What Is the Microsoft MAI Model Family?
MAI is Microsoft's first in-house model family, introduced at Microsoft Build on June 2, 2026. Microsoft presented seven first-party models spanning reasoning, coding, image generation, voice and transcription, and the usage bills against Azure rather than arriving as a separate vendor invoice.
That billing detail is what separates MAI from every other 2026 model launch. OpenAI, Anthropic and Google credits arrive as their own balances on their own accounts. MAI usage draws down the same Azure balance a startup may already be holding and underspending.
The models named at the keynote were MAI-Thinking-1 for reasoning, MAI-Code-1-Flash for coding, MAI-Image-2.5 for image generation and editing, MAI-Voice-2 for speech synthesis and MAI-Transcribe-1.5 for transcription.
One caveat sits over everything below. The keynote itself is officially confirmed. The per-model prices are not: the figures in circulation come from Azure pricing listings and Microsoft Community Hub posts, which is semi-official at best. Read every number here as reported rather than confirmed, and check the live Azure pricing page before budgeting on it. AI Perks tracks the credit programs that offset these bills.
How Much Do Microsoft MAI Models Cost?
Reported MAI prices run from $0.36 per hour of audio for transcription up to $47 per million image output tokens for the full image model. MAI-Thinking-1 was introduced in private preview with pricing still to be announced, so the reasoning model has no list price to plan against.
| Model | Task | Reported price | How well sourced |
|---|---|---|---|
| MAI-Thinking-1 | Reasoning | Not announced at introduction | Private preview, no price to verify |
| MAI-Code-1-Flash | Coding | Bundled into GitHub Copilot, no separate token rate found | Reported, weakly sourced |
| MAI-Image-2.5 | Image generation and editing | $5 / 1M text in, $8 / 1M image in, $47 / 1M image out | Reported, semi-official listing |
| MAI-Image-2.5 Flash | Faster image generation | $1.75 / 1M text and image in, $33 / 1M image out | Reported, semi-official listing |
| MAI-Voice-2 | Text to speech | $22 / 1M characters | Reported, semi-official listing |
| MAI-Transcribe-1.5 | Speech to text | $0.36 / hour of audio | Reported, semi-official listing |
Three structural points matter more than the individual numbers. Image models bill per token, not per image. A 4K render and a thumbnail are not the same purchase, and image output is the dominant cost line at a reported $47 per million tokens on the full model against $33 per million on Flash. The Flash variant also collapses text and image input into one reported $1.75 per million rate.
Voice bills per character, so the cost is fixed by script length before you generate anything. Transcription bills per hour of audio, the only line here that converts directly into a budget forecast: a thousand hours is about $360 at the reported $0.36 rate, and a backlog of 500 podcast episodes at 45 minutes each works out to 375 hours, or roughly $135.
Reasoning has no number at all yet, so any MAI-Thinking-1 token price quoted elsewhere is a guess until Microsoft publishes one.
How Does MAI Image Pricing Compare?
On reported figures, MAI-Image-2.5 matches GPT Image 2 exactly on input pricing and costs more on output, while Google's Nano Banana 2 inverts the shape entirely with very cheap input and the most expensive output tokens of the three.
| Model | Reported input | Reported output | Free path |
|---|---|---|---|
| MAI-Image-2.5 | $5 / 1M text, $8 / 1M image | $47 / 1M image tokens | General Azure trial credits |
| MAI-Image-2.5 Flash | $1.75 / 1M text and image | $33 / 1M image tokens | General Azure trial credits |
| GPT Image 2 (OpenAI) | $5 / 1M text, $8 / 1M image, $2 / 1M cached image | $30 / 1M image tokens | ChatGPT Free tier, rate limited |
| Nano Banana 2 (Google) | $0.50 / 1M text and image | $60 / 1M output tokens | Gemini app and Google AI Studio |
The identical reported input rates on MAI-Image-2.5 and GPT Image 2 are the interesting part. Input pricing appears to have converged, and the competitive difference has moved to the output side, where the reported spread runs from $30 to $60 per million tokens. Two discount levers exist only outside MAI on current sourcing: OpenAI's reported $2 per million cached image input rate, and Google's batch pricing at 50 percent off. No equivalent cached or batch rate has been confirmed for MAI.
Google also publishes per-image estimates of roughly $0.045 to $0.151 depending on output resolution between 0.5K and 4K, which is the clearest view of what token-based image billing costs per asset. Microsoft publishes no equivalent per-image figure we could verify.
Treat the launch dates of the comparison models with care. GPT Image 2 is reported by most sources as arriving on June 24, 2026, with at least one placing it in April 2026, and Nano Banana 2 is reported around May to June 2026 with no confirmed date. Credits for all three sit inside general startup programs rather than model-specific offers, and those are tracked at getaiperks.com.
What Is the MAI Model Inside GitHub Copilot?
MAI-Code-1-Flash was reported to have become the default model across GitHub Copilot tiers around August 2026. The exact date is not confirmed, and this is the most weakly sourced claim on this page.
Reported specifications put it at a 5 billion parameter coding model scoring 51 percent on SWE-Bench Pro. Neither figure is confirmed by a verifiable Microsoft page, so use them for orientation only.
The commercially relevant part is cleaner. No separate API token rate for MAI-Code-1-Flash was found. Inside Copilot it is bundled into the subscription, so its marginal cost is zero once the seat is paid for, and GitHub Copilot has a free tier, which makes a MAI coding model the most accessible member of the family.
Coding spend on MAI is therefore a subscription question, not a token question. If you are optimising a token budget, the image, voice and transcription models are where the arithmetic lives. Credits for the rest of the stack are tracked at AI Perks.
What Microsoft Has Not Published
The gaps here are as decision-relevant as the prices, and two of them will shape a build decision before a single request is served.
- No list price for MAI-Thinking-1. It was introduced in private preview with pricing still to be announced, and none has been confirmed since.
- No model-specific free tier. General Azure free-trial credits apply to MAI the way they apply to any other Azure service. No per-model free allowance has been confirmed for MAI-Thinking-1, MAI-Image, MAI-Voice or MAI-Transcribe.
- No separate token rate for the Copilot coding model. Any token price attached to it refers to API access rather than to what a Copilot subscriber pays, and no such rate has been confirmed.
- No single confirmed price page for the family. The figures sit across Azure pricing listings and community posts rather than one authoritative page.
How Azure Credits Cover MAI Usage
MAI bills to Azure, so an Azure credit balance is the funding mechanism. There is no separate MAI credit program: the general Microsoft startup track covers model usage the same way it covers storage or compute.
| Credit program | What it covers | Where it is tracked |
|---|---|---|
| Microsoft for Startups Founders Hub | General Azure and AI credits, usable on MAI through Azure billing | AI Perks Guide |
| Azure free trial credits | General trial balance, with no MAI-specific allowance confirmed | AI Perks Guide |
| OpenAI for Startups and Google for Startups Cloud | General programs sitting behind the comparison models above | AI Perks Guide |
Because the balance is general rather than model-specific, sequencing decides how far it goes. On reported rates, transcription and the Flash image variant consume a balance slowest per unit of useful output, so they are the natural place to start while expensive output-token workloads are still being validated.
Credits also do not change a list price. They change who pays it. That matters when a reported figure later turns out to be wrong: a credit balance absorbs a price correction quietly, and a cash budget does not.
Frequently Asked Questions
How much does MAI-Transcribe-1.5 cost per hour?
It is reported at $0.36 per hour of audio, so a thousand hours works out to about $360. That figure comes from Azure pricing listings rather than a confirmed Microsoft price page, so verify it before committing a budget. Transcription is the only MAI line that converts directly into hours of recorded material.
Does MAI-Thinking-1 have a published price?
Not at introduction. Microsoft presented it in private preview with pricing still to be announced. Reported architecture details describe a mixture-of-experts model with roughly 35 billion active parameters and a 256K context window, but no list price accompanied those figures and none has been confirmed since.
Which MAI model runs inside GitHub Copilot?
MAI-Code-1-Flash was reported to have become the default model across GitHub Copilot tiers around August 2026, with the exact date unconfirmed. Reported specifications put it at 5 billion parameters and 51 percent on SWE-Bench Pro. It is bundled into the Copilot subscription, with no separate API token rate found.
Is there a free tier for Microsoft MAI models?
Not a model-specific one. General Azure free-trial credits apply, and GitHub Copilot's free tier is the one genuinely free route to a MAI model. Microsoft has not confirmed a per-model free allowance for MAI-Thinking-1, MAI-Image, MAI-Voice or MAI-Transcribe.
Is MAI-Image-2.5 cheaper than GPT Image 2?
Not on reported figures. Both are listed at $5 per million text input tokens and $8 per million image input tokens, but MAI-Image-2.5 is reported at $47 per million image output tokens against $30 for GPT Image 2, which also carries a reported $2 per million cached image input rate with no confirmed MAI equivalent. MAI-Image-2.5 Flash, reported at $33 per million output tokens, is the closer comparison.
Do Microsoft for Startups credits cover MAI models?
Yes, indirectly. MAI bills to Azure, so an Azure credit balance funds MAI usage exactly as it funds any other Azure service. There is no separate MAI credit program to look for. The Microsoft for Startups Founders Hub is the general program, tracked alongside comparable offers at AI Perks.
Microsoft built the models. Azure credits pay for them.