MAI-Voice-2.1-Flash '60% cheaper': what the claim compares to

Microsoft's announcement prices MAI-Voice-2.1 at $22 per million characters and MAI-Voice-2.1-Flash at $15 per million, and describes Flash as about 60% cheaper than comparable models. The 60% is against other vendors' models, not against MAI-Voice-2.1. Against the standard model the saving is 1 - 15/22 = 31.8%. Sume's TTS list is $0.0475 per 1,000 characters, which is $47.50 per million before cent rounding, 3.17 times the Flash rate.

Recomputing the numbers

Per character, MAI-Voice-2.1 is $0.000022, Flash is $0.000015 and Sume is $0.0000475. The ratio of Sume to Flash is 0.0000475 / 0.000015 = 3.17. A 20,000-character script, the largest Sume accepts per job, costs $0.44 on MAI-Voice-2.1, $0.30 on Flash and $0.95 on Sume. All three are fractions of a dollar per job, so the gap matters only at volume.

What the 60% does not tell you

The comparison set is Microsoft's own, so this post cannot reproduce the 60% figure; check it against the models you actually use. The page also lists Flash with a 45-second audio limit per request, which shapes the real cost of a long script: you pay per character, but you also split and stitch. Sume accepts up to 20,000 characters and 1,200 seconds of audio per job, which suits one-pass narration.

When the rate decides, and when it does not

At 1 million characters a month, Flash costs $15 and Sume $47.50, a difference of $32.50. That is real at 100 million characters and trivial at 100,000, where it is 3 cents against 14 cents of list price. Decide on the rate when you generate in bulk. Decide on workflow when you generate in dozens: voices you own, source-bound receipts, one webhook and one invoice across video, music and speech.

A quick calculator

For any monthly volume V in characters, Flash costs V x 0.000015 dollars, MAI-Voice-2.1 costs V x 0.000022, and Sume costs V x 0.0000475 before per-job rounding up to whole cents. At 10 million characters that is $150, $220 and $475. At 100 million it is $1,500, $2,200 and $4,750. Sume's per-job cent rounding adds at most 1 cent per job, which is small next to the rate at large volume and large for very short jobs.

for v in (100_000, 10_000_000, 100_000_000):
    flash, std, sume = v * 15 / 1e6, v * 22 / 1e6, v * 47.5 / 1e6
    print(f"{v:>11,} chars  Flash ${flash:,.2f}  MAI-Voice-2.1 ${std:,.2f}  Sume ${sume:,.2f}")

How to test a 'cheaper than' claim

Ask three questions of any such claim. Which models were compared? On which date were their prices read? And is the unit the same, characters or tokens or audio seconds? A saving quoted against a basket you do not use tells you nothing about your bill. Replace it with a calculation on your own volume, as above, and write the date next to the result.

Also count what the price excludes. Per-character prices do not include your time spent splitting 45-second pieces back into a single narration, nor retries on failures, nor the receipt that proves which text was spoken.