$0.75 per million input tokens. That price, buried in GitHub's documentation, is the real story of Microsoft's Build 2026 keynote on June 2, 2026. Microsoft showed seven in-house "MAI" models that day, but only one is already writing code in front of developers: MAI-Code-1-Flash, now rolling out inside GitHub Copilot in Visual Studio Code. It is Microsoft's first coding model built without OpenAI's technology, and the company is pitching it as a step toward "long-term independence" from the labs it also bankrolls.
The interesting question isn't whether it beats GPT-5.5. It's whether a cheap model shipped as a Copilot default quietly reroutes millions of coding requests away from OpenAI. That is the move underneath the benchmark slides.
❝Want the deeper analysis? See our companion guide: Fine-Tuned Mid-Tier vs Frontier: When a Smaller Model Wins (And Costs 10x Less).
What Happened
Microsoft announced seven MAI models at Build: MAI-Thinking-1 (reasoning), MAI-Code-1-Flash (coding), MAI-Image-2.5 and MAI-Image-2.5 Flash (image), MAI Transcribe 1.5 (transcription across 43 languages), and MAI-Voice-2 plus MAI-Voice-2 Flash (speech). All are described as the company's first in-house foundation models trained without OpenAI technology.
MAI-Code-1-Flash takes a written description and produces application and website source code. It began rolling out on June 2 to a fraction of individual GitHub Copilot users in VS Code, across the Free, Pro, Pro+, and Max tiers, with no additional setup required. CLI support and a standalone API are listed as later rollouts.
One detail Microsoft has not cleanly resolved: the model's size. The published model card lists 137 billion parameters with a sparse Mixture-of-Experts (MoE) architecture, while the family announcement describes it as a 5-billion-parameter model. Microsoft has not said whether those figures refer to active parameters, total parameters, or different variants — so treat the parameter count as unconfirmed for now.
On benchmarks, Microsoft positions MAI-Code-1-Flash against Claude Haiku 4.5. It reports 51.2% on SWE-Bench Pro versus Haiku's 35.2% (a 16-point lead), and 71.6% on SWE-Bench Verified versus 66.6% — while, Microsoft says, solving harder tasks with up to 60% fewer tokens. On instruction-following benchmarks the company claims wider margins still (a +28.9-point lead on IF Bench). These are Microsoft's own published figures; independent third-party benchmarking was not available at launch.
Why It Matters
The strategic backdrop is the reworked Microsoft–OpenAI relationship. In April 2026 the two companies amended their partnership: Microsoft's exclusive license to OpenAI's intellectual property ended, Microsoft's revenue-share obligation to OpenAI was removed, and OpenAI's capped revenue share to Microsoft was preserved through 2030. That amendment is what makes proprietary frontier models legally possible for Microsoft — and MAI-Code-1-Flash is the first coding product to show up on the other side of it.
The developer-facing argument is cost. By running its own models on Azure rather than licensing them, Microsoft avoids paying royalties to partners like OpenAI, and it says those savings flow to developers. The pricing now in GitHub's documentation — $0.75 per million input tokens and $4.50 per million output tokens (with cached input at $0.075) — undercuts frontier models substantially, though the model card still marks pricing "to be finalized."
The most eye-catching number came from Mustafa Suleyman, CEO of Microsoft AI, who said on LinkedIn that an MAI model tuned to consulting firm McKinsey's tasks outperformed GPT-5.5 on quality "while being 10x lower on cost." That claim is Microsoft's, it is about a custom-tuned MAI model rather than the base MAI-Code-1-Flash, and it has not been independently verified — but it signals the company's broader bet: that domain-tuned mid-tier models can beat general-purpose frontier models on specific workloads.
Competitive Analysis
MAI-Code-1-Flash enters a crowded coding-model field, and Microsoft chose its comparison carefully. By benchmarking against Claude Haiku 4.5 — Anthropic's small, fast tier — rather than a frontier model like Claude Opus 4.8 or GPT-5.5, Microsoft is staking out the cheap-and-fast segment, where price-per-token and latency matter more than raw ceiling capability. On price, the contrast is stark: at $0.75/$4.50 per million tokens, MAI-Code-1-Flash sits well below Claude Opus 4.8's $5/$25, and below premium GPT tiers. For Anthropic, the immediate pressure lands on Haiku, whose value proposition has been exactly that low-cost, high-throughput coding niche. For OpenAI, the threat is structural rather than benchmark-by-benchmark: its single largest distribution channel, GitHub Copilot, now ships a default-eligible Microsoft alternative, and every developer routed to MAI is a developer not generating OpenAI royalties. The open question is durability — Microsoft's numbers are first-party, the model is still reaching only a fraction of users, and there is no public API yet for outside teams to validate the claims at scale.
What To Watch
Watch for the first independent benchmarks once the model reaches broader availability, clarification on the 5B-versus-137B parameter question, and whether the standalone API ships with the pricing GitHub currently lists. The bigger signal is adoption inside Copilot: if Microsoft makes MAI-Code-1-Flash the default for a meaningful share of coding requests, that is the clearest measure yet of how serious its move away from OpenAI dependence really is.
Sources
Written by the vybecoding.ai editorial team
Published on June 9, 2026