GitHub Cut Its Cheapest Copilot Coding Model's List Price 73% and Gave You Until September 10 to Move Off the Old One.
GitHub announced two things on August 11. MAI-Code-1.1-Flash is available in Copilot, at a list price 73% below the model it replaces. And MAI-Code-1-Flash will be removed from every Copilot surface on September 10.
Those are one announcement. A 73% price cut arriving with a four-week deprecation clock attached is a migration with a discount stapled to it, and the useful question is not whether the new model is cheaper per token. It is whether your bill goes down, which is a different question with a different answer.
What actually shipped
MAI-Code-1.1-Flash is Microsoft's in-house coding model, the successor to MAI-Code-1-Flash, which arrived for Copilot Business and Enterprise on June 26. The 1.1 release adds native vision for image understanding and, per GitHub's changelog, improves coding quality, instruction following, tool use and performance.
The 73% figure is a list price comparison against 1-Flash. GitHub frames it as making the model cost-effective for lightweight coding workflows that need a balance of capability and cost. That framing is doing work, and I want to pull on it.
The deprecation is the harder deadline. On September 10, 1-Flash goes away across Copilot. If you are on Enterprise, an administrator may need to enable 1.1-Flash through model policies in Copilot settings before anyone can select it, which is the kind of step that turns a non-event into a Monday morning outage for a team.
Why 73% cheaper does not mean 73% less
Here is the thing that took me too long to internalise about cheap models: per-token price and per-task cost diverge, and they diverge hardest exactly where cheap models are marketed hardest, which is agentic loops.
In a single-shot completion, a cheaper model is straightforwardly cheaper. You send tokens, you get tokens, you pay less per token, you pay less. Fine.
In an agent loop, the model makes a decision, calls a tool, reads the result, and decides again. A weaker model gets a step wrong more often. Each wrong step costs you the tokens for that step, the tokens for the tool result it misread, and the tokens for however many turns it takes to recover, all with your full context re-sent each time. The failures are not free and they are not linear. One bad tool call in a long-context loop can cost more than the twenty correct calls around it, because context grows and every subsequent turn pays for it.
So the honest version of the comparison is not "73% off." It is "73% off per token, multiplied by however many more tokens this model needs to finish the same job." If the cheaper model needs three times the turns, you saved nothing and spent more wall-clock time.
I am not claiming 1.1-Flash is worse in loops. GitHub says tool use improved, and it is a newer model, so it may well be better at exactly this. My claim is narrower: the 73% number tells you nothing about that, and the people who will be disappointed in September are the ones who read it as a bill forecast.
The three-week window and what breaks
Between now and September 10, the thing to check is where the string MAI-Code-1-Flash appears in anything you own.
Model names get pinned in more places than people expect. Editor settings and workspace config that select a default model. Scripts and CI jobs that call the API with an explicit model parameter. Any Copilot extension or custom agent config you wrote. And on Enterprise, the model policy list itself, which is the one that bites teams, because the model can be deprecated centrally while your policy still lists only the old one, leaving people with a dropdown that has nothing usable in it.
GitHub says no action is required to remove the old model once deprecated, which is true and is not the same as saying no action is required.
The broader lesson is that a pinned model name is now a maintenance liability with a half-life measured in months. 1-Flash shipped to Copilot Business and Enterprise on June 26 and is dead on September 10. That is under eleven weeks. If your architecture assumes a model name is stable, it is going to keep breaking, and the fix is the boring one: put the model name in config, not in code, and give yourself one place to change it.
Where a flash-tier model is genuinely the right call
I do not want this to read as anti-cheap-model, because the tier is real and useful.
Flash-tier models are the right choice for work where the task is well-specified and verification is cheap. Renaming variables across a file. Writing a first-draft test for a function whose contract you already wrote. Reformatting, mechanical refactors, filling in boilerplate you can eyeball in five seconds. Anything where you would notice a wrong answer instantly is a good candidate, because the cost of a mistake is one glance.
They are the wrong choice for work where you cannot cheaply tell whether the output is correct. Debugging something you do not understand yet. Multi-file changes where the failure shows up three files away. Anything where you would end up trusting the output because checking it is as much work as doing it. Those are the jobs where a stronger model earns a 10x per-token premium easily, because your time is the expensive input, not the tokens.
What I would actually do
Grep for the old model name today, not in the first week of September. That is a five-minute job now and an annoying one under deadline.
Then, before you switch anything permanently, measure the one number that matters, which is cost per completed task, not cost per million tokens. Take three real jobs from your actual backlog, run them through both models, and record what each cost end to end including retries. Three jobs is not a benchmark and I am not pretending it is. It is enough to tell you whether the 73% is showing up in your bill or evaporating into extra turns, which is the only thing you need to decide.
Where I could be wrong: I am treating list price as a weak signal, and for a lot of Copilot users it is not a signal at all, because they are on a subscription tier where the model selection does not change what they pay directly. If that is you, this entire cost analysis is irrelevant and the only thing that matters is whether 1.1-Flash does your work better than 1-Flash did. Which, since the old one disappears on September 10 regardless, you are going to find out either way. The grep is still worth doing.
Author
Lukas
@lukcombinator