On June 2 at Build 2026, a new option quietly appeared in my GitHub Copilot model picker inside VS Code: MAI-Code-1-Flash. It is part of a wave of seven new MAI models, and it matters for one reason beyond the benchmarks. These are the first Microsoft foundation models trained from the ground up without OpenAI technology, no distillation from third-party models, built on clean and traceable enterprise data.
Microsoft picked a specific target to measure against: Claude Haiku 4.5, the small, fast model a lot of us already lean on for everyday coding loops. So let me run the matchup the way Microsoft framed it, round by round, and call a winner each time.
First, the scoreboard Microsoft published, all run in the same production Copilot harness developers actually use.
| Benchmark | MAI-Code-1-Flash | Claude Haiku 4.5 |
|---|---|---|
| SWE-Bench Pro (pass rate) | 51.2% | 35.2% |
| IF Bench (instruction following) | +28.9 over Haiku | baseline |
| Advanced IF (rubric-based) | +14.5 over Haiku | baseline |
| Adversarial reasoning (adjusted) | 85.8% | lower overall |
| Tokens on SWE-Bench Verified | up to 60% fewer | baseline |
Round 1: Raw coding accuracy
The headline number is SWE-Bench Pro, a set of diverse real-world tasks rather than tidy textbook problems. MAI-Code-1-Flash lands at 51.2% versus 35.2% for Haiku, a 16-point gap. That is a wide margin for two models pitched at the same fast, cheap tier. Microsoft reports higher pass rates on all four of the SWE-Bench style evaluations it tested, not just one cherry-picked board.
Winner: MAI-Code-1-Flash. Sixteen points on real tasks is not noise.
Round 2: Token efficiency
Here is the part I care about most as someone who pays per token. MAI-Code-1-Flash was trained with what Microsoft calls adaptive solution length control. It stays terse on easy requests and only spends a bigger reasoning budget when a task genuinely needs it. The claim: solving harder problems with up to 60% fewer tokens on SWE-Bench Verified.
Fewer tokens is not just a billing line. It means lower latency, so an interactive edit loop feels snappier. If you want to reason about whether a swap is worth it, the math is simple:
type ModelRun = { passRate: number; tokensPerTask: number; pricePerMTok: number };
function returnOnToken(m: ModelRun): number {
// successful tasks per dollar spent
const dollarsPerTask = (m.tokensPerTask / 1_000_000) * m.pricePerMTok;
return m.passRate / dollarsPerTask;
}
const flash = { passRate: 0.512, tokensPerTask: 40_000, pricePerMTok: 0.5 };
const haiku = { passRate: 0.352, tokensPerTask: 100_000, pricePerMTok: 0.5 };
console.log(returnOnToken(flash) > returnOnToken(haiku)); // higher pass rate, fewer tokens
The token numbers above are illustrative, but the shape is the point: when accuracy goes up and tokens go down at once, return on token improves on both axes. That is rare.
Winner: MAI-Code-1-Flash, and not by a little.
Round 3: Following instructions
Coding agents live or die on whether they do exactly what you asked, especially across multi-turn sessions. Microsoft reports a +28.9 lead on IF Bench for precise instruction following and +14.5 on the rubric-based Advanced IF. Instruction adherence is also what makes agentic tool use reliable, since every wrong assumption compounds across steps.
Winner: MAI-Code-1-Flash, though Advanced IF is the closer of the two.
Round 4: Does it actually reason, or just remember?
This is where I appreciated Microsoft being honest. They built a 186-question, 34-category benchmark of adversarial traps: inverted classics like a flipped Monty Hall, impossible tasks, underdetermined prompts. The idea is to separate real reasoning from memorized answers. MAI-Code-1-Flash hit 85.8% adjusted accuracy and beat Haiku overall.
The caveat Microsoft published anyway: on Einstellung traps, where a familiar method blinds you to a simpler one, the model stayed below 50%. Both models still get anchored. So do humans.
Winner: MAI-Code-1-Flash, with an asterisk worth remembering.
Round 5: Where it lives
Architecture matters for who can run it. MAI-Code-1-Flash is a sparse Mixture-of-Experts model: 137 billion total parameters but only about 5 billion active per forward pass. That is how you get frontier-ish quality at small-model latency. It was trained directly against the GitHub Copilot harnesses used in production, so it learned the actual tools and systems it would operate inside, not a synthetic approximation.
Winner: MAI-Code-1-Flash for Copilot users, since it is tuned for exactly that environment. If you are not in VS Code Copilot, this round is a draw.
Here is the keynote where Satya Nadella and the MAI team walked through the in-house models. The Microsoft AI segment starts around the 1:47 mark.
The verdict
On Microsoft's own scoreboard, MAI-Code-1-Flash takes every round. I hold that loosely, because a vendor benchmarking against a competitor it chose is not a neutral referee. The harness is Microsoft's, the tasks are Microsoft's selection, and Haiku 4.5 is a deliberately small target. None of that is cheating, but it is context.
What I will trust is my own loop. The model is free inside Copilot right now, in the picker and routed through Auto, so the cost of checking is a few hours of real work. That is the test that counts: not the leaderboard, but whether my pull requests get smaller and my token bill gets lighter.
The more interesting signal is strategic. Microsoft just proved it can ship a competitive coding model with zero OpenAI DNA inside. Whatever happens with that partnership next, the company now has its own bench to play from. For those of us who build on top, more independent frontier labs means more leverage, and that is worth more than any single benchmark.