Every figure, date and claim the finished picture puts on screen, chased to a primary source. Where a number appears in a shot, the shot is named beside it.
MAI-Code-1-Flash is a coding model built by Microsoft AI and shipped into GitHub
Copilot on 2 June 2026.
Shots: 004-mai-code-1-flash, 005-not-just-a-picker, 078-treat-it-like-a-signal.
It was trained without distillation from other companies’ models. Microsoft’s wording is that it was “trained from the ground up on clean, traceable and enterprise-grade data, without distillation from third-party models”. This is the basis for the video’s claim that it is a genuinely first-party model rather than a rebadged one.
It is offered in the model picker and under the default auto picker.
Shots: 049-will-it-stop-external, 056-the-auto-picker, 081-let-microsoft-decide.
This is the single most load-bearing source in the video. The argument that a default can route work to a model nobody explicitly chose is not an inference here; Microsoft states that the model is available “in model picker and under default auto picker”.
Rollout by plan. At launch it began rolling out to Copilot Free, Student, Pro, Pro+ and Max, starting with a limited set of users in Visual Studio Code and expanding gradually. Copilot Business and Copilot Enterprise reached general availability on 26 June 2026, billed at provider list pricing under usage-based billing, and administrators must enable the MAI-Code-1-Flash policy before their users can select it.
Surfaces. By 18 June 2026 it was available on Copilot CLI, the Copilot cloud agent, the
GitHub Copilot app, Copilot Chat on GitHub, Visual Studio, GitHub Mobile, JetBrains IDEs,
Eclipse and Xcode, in addition to Visual Studio Code.
Shot: 024-surface-integration.
SWE-Bench Pro: 51.2% against Claude Haiku 4.5’s 35.2%.
Shot: 010-stay-to-the-end. These are the only two benchmark numbers the video renders,
and they are drawn as bars with the figures counting up beside them.
Token efficiency: up to 60% fewer tokens on SWE-Bench Verified, described by Microsoft
as “solving harder problems with up to 60% fewer tokens”. This is what the video’s cost
argument rests on when it says a first-party model can be tuned for the product’s own
usage pattern.
Shot: 032-route-cheap-reserve-expensive (as the reason the cheap route is viable).
Microsoft additionally reports higher pass rates on all four coding evaluations it ran (SWE-Bench Pro, SWE-Bench Verified, SWE-Bench Multilingual, Terminal Bench 2), and margins of +28.9 points on IF Bench and +14.5 points on Advanced IF for instruction following.
All of these are Microsoft’s own published evaluations of its own model against one named competitor. They are cited here as what Microsoft claims, which is what the video treats them as, and they are not independently reproduced.
Shots: 014-owns-everything, 044-becoming-the-stack, 047-the-boring-superpower.
microsoft/vscode.
https://github.com/microsoft/vscodemicrosoft/TypeScript. https://github.com/microsoft/TypeScriptShots: 023-not-a-model-wrapper, 024-surface-integration, 025-surface-controls.
The video lists editor integration, repository context, code review, agents, organization controls, usage metrics, billing and enterprise deployment. Each of these is a documented GitHub Copilot feature rather than a characterisation.
Shots: 039-the-third-reason, 040-if-they-change.
The video’s claim is narrow: that the two companies remain deeply linked, and that Microsoft would prefer its most important AI products not to depend entirely on another company’s roadmap. Microsoft and OpenAI announced a restructured partnership in October 2025, which Microsoft describes as continuing while giving each side more independence.
Shots: 045-cursor-and-claude-code, 046-openai-and-google.
The video characterises what each competes on rather than stating figures about any of them, and no number is rendered for any competitor.