GitHub Copilot Has 280 Million Users. Microsoft Just Started Paying Itself for Some of Them.
Microsoft launched seven in-house MAI models at Build 2026, putting MAI-Code-1-Flash into the GitHub Copilot auto picker. When Copilot routes a task to MAI instead of GPT-4o, Microsoft avoids an OpenAI API bill. Combined with the April 27 renegotiation that ended OpenAI's revenue share, this is a two-quarter execution of a plan to reduce a $13 billion dependency. The model benchmarks are competitive but not frontier. The financial architecture underneath them is the real story.
On April 27, 2026, Microsoft and OpenAI renegotiated their partnership. The exclusive license became non-exclusive. Microsoft stopped paying OpenAI a revenue share. OpenAI was freed to sell to AWS, Google Cloud, and Oracle. The arrangement that had defined the AI industry since 2019 was restructured into something more like a vendor relationship with history.
Five weeks later, at Build 2026, Microsoft launched seven in-house AI models under the MAI brand. One of them — MAI-Code-1-Flash, a 5-billion-parameter coding model — is rolling out in the GitHub Copilot model auto picker.
When the auto picker routes a Copilot task to MAI-Code-1-Flash instead of GPT-4o, Microsoft pays itself instead of OpenAI. GitHub Copilot has 280 million enrolled developers. The number of daily completions is not publicly disclosed. The dollar value of routing a fraction of those tasks in-house has not been quantified anywhere I can find in the coverage of this launch.
That number is the story.
What MAI-Code-1-Flash actually is
Five billion active parameters. 51.2% on SWE-Bench Pro, which is a 16-point lead over Claude Haiku 4.5 and represents competitive mid-tier coding performance. Microsoft claims 60% fewer tokens on equivalent tasks compared to alternatives — a cost efficiency claim that, if it holds, makes the routing decision even more favorable on a per-task basis.
The model is available on Fireworks AI, Baseten, and OpenRouter — three platforms that serve developers who have specifically decided not to be locked into a hyperscaler's model. This is a deliberate signal: Microsoft is not building a walled garden with its own models. It is competing in the open market.
What MAI-Thinking-1 actually is
MAI-Thinking-1 is the headline model: a sparse Mixture of Experts with 35 billion active parameters and approximately 1 trillion total parameters, with 97.0% on AIME 2025 (a math competition benchmark) and 52.8% on SWE-Bench Pro. Microsoft positions it as competitive with Claude Sonnet 4.6 — the mid-tier Anthropic model, not Opus 4.8 — based on its own blind human evaluation across 1,276 tasks.
The AIME 2025 number has a complication. BenchLM.ai's independent benchmark aggregation shows Moonshot AI's Kimi K2.5 Reasoning at 96.1% on AIME 2025, placing it ahead of MAI-Thinking-1 on the same measure. Microsoft's 97.0% is self-reported. The discrepancy is unresolved. Whether MAI-Thinking-1 actually holds the AIME 2025 record or not matters for the "we built a frontier reasoning model" narrative; it does not materially affect the Copilot routing economics, which are driven by MAI-Code-1-Flash.
The honest benchmark read: MAI-Thinking-1 at 52.8% SWE-Bench Pro is a competitive mid-tier model. GPT-5.5 is at 58.6%. Claude Opus 4.8 is at 69.2%. This is not a frontier model. The AIME score is impressive on a math benchmark. The coding and reasoning scores position MAI-Thinking-1 alongside GPT-4o tier performance, not GPT-5.5 tier.
Microsoft says it is preferred to Claude Sonnet 4.6 in human evaluations on its own testing platform. I would trust that directionally. I would not treat it as a settled competitive claim without independent replication.
The "no distillation" wedge
Microsoft's most legally and commercially significant claim: "We trained it from the ground up on clean, traceable, and enterprise-grade data, without distillation from third-party models."
This is aimed directly at regulated industries — finance, healthcare, legal — where AI procurement teams have been asking, for two years, what data their AI vendors used to train the models they are integrating into client workflows. Neither OpenAI nor Anthropic can make an equivalent claim with the same confidence. OpenAI's training data has been the subject of multiple copyright litigations. Anthropic's data provenance claims are cleaner but not accompanied by the same "commercially licensed, traceable" framing.
The claim is not externally verifiable. Microsoft is asking the market to trust it. For developers and researchers, benchmarks matter more than data claims. For banks, hospitals, and law firms deploying AI in regulated workflows, "traceable data" is a compliance checkbox. Microsoft is the only major AI provider that can credibly check that box at scale with its own models.
The two-quarter plan
The sequencing is what makes this legible as a strategy rather than a product launch.
April 27: Renegotiate commercial terms. End exclusivity. End revenue share. Retain non-exclusive license through 2032. Give OpenAI the multi-cloud freedom it needed to survive independently. This was not a breakup — it was a controlled deescalation that removed the structural dependencies on both sides.
June 2: Launch seven in-house models. Put the coding model into the product that serves 280 million developers. Make it clear, publicly, that the models are trained without OpenAI data.
The April move created the commercial conditions for independence. The June move is the product execution. The timing is not coincidental — five weeks is roughly the PR gap between "we restructured the deal" and "here is why we restructured the deal."
What this means for OpenAI
OpenAI's most important distribution relationship was GitHub Copilot. GitHub had the developer base; OpenAI had the model; Microsoft owned both. Now Microsoft is using that distribution for its own models.
The revenue share that ended in April was not disclosed in dollar terms. Satya Nadella described the original OpenAI relationship as central to Microsoft's AI strategy. Analysts estimated Microsoft's OpenAI-related infrastructure and investment commitments at over $13 billion. The revenue share was a portion of that flowing back to OpenAI.
OpenAI filed an S-1 for its IPO on June 8. The S-1 covers the period before the MAI launch and the April renegotiation was disclosed publicly. But the investor narrative for an OpenAI IPO has to include: its largest equity partner is now a model competitor with 280 million developers as a captive distribution channel. That is a different risk disclosure than the pre-April version of the OpenAI story.
I don't think Microsoft is trying to destroy its relationship with OpenAI. The two companies are too intertwined through Azure, the non-exclusive license, and shared enterprise customers to go to war cleanly. What Microsoft is doing is something more precise: ensuring that OpenAI cannot hold the Copilot distribution hostage in any future renegotiation, because Microsoft now has its own model that is competitive for that traffic.
That is what a $13 billion dependency looks like when you decide to start reducing it.