Skip to content

Compare models and cost

Compare the same task, language, and output format—not just model names. The app's feature map shows its four model choices.

The MAI family

Six featured MAI families: Voice for speech synthesis, Transcribe for audio to text, Image for generation and edits, Thinking for reasoning, Code for agentic coding through Copilot, and Cyber for defensive work through restricted MDASH. Frontier Tuning customizes models.

MAI-generated diagram. Full size · Prompts and revisions.

These are the currently featured families, checked 18 September 2026, not every historical version. Voice and Image include Flash variants. Code is offered through Copilot; Cyber has restricted MDASH access. Frontier Tuning is model customization, not a seventh model. Those three are not workshop runtime calls.

What the prices establish

Public Azure USD pay-as-you-go rates, checked 18 September 2026. These are normalized inference budgets, not measured app runs or a claim of equal quality. Cache savings, free allowances, negotiated rates, and Batch/Flex discounts are excluded.

Task and workload MAI Comparable Azure model Cost conclusion
Transcribe 1,000 audio minutes MAI-Transcribe-2: $1.67 at the advertised $0.10/hour promotion Whisper 001, Regional: $6.60 at $0.396/hour 74.7% lower if the MAI promotion applies.
Speak 1 million billable characters MAI-Voice-2-Flash: model-specific tariff not verified tts 001, Regional: $15 No verified MAI saving yet.
Generate 100 images at 1024 × 1024 MAI-Image-2.6-Flash: $1.75/million text-input tokens; $19/million image-output tokens GPT Image 1, medium: $4.274 under the token assumptions below MAI's token rates are lower; a per-image saving is not established.
Generate text with 1 million uncached input + 100,000 output tokens MAI-Thinking-1: $2.80 at $2/$8 per million input/output tokens GPT-5-mini: $0.45 at $0.25/$2 MAI costs $2.35 more for this token mix.

Transcription: the MAI announcement quotes $0.10/hour; Speech pricing marks a discount through 31 December 2026, while its generic MAI table still shows $0.36/hour. Confirm the applicable offer before budgeting. The comparison uses MAI in East US and Whisper Regional in North Central US. Whisper retail rate.

Speech: do not assign generic Neural/Neural HD prices to MAI-Voice-2-Flash without a confirmed tariff. The baseline is ordinary Azure OpenAI tts in North Central US, not tts-hd or token-metered gpt-4o-mini-tts. OpenAI pricing on Azure.

Images: text-to-image only, with no input image or web grounding. GPT Image 1 medium uses 1,056 output tokens/image at $40/million, plus an assumed 100 prompt tokens/image at $5/million: 100 × (100 × 5 + 1056 × 40) / 1,000,000 = $4.274. For MAI, independently assume 10,000 text-input tokens (100 per image), costing $0.0175. Identical prompts need not tokenize identically across models. MAI costs $0.0175 + 19 × Q / 1,000,000, where Q is its actual total image-output tokens. No verified MAI pixel-to-token formula or equivalent medium setting was found, so Q is not guessed. GPT image token table · MAI retail rates.

Text: output counts include reasoning, not just the visible mnemonic. GPT-5-mini is a smaller reasoning baseline for this short task, not an equal-capability claim. MAI pricing · Azure OpenAI pricing.

Image/text rows use Global Standard: MAI in East US, GPT Image 1 in East US 2, GPT-5-mini in East US. Versions: Image Flash 2026-07-31, GPT Image 1 2025-04-15, Thinking 2026-06-01, GPT-5-mini 2025-08-07. The older comparators remain listed; they are not presented as the latest models.

Try a controlled comparison

Use the app and the comparator's existing playground—no new adapter needed.

  1. Pick one feature and ten pairs in a language supported by both models. Reuse the same WAV for transcription, text/language for speech, or prompt and requested size for images.
  2. Run each model three times. Record useful results, errors, total latency, and its actual billable units. For text, include reasoning tokens. MAI image responses carry a usage object (for example num_output_tokens); log it as lesson 5 shows. To compare MAI-Image-2.6 with Flash, set MAI_IMAGE_DEPLOYMENT=mai-image in .env and restart Flask; for MAI-Voice-2, change a voice's :MAI-Voice-2-Flash suffix to :MAI-Voice-2.
  3. Compare cost per useful result, not just price per request. Keep voice, quality, deployment, region, and settings in your notes.

If the MAI voice tariff or actual image-output token count is unavailable, record cost as unverified rather than ranking it.

Storage and answer matching make no model call. Gateway, hosting, and Codespaces costs are separate.