Microsoft Launches Seven In-House MAI Models, Loosening Its Reliance on OpenAI
At Build 2026, Microsoft unveils the MAI family, trained 'from scratch' without distillation: an in-house alternative to OpenAI and Anthropic. Its headline numbers, however, remain vendor-claimed benchmarks, not yet reproduced by third parties.
At Build 2026 (June 2), Microsoft unveiled seven proprietary models in the MAI family, trained — the company says — "from the ground up on clean data, without distillation from third-party models". The flagship is MAI-Thinking-1, the company's first reasoning model; the lineup is rounded out by MAI-Code-1-Flash (5 billion active parameters, already integrated into GitHub Copilot and VS Code), MAI-Image-2.5 with its Flash variant, MAI-Transcribe-1.5 and MAI-Voice-2 (plus a Flash version on the way).
On MAI-Thinking-1's numbers the sources diverge and should be read with caution. The official page calls it only "medium-sized" and claims it is "preferred to Sonnet 4.6" in blind human evaluations conducted by Surge. Independent coverage adds the specs Microsoft has stated: a sparse Mixture-of-Experts architecture with ~35 billion active parameters out of roughly one trillion total, and a 256,000-token context window. On top of these comes a claim of parity with Claude Opus 4.6 on SWE-Bench Pro. These, however, are vendor-claimed benchmarks: the model is in private preview on Foundry, and "full reproduction of the results by independent labs has not yet occurred".
The move has a strategic center of gravity. "We believe the time has come for every company to move from consuming a frontier model to fully participating at the frontier", said Satya Nadella. Microsoft — which has committed 13 billion dollars to OpenAI and up to 5 billion to Anthropic — is not closing those alliances, but it is equipping itself with an in-house alternative that reduces dependence and shifts its contractual leverage.
Why it matters
- Entrepreneurs: A longtime OpenAI partner is building itself an in-house alternative: anyone negotiating AI contracts should reassess lock-in, pricing leverage and multi-vendor options over the medium term, because the supply chain for enterprise models is becoming more contestable.
- LLM builders / devs: Tunable weights, MAI-Code-1-Flash already inside GitHub Copilot and the promise of lower inference costs broaden your stack options; but the headline benchmarks are vendor-claimed and not yet reproduced by third parties, so validate them on your own workloads before adopting them.
ElevenLabs Unveils Dubbing v2, the AI Dubbing That Molds Itself to the Original Performance
ElevenLabs' new model doesn't just translate the words: it conditions on the speaker's performance to preserve tone, rhythm, and emotion across more than 90 languages.
On May 28, 2026, ElevenLabs announced Dubbing v2, an AI dubbing model that, unlike traditional systems based on the transcript alone, conditions directly on the original vocal performance: according to the company it preserves "tone, rhythm, delivery, and emotional intent," carrying the speaker's intonation from one language to another instead of reconstructing the voice from the text (ElevenLabs, Introducing Dubbing v2). The system supports more than 90 languages and introduces "sync-aware translation": it adapts the translation to a natural spoken delivery, automatically aligns timing, pauses, and rhythm to the original, and reduces manual adjustments.
The model is available in ElevenCreative and ElevenProductions. At launch, for the 7 days following May 28, 2026, ElevenLabs offered a free trial of 1 minute on the Free plan, 15 minutes on Starter, and 30 minutes on Creator+ plans; anyone evaluating the tool today should verify current availability and limits on the product/pricing page. API access is announced as imminent. The official page was updated on June 11, 2026.
Independent coverage tempers the enthusiasm with some material caveats. An analysis on audiobook localization notes that speed does not remove every constraint. The open question remains of consistent quality on long-form content without causing human review to balloon; moreover, the tool handles only the audio, not metadata or storefront copy. Finally, "localization at scale changes the compliance surface" and demands clarity on licensing and consent for voice use (Automateed). On the consent front, ElevenLabs' terms require uploading only voices to which you hold the rights, while the Tennessee ELVIS Act frames the voice as part of one's identity to be protected (margabagus, consent and voice cloning). The EU AI Act in turn provides for marking and disclosure obligations for synthetic content and audio deepfakes, but with general application from August 2, 2026: as of June 13, 2026 it is not yet a general obligation in force (Regulation (EU) 2024/1689, AI Act).
Why it matters
- End users: For those who consume content, Dubbing v2 promises to let you watch videos, courses, and audiobooks in your own language while keeping the real voice and emotion of the original, breaking down the language barrier without the detachment of traditional dubbing. The same mechanism, however, raises the stakes on consent and voice cloning: the ease of replicating a voice makes it central to understand who authorized what and how to recognize synthetic audio, a need that in the EU the AI Act's labeling obligation will make binding from August 2, 2026.
Anthropic Launches Mythos-Class Fable 5, Then Suspends Access After Three Days Under a US Export-Control Directive
A frontier model available in production, pulled by government order within ~72 hours of launch: the case study of the year on service continuity, regulatory risk, and vendor dependence.
On June 9, 2026, Anthropic made Fable 5 publicly available, a Mythos-class model — the most capable family, above Opus 4.8 — alongside Mythos 5, reserved for authorized partners. The material news is not so much the model as its controlled packaging: Fable 5 is Mythos with safeguards active, while Mythos 5 is the same underlying model but with certain safeguards removed or loosened in specific areas, particularly cyber for Glasswing partners. AI classifiers intercept high-risk requests (cybersecurity, biology/chemistry, distillation) and automatically reroute them to Claude Opus 4.8; the user is notified when the fallback kicks in. According to Anthropic, more than 95% of sessions trigger no fallback at all. Price: $10 per million input tokens and $50 per million output tokens — "twice that of Opus 4.8" according to TechCrunch. Binding condition: mandatory 30-day retention on all Mythos-class traffic, even for zero-retention contracts.
Three days later, on June 12 at 5:21pm ET, the US government issued an export-control directive on national-security grounds, and Anthropic suspended access to both models. The directive required blocking access for "any foreign national, inside or outside the US, including foreign Anthropic employees"; to ensure compliance, Anthropic states it disabled the models for all customers, across every channel. The letter, again according to Anthropic, provided no specific details. The company says it understands the government's concern to involve a narrow rather than universal jailbreak — asking the model to read a codebase and fix its flaws. And it openly disagrees: similar capabilities are already available from other models, and a recall on this basis "would effectively halt every new deployment." Independent voices such as Aikido push back on the "AI hacker" narrative: vulnerability discovery is reportedly about 20% of the real problem. And demand for 0-days does not automatically rise with greater vulnerability discovery. Anthropic considers the order a misunderstanding, states it is working to restore access as soon as possible, and intends to share more technical details within 24 hours.
Why it matters
- ICT engineers / IT managers · LLM builders / devs: A production model can vanish in 72 hours by government order, without notice and for all users: service continuity depends on a regulatory risk outside your control. Anyone building on top of it must design for single-vendor lock the way they would for an availability incident, with fallbacks to alternative models already in place.
- LLM builders / devs: Two architectural details weigh on design: the classifiers reroute high-risk requests to Opus 4.8 — Anthropic says the user is informed when the fallback occurs — changing the model's expected behavior all the same; and the mandatory 30-day retention applies even to zero-retention contracts. Moreover, the precedent of a recall over a "narrow" jailbreak signals that compliance risk can hit any frontier deployment.