MAI-Code-1.1-Flash — 73% Cheaper, Off By Default for Orgs
Microsoft's new Copilot model cuts price by 73% and adds vision. Enterprise admins must flip an explicit policy toggle or their teams never see it.
Microsoft shipped MAI-Code-1.1-Flash into GitHub Copilot on August 11 — a sparse mixture-of-experts model with 138 billion total parameters (5 billion active), native vision support, and a 73% price cut over its predecessor. If you run a Copilot Business or Enterprise tenant, none of your developers can select it right now. The policy toggle ships disabled, and unless an admin walks into Settings → Copilot → Models and flips it, this model does not exist for your org.
TL;DR
- What: MAI-Code-1.1-Flash rolls out across all Copilot surfaces — $0.20/$1.20 per million input/output tokens, 256K context, image input
- Gate: Enterprise and Business admins must explicitly enable the model policy. It ships off.
- Why it matters: The model picker is quietly becoming an enterprise governance surface — data flows, retention, approved providers
- Action: Audit your Copilot model policy now, before a developer files the ticket for you
MAI-Code-1.1-Flash — What Happened
The model replaces MAI-Code-1-Flash, which Microsoft launched at Build in June. The architecture stays sparse MoE but scales up: 138B total parameters with 5B active per inference, a 256K-token context window, and — for the first time in Microsoft’s own coding models — image input. Developers can now paste screenshots, architecture diagrams, or UI mockups directly into Copilot prompts.
Microsoft’s vendor-reported numbers claim 25% greater token efficiency and 25% faster token streaming compared to the June model. On benchmarks: a 22% improvement on Terminal-Bench 2.1 for CLI tasks, 15% improvement on .NET workloads, plus production telemetry showing code survival up 4% and return visits up 9%. These are Microsoft’s own numbers from their own telemetry pipeline — treat them as directional, not gospel, until your team validates against your own codebase.
The pricing is the headline. At $0.20 per million input tokens and $1.20 per million output tokens (with $0.02 for cached inputs), this is materially cheaper than anything else in the Copilot model picker. Annual subscribers pay at a 0.25× premium request multiplier. For teams running high-volume agentic loops — the kind where you’d otherwise burn GPT-5.6 or Claude Opus credits — this is a legitimate routing option that could cut your Copilot bill substantially.
Surface availability is broad: Copilot CLI, cloud agent, Copilot Chat on GitHub, VS Code, Visual Studio, GitHub Mobile, JetBrains IDEs, Eclipse, and Xcode. This is not a VS Code-only beta. Microsoft is shipping this everywhere at once, which makes the admin gate even more significant — because every surface becomes a potential access point the moment an admin enables it.
Why This Matters
The pricing story is real, but it is not the interesting part. The interesting part is the governance pattern Microsoft is normalizing.
When GitHub Copilot first launched, model selection was not a thing. You got GPT-4, it worked, you moved on. Then came the model picker — Claude, Gemini, GPT-5.x variants — and suddenly developers had choices. Now Microsoft is adding its own models to that picker, at dramatically lower prices, but with an enterprise policy gate that most orgs have never configured.
This matters because the model picker in GitHub Copilot is quietly becoming what cloud region selection used to be: a routine-looking dropdown hiding real decisions about data handling, retention commitments, and approved provider surfaces. MAI-Code-1.1-Flash accepts image inputs — that is a new data-flow surface your security team has not reviewed. It runs on Microsoft’s infrastructure under Copilot for Business terms — but have you verified whether your org’s approved model-provider list includes Microsoft’s own models, or only the third-party options you explicitly vetted?
Most enterprise teams are about to discover this model exists when a developer tries to select it and cannot. That conversation — “Why can’t I use the cheap model?” — should have happened proactively, not reactively. The admin who enabled Claude Opus six months ago without updating the data-flow assessment is the same admin who now needs to decide whether vision-capable inference on code screenshots changes the risk profile.
The broader pattern here mirrors what we covered when Copilot moved to usage-based billing: Microsoft is shifting complexity from the product surface to the admin surface. The developer experience gets simpler (more models, lower prices, auto-selection for free users). The governance burden gets heavier. Free and Student users receive MAI-Code-1.1-Flash only through auto model selection — they cannot deliberately choose it. Paid individual users can select it freely. But for every managed seat in a Business or Enterprise tenant, it is invisible until someone with org admin permissions makes a deliberate choice.
This is not a bug. It is a design decision that acknowledges model selection is now a policy domain, not a preference domain. The question is whether your org treats it that way.
MAI-Code-1.1-Flash accepts image inputs. If your org has policies around screenshot sharing, UI mockup data, or visual representations of proprietary architecture — those policies now intersect with Copilot. Review before enabling.
Compare this to how MAI-Code-1-Flash launched at Build: that model rolled out to free and individual users first, with enterprise following. This time, Microsoft gates enterprise explicitly from day one. They learned that “available by default” creates compliance conversations after the fact. “Off by default” forces the conversation to happen first — in theory. In practice, most admins will not know the toggle exists until someone asks.
The competitive context matters too. At $0.20/$1.20 per million tokens with a 0.25× multiplier for annual subscribers, MAI-Code-1.1-Flash undercuts every frontier model currently available in Copilot’s picker. Microsoft is not just adding a model — they are creating a pricing tier designed to make their own model the default routing choice for cost-conscious teams. If your agentic workflows currently burn Claude Opus or GPT-5.x credits, the economic argument for switching the commodity workloads to MAI-Code-1.1-Flash is straightforward. But “cheap” and “approved” are different questions, and only one of them shows up on the invoice.
Before enabling MAI-Code-1.1-Flash for your org, verify four things: (1) image input policy — does your data classification permit screenshots in AI inference? (2) Source-code retention terms under your Copilot for Business agreement. (3) Whether “Microsoft MAI-Code” is on your approved model-provider list. (4) Whether existing architecture guardrails cover the model’s agentic actions across CLI and cloud agent surfaces.
The Take
I keep coming back to a simple observation: the GitHub Copilot model picker now has more governance implications than most teams’ cloud provider selection process, and it gets zero percent of the scrutiny. You have a change advisory board for switching from us-east-1 to eu-west-1. You have nothing for switching from GPT-5.x to a Microsoft first-party model that accepts image input.
The 73% price cut is not charity. Microsoft wants MAI-Code-1.1-Flash to become the default routing target for high-volume Copilot usage — the model that eats the commodity tokens while frontier models handle the hard problems. That is a sound architecture. But it only works if your org has actually decided it is acceptable, rather than discovering it six months later in an audit.
My recommendation: enable it. The economics are real, and the performance numbers — even discounted for vendor bias — suggest it handles CLI and routine coding tasks well enough to justify the cost savings. But enable it after running through the governance checklist, not before. And if your org does not have a model-selection policy yet, this is the forcing function. Build one now, because the next model Microsoft adds to the picker will not be the last.