Post claims Microsoft trained MAI-Thinking-1

Microsoft disclosed key technical details of MAI-Thinking-1 on June 2, but the viral social post that followed a day later added claims the company has not publicly backed. In a post on X dated June 3, user aamirpatni said Microsoft had trained a 35 billion-parameter reasoning model with a 256,000-token context window, at one-tenth the training cost of GPT-5.5 and with no OpenAI data. Microsoft’s own materials support parts of that description, including the model name, parameter count and context length, but not the cost comparison. Microsoft’s June 2 announcement described MAI-Thinking-1 as the company’s reasoning model and said it was trained “from the ground up” on “enterprise grade, clean and commercially licensed data,” without distillation from third-party models. The company also said AI-generated content was excluded from pre-training. Those statements line up with the social post’s broader claim that the model was built without OpenAI-derived training data, though Microsoft’s wording was narrower and specific to distillation and data provenance. (microsoft.ai) ### Which parts of the social post are confirmed by Microsoft? Microsoft’s June 2 product post said MAI-Thinking-1 supports a 256,000-token context window and is a “35B-active” sparse mixture-of-experts model with roughly 1 trillion total parameters. In its Build keynote transcript, Microsoft separately described the model as a “35B active parameter MoE with a 256K context window.” Microsoft also said the model is being integrated into Copilot, Teams, GitHub and Dynamics 365 Contact Centre, and is available in Foundry. (microsoft.ai) The company positioned it as a medium-sized model aimed at software engineering, math and enterprise deployment. ### Where does the “zero OpenAI data” claim come from? Microsoft’s June 2 blog post said MAI-Thinking-1 was trained without distillation from third-party models and on licensed data with AI-generated content excluded from pre-training. (microsoft.ai) That is the clearest public statement the company has made on how the model was built. The phrase “zero OpenAI data” appears to be a social-media paraphrase, not wording Microsoft used in the materials reviewed. (microsoft.ai) Microsoft did not, in the June 2 post, publish a line-by-line accounting of every dataset or explicitly frame the disclosure as a comparison with OpenAI. ### What about the claim it cost 10 times less than GPT-5.5 to train? OpenAI’s April 23 announcement for GPT-5.5 described the model’s capabilities, rollout and benchmark results, but did not disclose training cost. (microsoft.ai) Microsoft’s June 2 materials on MAI-Thinking-1 likewise did not publish a training-cost figure or a direct cost comparison with GPT-5.5. That means the “10x cheaper” claim cannot be verified from the companies’ primary public releases reviewed here. (microsoft.ai) Without a Microsoft filing, blog post, keynote remark or OpenAI cost disclosure to compare against, the number remains an unsupported assertion from the June 3 social post. ### Did Microsoft present MAI-Thinking-1 as a direct GPT-5.5 challenger? Microsoft’s June 2 post said MAI-Thinking-1 “matches leading models on key software engineering benchmarks” and is preferred to “Sonnet 4.6” in the company’s blind human side-by-side evaluations. (openai.com) In the same post, Microsoft said the model is “toe-to-toe with Claude Opus 4.6 on SWE-Bench Pro.” OpenAI’s April 23 GPT-5.5 release presented GPT-5.5 as the company’s newest frontier model and published benchmark tables against earlier OpenAI models and some outside systems. (microsoft.ai) But neither company, in the materials reviewed, published a direct MAI-Thinking-1-versus-GPT-5.5 training-cost comparison. ### What can be said with confidence right now? June 2 is the key date for the underlying product release: that is when Microsoft published its MAI-Thinking-1 announcement and Build keynote transcript. (microsoft.ai) June 3 is the date the viral X post reframed that launch into a broader claim about cost, independence from OpenAI and competitive positioning. The next verifiable step is straightforward: any confirmation of the “10x cheaper” figure would need to come from Microsoft in a technical paper, model card, executive interview or filing, or from a comparable OpenAI disclosure on GPT-5.5 training cost. (openai.com) As of June 4, the public primary sources reviewed support the model specs and the no-distillation claim, but not the cost claim. (microsoft.ai)