# MAI Code 2 vs Flash: Access, Pricing Claims and an Evaluation Plan

How to choose between Microsoft's new coding models, understand the promotional-price benchmark claims and run a useful repository-level evaluation.

By Paras Madan | 2026-11-18


MAI Code 2 and MAI Code 2 Flash give developers a new choice inside GitHub Copilot. The useful question is not which model has the larger headline score. It is which one completes your kind of change with the least total effort, including review and repair.

I would treat the launch as a reason to run a controlled comparison, not to replace an existing coding workflow immediately. Access, evaluation settings and promotional pricing all affect what the announcement means in practice.

## Code 2 and Flash have different jobs

Microsoft positions Code 2 for difficult debugging, multi-file implementation and repository-wide reasoning. Flash targets frequent, smaller tasks such as explanations, targeted edits and routine fixes. Both are available through GitHub Copilot; Flash also has a limited-access Foundry preview. Other distribution channels are described as coming soon in the [launch announcement](https://microsoft.ai/news/introducing-mai-code-2-and-mai-code-2-flash/).

That suggests a sensible starting split: try Flash on well-scoped changes and Code 2 on work where understanding relationships across files matters. This is an evaluation strategy, not a claim that the larger model always wins. A small task can contain a subtle dependency, and a large diff can be mechanically straightforward.

## The cost comparison includes a promotion

Microsoft reports Code 2 at 98% of GPT 6.1 Sol's average performance at 40% of its cost across four evaluations. The footnotes matter: MAI's comparison uses a 50% promotional price and maximum reasoning effort. Those percentages are not a permanent rate card or a promise about your Copilot bill.

Before making a budget decision, record the actual price and usage rules for the surface you use. Keep promotional and ordinary pricing separate in your spreadsheet. A benchmark cost ratio cannot tell you what happens when the promotion ends, your request mix changes or your team needs several attempts to finish a task.

## Large context is a limit, not a guarantee

The [model card](https://microsoft.ai/pdf/MAI-Code-2-And-MAI-Code-2-Flash-Model-Card.pdf?v=3) specifies text and image input, a 922,000-token input limit, 128,000-token output limit and 1,050,000-token combined context. It also warns that quality can degrade near the maximum context length. The models are hosted offerings, not publicly released weights.

For a repository evaluation, start with the files needed for the change rather than filling the available window. Compare that with a broader-context run only when the additional material has a clear purpose. Otherwise, you risk paying to add distraction without learning whether the model can find and use the important dependency.

## Read the benchmark setup before the ranking

The card describes coding evaluations using GitHub Copilot CLI without subagents, maximum reasoning and means across three runs. These are single-attempt results, not the success rate after picking the best of several answers. It also notes that longer sessions can drift or stop before completing verification.

Your agent may use a different harness, tool set or retry policy. Keep those variables fixed when comparing models. If one gets more attempts or a richer environment, the result describes two workflows rather than two interchangeable model choices. That can still be useful, but label it honestly.

## Build a small test from completed work

Choose historical tasks with known outcomes: a localized bug, a cross-file behavior change and a test failure requiring diagnosis. Remove the final patch from the starting state. Give each model the same request, available tools and stopping rules.

Score more than whether the tests pass. Record unintended changes, missed requirements, reviewer corrections and time to an acceptable patch. Include unsuccessful runs in the cost calculation. A model that writes an impressive first diff but needs repeated repair may be a poor default for routine work.

## Decide where each model belongs

Use the results to define routing rules you can explain. Flash might become the default for bounded edits, with escalation when an attempt fails a specific check. Code 2 might earn its place on dependency-heavy work. Neither outcome needs to be universal.

Keep human review and normal release gates. The useful win is a repeatable improvement in completed work, not a new model name in the editor.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams explain technical products through clear, practical launch stories.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
