveils MAI-Code-1.1-Flash: Enhanced Coding Model for Competitive Edge > Microsoft's upgraded MAI-Code-1.1-Flash aims to outshine competitors with improved performance and efficiency. > By Dave · 2026-08-12 > Source: https://otf-kit.dev/blog/mai-code-1-1-flash-upgrade ## A 75% price cut that's actually earned Microsoft dropped MAI-Code-1.1-Flash on August 12, and the headline number is the one that matters: the new model costs [one-quarter of what its predecessor did](https://www.neowin.net/news/microsoft-releases-mai-code-11-flash-coding-model-to-better-compete-with-chinese-models/) — and it isn't a discount. It's a vertical gain on every axis. Same vendor, same workloads, 22% better on Terminal-Bench 2.1, 15% better on .NET tasks, 25% faster token streaming, 25% fewer tokens for the same answer. When the cheaper thing is also the better thing, that's a real enable for anyone running Copilot at scale. The narrative around it is sharper. Back in June, Microsoft shipped its first inference-efficient coding model tuned for GitHub Copilot workloads. Within weeks, GLM-5.2, Kimi K3, and OpenAI's GPT-5.6 Luna had pulled ahead on cost-or-quality. MAI-Code-1.1-Flash is the response — and it's the right one. Honest appraisal first: it's an impressive turnaround from a vendor that was, six weeks ago, getting out-shipped by open Chinese models on raw capability-per-dollar. Worth a moment in the spotlight. ## What changed under the hood The improvements are concrete, not vibes. Lining them up against the older MAI-Code-1-Flash: | Category | Metric | MAI-Code-1.1-Flash | MAI-Code-1-Flash | |---|---|---|---| | Performance | Terminal-Bench 2.1 (CLI) | +22% | baseline | | Performance | .NET tasks | +15% | baseline | | Performance | Token streaming | +25% | baseline | | Performance | Tokens per task | −25% | baseline | | Pricing | Input $ / 1M | $0.20 | $0.75 | | Pricing | Cached input $ / 1M | $0.02 | $0.075 | | Pricing | Output $ / 1M | $1.20 | $4.50 | The "1/4x price" claim is the average of those three token rates. The cached-input rate in particular — $0.02 per million — is the line that quietly makes agent loops economical. A 30-turn debugging session that lived mostly on cached reads now costs a rounding error. Microsoft attributes the gains to training and serving efficiency improvements. The serving piece is the one that matters for latency: 25% faster streaming means the first token shows up faster, and that's the number a developer actually feels in the editor. ## Pricing, indexed For annual Copilot subscribers, MAI-Code-1.1-Flash is charged at a 0.25x premium request multiplier. Translated: Copilot subscribers get a discount on the already-cheaper model. The flat per-token rates if you're hitting the API outside a subscription: ```bash # Per 1M tokens, USD input $0.20 cached_input $0.02 output $1.20 ``` Free and Student Copilot users get the model through automatic model selection — they don't pick, the router picks for them. Paid users (Pro, Pro+, Business, Enterprise) can manually select it from the model picker. That's the rollout split, and it's the kind of tier-aware rollout that suggests Microsoft has actually thought about the upgrade path for pro users who'd notice the gain. ## How to actually use it today It's rolling out across the whole Copilot surface: VS Code, Visual Studio, JetBrains IDEs, Copilot CLI, GitHub Mobile, and other supported experiences. The mechanics differ by tier, so here's the concrete path: **Free / Student** — no action. The router picks MAI-Code-1.1-Flash where appropriate. **Paid users** — open the model picker in any supported IDE and select it: ```text 1. Open the Copilot Chat panel 2. Click the model selector at the top of the panel 3. Choose "MAI-Code-1.1-Flash" from the list 4. Pin it as the default if you want it across all sessions ``` **Copilot CLI** — the terminal is where the 22% improvement is most visible, since Terminal-Bench 2.1 is a CLI benchmark. Pass the model id on the command line; the exact flag varies by Copilot CLI version, so check `gh copilot --help` for the current syntax in your install, then set it as the default for repeat use: ```bash # Confirm what's available gh copilot --help # One-shot question with the new model gh copilot ask --model mai-code-1.1-flash "what does this regex do?" # Set as default for the session gh copilot config set default-model mai-code-1.1-flash # Verify gh copilot config get default-model ``` If you're running automation through `gh copilot` today, swap the model and re-run your slowest prompts. The deltas will show up in the same places the benchmarks highlighted: terminal reasoning, multi-step shell tasks, .NET projects. ## The vision addition is small but real Beyond the coding win, MAI-Code-1.1-Flash adds native vision support for understanding images. Not a coding feature, exactly — a UI-asset feature. Drop a screenshot of a broken layout into the chat and the model can read it. For frontend developers shipping visual debugging flows, this is the thing the original Flash was missing. It's not multimodal in the "draw me a wireframe" sense. It's multimodal in the "I can OCR a stack trace and read your error toast" sense. Useful, narrow, honest. ## The retirement date is your migration deadline GitHub is retiring MAI-Code-1-Flash across all Copilot experiences on September 10, 2026 — roughly a month from the August 12 release. If you have any pinned workflows on the old model, you have one full release cycle to migrate. The 1.1 variant is a drop-in replacement at the prompt surface, so this is a string change, not a rewrite. A practical migration checklist: ```text 1. Audit any pinned model references in repos, CI scripts, agent configs 2. Replace the old model id with the new one (use the Copilot model picker or `gh copilot config get default-model` to confirm spelling) 3. Re-run your evaluation suite — the 22% / 15% deltas are net positives, but any prompt tuned to the old failure modes may need a refresh 4. Update internal docs and any pricing calculators 5. Set a calendar reminder for Sept 10 to confirm the old model is gone ``` Step 3 is the one people skip. A 22% improvement on a benchmark doesn't mean every prompt improves — it means the average improves. Specific failure modes may shift. Re-run your evals. ## The part that doesn't churn The win here is real and worth shipping. Take the model swap, run the 75% cost cut, re-run your agent evals, retire the old id by September 10. That's the move. The companion thought, said without framing it as a reason to avoid the model: the model is going to change again. Microsoft's roadmap will iterate. OpenAI will ship something. Anthropic will ship something. A Chinese lab will ship something cheaper. The last twelve months have been a clinic in churn at the model layer. The part that doesn't churn is the durable layer underneath: the components that have to render the same on web, iOS, and Android; the auth that has to survive a model rewrite; the data layer that the agent writes to; the build that has to ship on every platform. Every quarter the model moves, and the only thing that protects your velocity is making sure the surface area the model touches is small, stable, and owned. One API for a button. One API for a form. One API for a session. The model swaps; the components don't. ![a stack where the model name is a config string and the components are the durable substra](https://cdn.otf-kit.dev/blog/mai-code-1-1-flash-upgrade/inline-1.png) Ship MAI-Code-1.1-Flash today. The moment the next version lands, that'll be a one-line change too — and that's by design.