Two claims, two kinds of evidence
Microsoft published one claim and reportedly made another. Keeping them apart is the whole story.
The published claim is about training efficiency against open-weight models. Microsoft says MAI-Base-2 beats every open-weight model it tested on its held-out evaluations, and that its training recipe is far cheaper per unit of performance than DeepSeek's. The method is spelled out: the team fitted scaling laws on smaller models, estimated how much compute its recipe would need to match DeepSeek-V4-Pro-Base's scores, and compared that estimate with DeepSeek's actual training compute. The result is a range, 4.2 to 12.3 times, depending on which evaluation is used. That figure is a modeled estimate from scaling-law fits — not a direct comparison of two finished training runs. The same applies to the 6.3–12.7x improvement Microsoft claims over its own MAI-Base-1.
The reported claim is larger and rests on nothing public. According to Business Insider's account of a staff town hall on 2026-10-15, Suleyman said Microsoft now has a stronger pre-trained base model than the one behind GPT-5.6, built with a quarter of the compute OpenAI used. Business Insider's report distinguishes the GPT-5.6 product label from an internal nickname for that model — neither name is independently confirmed as the compared model's settled identity. He said the comparison used technical information Microsoft holds about OpenAI's models. No scores, evaluation names, or margins were published.
Read the public numbers closely
The public case is unusually detailed for a lab note, and it still carries the usual discounts.
The evaluations are Microsoft's own. Held-out text was measured in bits-per-byte, which reduces tokenizer sensitivity when comparing models with different tokenizers; the code test used Microsoft's internal codebase; the STEM, maths, and Q&A sets were commissioned from vendors and decontaminated against Microsoft's training data. That hygiene deserves credit. It does not make the comparison independent: the evaluations remain Microsoft's own and cannot be rerun by outsiders.
The efficiency figure is an estimate, not a measurement of two finished runs. Microsoft fitted its own scaling curves, with MAI-Base-2's final result included in the fit, and read off how much compute its recipe would need at DeepSeek's score level.
DeepSeek-V4-Pro-Base was nevertheless trained on substantially fewer total FLOPs than MAI-Base-2. A recipe can be more efficient per point of score while the finished model still costs more to train, because Microsoft spent the saving on a larger model. Microsoft states this itself in its own post.
Microsoft also reports 21% average model FLOPs utilization and 89% goodput for the training run, with its MFU calculation excluding certain rematerialization and optimizer FLOPs.
The comparison set also mixes categories. Three of the four rival models named are available only as post-trained checkpoints, which Microsoft flags as its own caveat, because post-training can shift pre-training-style scores in either direction.
The part nobody can check
A base model is not a product. Bits-per-byte on held-out text says something real about pre-training quality, and very little about how the finished, post-trained model will behave in Copilot or in a customer's hands. Pre-training quality is not instruction-following, agentic performance or safety behaviour. Microsoft has not published a single finished-model benchmark for MAI-Base-2, because that model does not exist yet: pre-training is complete, the model is in post-training, and there are no weights, no API endpoint and no release date.
The town-hall comparison adds a second problem. Microsoft is OpenAI's largest investor and a privileged partner, so it plausibly does hold technical detail about OpenAI's base models that no other lab has. That makes the claim harder to dismiss and, on the current public record, impossible to verify. Microsoft could publish a sufficiently detailed comparison later; it has not done so. A comparison run privately, on a partner's undisclosed model, reported second-hand from an internal meeting, is not evidence a reader, a customer, or a competitor can weigh today. It stands as a reported internal remark — a signal about how Microsoft wants its staff, and now the market, to see its position relative to OpenAI.
What to watch
Three things will turn this from a claim into a fact pattern. First, the post-trained model's public benchmarks, ideally run by someone other than Microsoft. Second, whether the Foundry licensing Microsoft describes actually ships weights to trusted partners, and on what terms — licensing through Foundry is a stated plan, not a shipped product. Third, whether the efficiency story survives contact with inference costs, which is where a 2.8-trillion-parameter model has to earn its keep. Preparations for MAI-Base-3 are already under way, Microsoft says. The race it is describing is about cost per unit of capability. On the public record, it has shown a cheaper recipe. It has not yet shown a better model anyone else can test.
Sources
- [1] Microsoft AI (primary), "Frontier Pre-training at Microsoft AI", Lab Notes, 15 October 2026 — model specs, training run, evaluation method, efficiency estimates, MFU and goodput figures, Foundry licensing plan, post-training statusRead source
- [2] Business Insider (Ashley Stewart), Oct 15, 2026 — reporting of Mustafa Suleyman's staff town-hall remarks on the GPT-5.6 base-model comparison and quarter-compute claim; no scores published (link via Biztoc syndication record of the Business Insider story)Read source