Back to Developer ZoneOfficial Pricing Official Pricing Official Pricing Official Pricing Official Pricing Official Pricing Official Pricing
Microsoft Models
Explore all 7 models from Microsoft with detailed pricing, pros & cons, and developer recommendations.
7
Models
$0.360
Lowest Input
262K
Max Context
4
Quality Tiers
Quick Recommendations
Best Value: MAI-Transcribe-1.5 ($0.360/1M)
Best Quality: MAI-Image-2.5
Best for Reasoning: MAI-Thinking-1
MAI-Thinking-1
ReasoningComplex reasoning, math, agentic coding
When to use: For complex multi-step reasoning, math problems, and agentic coding where accuracy matters more than raw speed.
Upgrade Highlights
- ◆~1T total params, 35B active MoE — competitive at mid-weight
- ◆AIME 2025: 97.0%, AIME 2026: 94.5% — advanced math reasoning
- ◆SWE-bench Pro: competitive with Claude Opus 4.6
- ◆Faster first-token latency than o3 for real-time reasoning
- ◆Clean, traceable enterprise-grade data — no distillation
Input Price
$5.00
per 1M tokens
Output Price
$20.00
per 1M tokens
Cached Input
$1.00
per 1M tokens
Batch Input
—
per 1M tokens
Context Window: 262K
Max Output: 32,000 tokens
Knowledge Cutoff: 2026-03
VisionFunction CallingFine-tuningJSON Mode
Pros
- 35B active / ~1T total params — strong for its weight class
- AIME 2025: 97.0%, AIME 2026: 94.5% — top math reasoning
- No distillation — trained from scratch on clean data
- Preferred to Sonnet 4.6 in blind human evaluations
Cons
- No vision support yet
- New model — limited independent benchmarks
- Azure-centric availability
- Higher cost than non-reasoning models
Performance
Output Speed~55 tok/s
Rate Limit3,000 RPM
Multimodal
Image InputImage OutputAudio InputAudio Output
Benchmarks
AIME 2025
97.0%
AIME 2026
94.5%
GPQA Diamond
82.0%
MAI-Code-1-Flash
LiteFast agentic coding, IDE integration
When to use: Best for fast code completion, inline suggestions, and everyday coding assistance in IDE environments.
Upgrade Highlights
- ◆5B active params — comparable to Haiku but 60% cheaper
- ◆$0.75/M input — cheapest coding model in market
- ◆60% fewer tokens per task — inference-efficient design
- ◆Native GitHub Copilot and VS Code integration
- ◆HumanEval: 92%+ — strong coding accuracy at this size
Input Price
$0.750
per 1M tokens
Output Price
$3.00
per 1M tokens
Cached Input
$0.150
per 1M tokens
Batch Input
—
per 1M tokens
Context Window: 131K
Max Output: 16,384 tokens
Knowledge Cutoff: 2026-03
VisionFunction CallingFine-tuningJSON ModeFree Tier
Pros
- 60% fewer tokens per task vs alternatives
- 5B active params — ultra-low latency
- Deeply integrated into GitHub Copilot and VS Code
- Cheaper than Haiku at $0.75/M input
Cons
- Smaller model — weaker on complex multi-file tasks
- No vision support
- Limited to coding tasks — not general purpose
- 128K context is smaller than flagship models
Performance
Output Speed~150 tok/s
Rate Limit10,000 RPM
Multimodal
Image InputImage OutputAudio InputAudio Output
Benchmarks
HumanEval
92.0%
SWE-bench Lite
55.0%
MAI-Image-2.5
FlagshipEnterprise image generation and editing
When to use: For brand-safe, policy-compliant image generation and editing in enterprise and Microsoft 365 workflows.
Upgrade Highlights
- ◆Unified text-to-image + image editing in one model
- ◆Arena ELO: surpasses Nano Banana Pro
- ◆Superior text rendering in generated images
- ◆Enterprise compliance: private Azure deployment
- ◆Flash variant available for cheaper, faster generation
Input Price
$5.00
per 1M tokens
Output Price
$47.00
per 1M tokens
Cached Input
—
per 1M tokens
Batch Input
—
per 1M tokens
Context Window: 8K
Max Output: 0 tokens
Knowledge Cutoff: 2026-03
VisionFunction CallingFine-tuningJSON Mode
Pros
- World-class text-to-image and image editing in one model
- Surpasses Nano Banana Pro Arena score
- Enterprise-grade: private Azure deployment available
- Clean text rendering within images
Cons
- Expensive output at $47/M tokens
- Not as artistic as Midjourney
- Azure-centric availability
- No function calling
Performance
Output Speed~20 tok/s
Rate Limit1,000 RPM
Multimodal
Image InputImage OutputAudio InputAudio Output
Benchmarks
GenEval
88.0%
MAI-Image-2.5-Flash
Mid-tierFast, cost-efficient image generation
When to use: For high-volume image generation where speed and cost matter more than absolute quality.
Upgrade Highlights
- ◆$1.75/$15 — 65% cheaper than MAI-Image-2.5
- ◆Optimized for high-throughput generation
- ◆Free tier available for development
- ◆Same model family as MAI-Image-2.5
Input Price
$1.75
per 1M tokens
Output Price
$15.00
per 1M tokens
Cached Input
—
per 1M tokens
Batch Input
—
per 1M tokens
Context Window: 8K
Max Output: 0 tokens
Knowledge Cutoff: 2026-03
VisionFunction CallingFine-tuningJSON ModeFree Tier
Pros
- Much cheaper than standard MAI-Image-2.5
- Fast generation for high-volume workflows
- Same quality as MAI-Image-2.5 for most prompts
- Free tier available
Cons
- Lower quality on complex multi-object prompts
- No function calling
- Azure-centric
- Less detail in fine textures
Performance
Output Speed~40 tok/s
Rate Limit5,000 RPM
Multimodal
Image InputImage OutputAudio InputAudio Output
Benchmarks
GenEval
82.0%
MAI-Transcribe-1.5
FlagshipEnterprise speech-to-text, real-time transcription
When to use: For enterprise meeting transcription, call center analytics, medical dictation, and multilingual captioning.
Upgrade Highlights
- ◆SOTA accuracy: beats Whisper Large v3 on English and multilingual ASR
- ◆5x faster than competing transcription models
- ◆43 languages with built-in domain vocabulary support
- ◆Speaker diarization: identifies who said what
- ◆$0.36/hour — competitive pricing for enterprise STT
Input Price
$0.360
per 1M tokens
Output Price
$0.0000
per 1M tokens
Cached Input
—
per 1M tokens
Batch Input
—
per 1M tokens
Context Window: 0
Max Output: 0 tokens
Knowledge Cutoff: 2026-03
VisionFunction CallingFine-tuningJSON Mode
Pros
- Best transcription accuracy in class — SOTA on ASR benchmarks
- 5x faster than competing models
- 43 languages with domain-specific terminology
- Speaker diarization and real-time streaming
Cons
- Specialized model — transcription only
- Priced per audio-hour, not tokens
- Azure-centric availability
- No text generation capabilities
Performance
Output Speed—
Rate Limit—
Multimodal
Image InputImage OutputAudio InputAudio Output
Benchmarks
WER (LibriSpeech)
1.8%
WER (Common Voice)
5.2%
MAI-Voice-2
FlagshipHigh-quality TTS, voice cloning, conversational AI
When to use: For customer service voice bots, accessibility tools, e-learning content, and real-time conversational AI.
Upgrade Highlights
- ◆15 languages with natural prosody and tone control
- ◆Voice cloning: adapt from short audio samples
- ◆Sub-300ms latency — viable for real-time dialogue
- ◆Strong misuse safeguards built in
- ◆Flash variant available for ultra-low-cost deployment
Input Price
$2.00
per 1M tokens
Output Price
$10.00
per 1M tokens
Cached Input
—
per 1M tokens
Batch Input
—
per 1M tokens
Context Window: 0
Max Output: 0 tokens
Knowledge Cutoff: 2026-03
VisionFunction CallingFine-tuningJSON Mode
Pros
- Natural-sounding speech across 15 languages
- Voice cloning from short audio samples
- Sub-300ms latency for real-time conversation
- Enterprise compliance and Azure integration
Cons
- Specialized model — speech synthesis only
- Voice cloning requires enterprise gating
- Less flexible cloning than ElevenLabs
- Azure-centric availability
Performance
Output Speed—
Rate Limit—
Multimodal
Image InputImage OutputAudio InputAudio Output
Benchmarks
MOS (Naturalness)
4.5/5.0
MAI-Voice-2-Flash
Mid-tierUltra-low-cost speech synthesis
When to use: For high-volume TTS applications where cost efficiency matters more than absolute voice quality.
Upgrade Highlights
- ◆$0.50/$2.50 — 75% cheaper than MAI-Voice-2
- ◆Free tier for development and testing
- ◆Same 15-language coverage as MAI-Voice-2
- ◆Optimized for high-throughput, low-latency deployment
Input Price
$0.500
per 1M tokens
Output Price
$2.50
per 1M tokens
Cached Input
—
per 1M tokens
Batch Input
—
per 1M tokens
Context Window: 0
Max Output: 0 tokens
Knowledge Cutoff: 2026-03
VisionFunction CallingFine-tuningJSON ModeFree Tier
Pros
- Ultra-efficient TTS at 75% lower cost
- Free tier available
- Same language coverage as MAI-Voice-2
- Low latency for real-time applications
Cons
- Slightly lower naturalness than MAI-Voice-2
- No voice cloning in Flash variant
- Specialized model — speech only
- Newer model — limited independent reviews
Performance
Output Speed—
Rate Limit—
Multimodal
Image InputImage OutputAudio InputAudio Output
Benchmarks
MOS (Naturalness)
4.2/5.0
Side-by-Side Comparison
| Model | Tier | Input | Output | Context |
|---|---|---|---|---|
| MAI-Thinking-1 | Reasoning | $5.00 | $20.00 | 262K |
| MAI-Code-1-Flash | Lite | $0.750 | $3.00 | 131K |
| MAI-Image-2.5 | Flagship | $5.00 | $47.00 | 8K |
| MAI-Image-2.5-Flash | Mid-tier | $1.75 | $15.00 | 8K |
| MAI-Transcribe-1.5 | Flagship | $0.360 | $0.0000 | 0 |
| MAI-Voice-2 | Flagship | $2.00 | $10.00 | 0 |
| MAI-Voice-2-Flash | Mid-tier | $0.500 | $2.50 | 0 |