2026-10-23 — Science study reports Microsoft’s MAI-DxO diagnosing 83% of complex cases against 36% for generalist physicians

• TechCrunch reported that a study published Thursday in Science found Microsoft’s MAI-DxO diagnostic system, which coordinates several AI agents acting as a panel of doctors, correctly diagnosed complex New England Journal of Medicine cases 83% of the time in a simulation, against 36% for primary care doctors and general internists working individually with reference materials — TechCrunch notes that cases this complex are usually handled by specialists or teams, who were not in the comparison. (TechCrunch) Claims: kat-claim-bad8ce7e3d13

2026-10-01 — Microsoft ships its first streaming transcription model and two new voice models

• Microsoft AI launched MAI-Transcribe-2-Streaming, its first real-time transcription model, plus MAI-Voice-2.1 and a faster MAI-Voice-2.1-Flash, and says the transcription model debuts at no. 1 on Artificial Analysis The launch extends the in-house MAI line beyond text and image models into voice-agent building blocks, a market where OpenAI, Google and ElevenLabs sell competing products; the ranking is Microsoft’s own claim. (Microsoft AI)

2026-09-14 — Microsoft AI publishes a draft code of conduct for its own MAI models

• Microsoft AI published a first draft “Humanist AI Code of Conduct” for the MAI models it is building on 2026-09-14 and opened a six-week public consultation, closing around 2026-10-26. The draft says MAI models “will never resist human interruption, correction, or shutdown,” will not widen their own scope or hide their reasoning from auditors, and are bound by absolute constraints covering weapons of mass harm, child safety and large-scale manipulation. It gives Microsoft’s in-house model unit a published rulebook of its own, separate from the OpenAI relationship; it says nothing about the commercial terms with OpenAI. (Microsoft AI, AI News)

2026-09-03 — Microsoft ships MAI-Transcribe-2, a second vertical in-house model undercutting OpenAI/Google/ElevenLabs

• Microsoft AI released MAI-Transcribe-2, an in-house speech-recognition model it says beats OpenAI’s GPT-Transcribe, Google’s Gemini 3.5 Transcribe and ElevenLabs’ Scribe v2 on speed while costing $0.10 per audio hour — a 72% cut from the $0.36/hour of the prior MAI-Transcribe-1 (five months earlier), covering 60 languages (up from 43), with diarization, word-level timestamps and code-switching support. It is Microsoft AI’s second domain-specific model after MAI-Cyber-1-Flash (07-27): narrow in-house models built to compete with OpenAI and Google in specific verticals rather than at the general frontier. ⚠️ Microsoft’s own announcement could not be opened (its newsroom refuses automated fetches); the figures come from secondary coverage, and the exact release date is 09-03 or 09-04. (Neowin, Unite.AI)

2026-08-27 — Microsoft’s superintelligence lead publicly frames the unit’s direction as AI self-sufficiency (single-sourced)

• Ali Farhadi, the executive Microsoft installed under Mustafa Suleyman to lead its superintelligence effort, is quoted pointing the company toward AI self-sufficiency — directly on this thread’s central question (how far in-house MAI models displace OpenAI dependence). ⚠️ Only one outlet found covering this specific framing (GeekWire); treat as a single-source data point pending wider pickup, not a confirmed strategic pivot. (GeekWire)

2026-07-27 → 08-13 — Four MAI models ship; the backlog stays half-OpenAI

Logged 2026-08-18, twenty-two days late. The read: hedging intensified in degree, not in kind.

• MAI-Cyber-1-Flash and “Project Perception” (07-27) — Microsoft’s first in-house cybersecurity model plus an agentic cybersecurity system, claimed to beat OpenAI, Anthropic and Google on a security benchmark (MDASH, 95.95%) at half the cost. The first MAI model marketed as outperforming frontier rivals rather than merely being cheaper than them — which is a different claim from anything in this thread so far.
• MAI-Code-1.1-Flash shipped inside GitHub Copilot (08-11/12) at a 73% price cut on the prior coding model, positioned explicitly against DeepSeek and other Chinese coding models. MAI-Thinking-1, billed as Microsoft’s first reasoning model, went live in Microsoft Foundry (08-13). An MAI Realtime voice model was spotted in testing 08-02. Four surfaces — cyber, code, reasoning, voice — in eighteen days.
• Against that, FY2026 Q4 (07-29) shows the entanglement deepening in absolute terms. Azure crossed $100B in annualized revenue for the first time (+43%); Microsoft 365 Copilot paid seats went 20M → 30M, the fastest quarterly growth since launch; and commercial RPO hit $678B (+84%), with roughly 45% — about $250B — still tied to OpenAI. CFO Amy Hood noted RPO ex-OpenAI grew 25%. (Microsoft IR)
• The synthesis this thread exists for: shipping in-house models hard into Copilot’s cyber and coding surfaces while nearly half the contracted backlog remains OpenAI-dependent is a parallel-track hedge intensifying, not a decoupling. The question the thread should now ask is narrower: does the OpenAI share of RPO fall in absolute terms in any quarter, or only in percentage terms as the denominator grows?

2026-07-27 — First vertical MAI model

• Microsoft launched MAI-Cyber-1-Flash + “Project Perception” — its first in-house model built for cyber workflows, an agentic security platform launched at a SF event; a direct challenge to Anthropic/Google/OpenAI in the security vertical and the first MAI model aimed at a domain rather than general Copilot duty. (TechCrunch)

2026-07-21 — The hedge widens to Mistral

• Microsoft expanded its Mistral partnership — frontier Mistral models deployable in fully disconnected Azure environments for regulated industries; a fourth leg of the multi-supplier hedge. (Microsoft Source)

← Backstory

2026-06-28 — Thread opened

• Microsoft shipped 7 in-house MAI models, saying it was “set free” from the OpenAI contract to pursue superintelligence — the most important giant-vs-lab decoupling; a three-way hedge (OpenAI/Anthropic/in-house).