Microsoft

As of Sept 2026 the MAI line spans text reasoning (MAI-Thinking-1), image generation/editing (MAI-Image-2.5 and 2.6 series), text-to-speech (MAI-Voice-2), and speech-to-text (MAI-Transcribe-2), with a security-domain model (MAI-Cyber-1-Flash) appearing in Azure billing meters. Microsoft also publishes other in-house open-weight models outside the Phi/MAI lines (Fara-1.5 computer-use agents, MagenticBrain-14B, VibeVoice ASR, BitNet, Aurora, plus a large healthcare/science model set) — these are out of scope here but are in the same Foundry catalog.

A Access and limits

All hosted models require a paid Azure subscription plus a Microsoft Foundry (Azure AI Foundry) resource/project; authenticate with a resource API key or a Microsoft Entra ID bearer token scoped to https://cognitiveservices.azure.com/.default. Phi models deploy as serverless per-token endpoints (OpenAI-compatible chat completions) or as managed compute. MAI text/image models use dedicated Microsoft-managed endpoints: https://<resource>.services.ai.azure.com/mai/v1/chat/completions, /mai/v1/images/generations and /mai/v1/images/edits.

MAI-Voice and MAI-Transcribe are NOT Foundry Models — they run through Azure Speech in Foundry Tools (SSML voice names like en-US-Harper:MAI-Voice-2 via the Speech SDK/REST, and the Fast Transcription API with enhancedMode.model set to "MAI-Transcribe-2"). PRICE TIERS: every price recorded here is the BASE Global Standard rate. Azure adds roughly +10% for Data Zone Standard / regional deployments on Phi and MAI-Image-2.5-series meters, and a further premium (about +20% to +25% over global) in a small number of premium-priced regions.

Phi fine-tuning is billed separately at $3.00 per 1M training tokens plus $0.80/hour for hosting a fine-tuned deployment. Cached (prefix-cache) input is billed at 10% of the input rate on MAI-Thinking-1 ($0.20/1M) and MAI-Cyber-1-Flash ($0.06/1M). MAI-Thinking-1 is Global Standard only (no PTU) with default rate-limit tier 0 — you must request Medium (100K TPM / 100 RPM) or High (250K TPM / 250 RPM) quota.

MAI-Image models default to 0 RPM and step up to 12 RPM at tier 6 via the quota request form. MAI-Voice instant voice cloning / Personal Voice is gated behind the Custom Neural Voice Limited Access Review (aka.ms/customneural). Open-weight Phi checkpoints are free under the MIT licence on huggingface.co/microsoft with no key required; several are also mirrored to ONNX/GGUF and to Foundry Local for on-device use.

Prices verified 2026-09-05 against the official Azure Retail Prices API (prices.azure.com), which is the machine-readable source behind the azure.microsoft.com pricing pages.

C Calling the API

Microsoft accepts requests in the OpenAI Chat Completions format, so any OpenAI-compatible client works by changing the base URL. Example uses Phi-4-mini-instruct.

curl https://<resource-name>.services.ai.azure.com/chat/completions \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Phi-4-mini-instruct",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

from openai import OpenAI

client = OpenAI(
    base_url="https://<resource-name>.services.ai.azure.com",
    api_key=os.environ["API_KEY"],
)

response = client.chat.completions.create(
    model="Phi-4-mini-instruct",
    messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)

A compatibility layer is not always the provider's full API — caching, strict tool schemas and structured output are commonly unsupported there. Check the official docs before relying on an advanced feature.

D Official references

• Homepage
• API documentation
• Pricing