Microsoft MAI-Base-2: Pricing and API Availability

Microsoft’s frontier pre-training announcement is a training-economics story, not a customer price cut. Its October 15 technical report introduces MAI-Base-2 as the foundation for future models. It is still undergoing post-training, not launching as a generally available assistant today.

No public MAI-Base-2 input/output token price or exact-model endpoint is verified. Microsoft’s model pricing page and public Foundry catalog do not list a MAI-Base-2 rate or callable route in our checks. Unknown does not mean free. Our Microsoft AI pricing page reads a daily-maintained structured availability record; existing priced rows in the canonical API feed remain unchanged.

What changed: Microsoft’s next frontier foundation

MAI-Base-2 is a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active parameters. Microsoft reports training it on 16,000 GB200 GPUs for 38 days during August and September 2026, using a 50 trillion-token training mixture.

The run averaged 21% model FLOPs utilization and 89% goodput, Microsoft’s measure of time when training progressed. These are different metrics, not interchangeable efficiency percentages. Microsoft reports native vision pre-training and a corpus roughly 36% larger than MAI-Base-1 by unique training tokens. Preparations for MAI-Base-3 are already underway.

Microsoft also says its organic-data approach does not involve distillation from other models. It plans to license weights to trusted partners through Foundry; this is not an unrestricted open-weight release, and the post supplies no commercial license terms.

Why the efficiency claim is not cheaper API pricing

Microsoft fits scaling laws to estimate the compute its recipe would need to match DeepSeek-V4-Pro-Base on four held-out evaluations. Its reported 4.2–12.3× FLOPs-efficiency advantage describes those estimates—not a measured reduction in the total training invoice, inference cost, or customer bill. Microsoft says it used the saving to train a larger model.

The evaluations measure bits per byte, a tokenizer-independent held-out text metric where lower is better. Microsoft reports outperforming the open-weight models it tested, while noting that some comparison checkpoints were post-trained. That does not establish superiority over every proprietary assistant, coding benchmark, or production workload.

Engineering changes include improved mixture-of-experts routing, the Muon optimizer, and activation gradient clipping. These can improve training economics; they do not establish serving costs. The report discloses neither the full training bill nor MAI-Base-2 inference throughput.

Buying question	What Microsoft disclosed	Budget consequence
Can I buy MAI-Base-2 API tokens?	Post-training underway; no API rate card	No confirmed replacement cost
Can I license the weights?	Trusted-partner Foundry licensing intent	Wait for eligibility and contract terms
Did existing API rates fall?	No rate change announced	Keep current production assumptions

The generated pricing table is a baseline for available APIs, not pricing for Microsoft’s unreleased model. DeepSeek’s serving model is not the base checkpoint used in Microsoft’s evaluation. Its peak rates apply 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays; the displayed off-peak rates are half the peak rates. GPT pricing shown is Standard at up to 272,000 input tokens. Check DeepSeek pricing and OpenAI pricing for full billing conditions.

Labs availability blocker

MAI-Base-2 is not included in our Labs rankings. Microsoft says post-training is ongoing; no public exact-model endpoint, downloadable checkpoint, license terms or authorized Labs route is verified. A fair result needs pinned model and serving settings, fixed tasks and acceptance criteria, raw responses, retries, latency, token counts, and complete API bills or allocated serving costs. Microsoft’s held-out bits-per-byte comparisons are vendor evidence, not our 49-task text score. See the explicit Labs blocker.

What this means—and what to do now

Microsoft gains strategic control over the data, model recipe, infrastructure, and eventual distribution. Foundry customers could gain another model and a trusted-partner licensing route, but eligibility, price, deployment rights, and availability remain unannounced here.

Rival model suppliers face potential competition, not a confirmed pricing loss. Buyers expecting an immediate discount have no new rate to budget against. A sparse model’s active parameter count alone cannot establish affordable self-hosting: memory, interconnect, utilization, and operational support still matter.

Practical next steps

1. Keep the production baseline. Do not migrate or cut forecasts based on a training-efficiency multiple.
2. Prepare a repeatable evaluation. Track quality, latency, tokens, retries, and reviewer time per accepted result. Re-run it when a post-trained model becomes accessible.
3. Ask Foundry for commercial specifics. Request availability, regional endpoints, licensing restrictions, serving rates or capacity requirements, and support obligations.
4. Price measured usage. Put actual monthly input and output tokens into the token calculator; adjust for peak windows, context tiers, caching, and tools. Our API cost guide explains the method.

For a hosted open-model pilot today, compare available models on Novita. This is not MAI-Base-2 access or evidence of equivalent performance.

Affiliate disclosure: This article contains an affiliate link. We may earn a commission at no extra cost to you.

Official source and freshness checks run daily; changes require review and republishing of matching canonical/API availability records, not blind date refreshes. Sources checked October 15, 2026: Microsoft’s technical report, model pricing page and Foundry catalog, OpenAI’s rate card, and DeepSeek’s rate card.