https://www.ft.com/content/f1ec830c-2f08-4b1a-b70f-7330f260753c

Microsoft has announced their intention to train their own frontier LLM / foundation model, independent of OpenAI.

• Mustafa Suleyman, CEO of Microsoft AI, stated that Microsoft is now prioritizing the development of its own in-house "frontier foundation models." He described this as a move toward "true AI self-sufficiency" to reduce the company's near-total dependence on OpenAI for its flagship products.

• The reports highlight a model lineup referred to as MAI (including the MAI-1 preview mentioned in earlier contexts). These models are being trained on massive "gigawatt-scale compute" clusters and are expected to be released "sometime this year" (2026).
  

Market resolves to YES if Microsoft succeeds in training and releasing their own frontier-level language model before the end of 2026. Release = make available for use by the public, e.g. in a web app or via API.

What counts as a frontier language model?

The most ambiguous part of this question is what counts as a frontier model. This will be mostly based on vibes. As guidance, I would say as of February 2026, the following companies definitely have released frontier or frontier-competitive models at some point, for the purposes of this market:

• OpenAI

• Anthropic

• Google / Deepmind

• Deepseek (R1 at the time of its release)

• X.AI
  

In addition, I think there is a reasonable case to be made that the following companies have released frontier-competitive models:

• Meta (Llama 4 doesn't count in my mind due to the cheating shenanigans, but Llama 3 and its successors were arguably close enough to frontier at time of release to count. The frontier was very thin back then. We'll see if Avocado is Truly Frontier)

• Zhipu (Z.AI) - the GLM 4.7/5 seem competitive with recent versions of Opus on benchmarks and recent versions of Sonnet for real-world use.

• Moonshot - Kimi 2 / 2-thinking / 2.5 were the leading open weights models at certain points.

I would consider a model to be frontier if one of the following is true:
1. It's closed source and competitive with or better than the leading open source model at the time of its release.
2. It's closed source and competitive with or better than the latest Opus or Sonnet model from Anthropic, the latest Pro or Flash model from GDM, or the latest flagship thinking model from OpenAI at the time of its release.
3. It's open weights and competitive with or better than the third-best open weights model (not counting itself) at the time of its release. (open models can have a little affirmative action, as a treat)

The model must be trained from scratch by Microsoft, without involvement from OpenAI or other incumbent frontier labs. A finetune of a base model or checkpoint trained by another company doesn't count.