Millis AI Review (2026)

Millis advertises two cents a minute. Its own pricing page then works through a ten-minute call and totals it at $0.6615.

That is 6.6 cents a minute, 3.3 times the advertised rate, computed by the vendor, printed by the vendor, and still with no telephony in it.

This site's whole method is to add the published list price of what an advertised rate excludes, and Millis is the widest gap it has found. The unusual part is that the finding needs no adversarial arithmetic at all. The vendor got there first and published the answer.

What the two cents actually buys

The pricing page is direct about it: the base rate is charged for using the platform, and it is in addition to the fees for the language model, synthesis and speech recognition. All four components sit outside the headline.

Which corrects something this site had wrong. Our record called this a bring-your-own-key model, and that is true of the language model alone. Millis resells speech recognition and synthesis and meters them itself — its own webhook payload bills speech at $0.0043 with the provider named as Millis. The cost is not moved to your vendor accounts. It is rebilled to you at Millis's rates.

Telephony is the one genuine exclusion. You import a number from Twilio, Vonage, Plivo or Exotel, or route a SIP trunk, and that carrier bills you directly.

The vendor's arithmetic against ours

This site computes Millis at $0.0665 a minute, using its own reference component rates. Millis computes $0.06615, using its own.

That is worth stating plainly because the site's headline claim rests on this entity. If the widest gap on the site were an artefact of our assumptions, this is where it would show. It does not. Two methods with nothing in common but the vendor's published rates agree to within a fifteenth of a cent.

The number that actually decides your bill

Not the base rate. The voice.

Synthesis is billed per thousand characters and Millis publishes a menu spanning a factor of seventeen. Hold every other assumption in the vendor's own example constant and change only which voice you click:

So the advertised two cents predicts almost nothing. One dropdown moves the real cost by four times, and it is the same dropdown every demo leaves on the best-sounding option.

This also explains a movement in the vendor's own numbers. The base rate has been $0.02 since January 2025, unchanged for nineteen months. Over the same period the worked example fell from $0.8615 to $0.6615, a 23% drop, entirely because the ElevenLabs rate Millis resells came down. The advertised price is the one number that has never moved and the one number that never mattered.

Where the page disagrees with itself

The synthesis table derives roughly five cents a minute for ElevenLabs from a stated assumption of about 500 characters in a typical minute of voice interaction. The worked example below it uses 4,000 characters across ten minutes, which is 400.

Run the page's own 500-character assumption through the page's own example and the total comes out at $0.5115, or 2.6 times the headline rather than 3.3. The two methods on one page disagree by 23%, and the difference is whether a minute means a minute of call or a minute of the agent talking. On the largest line of the bill, that ambiguity is not a detail.

The model menu is also dated. It prices GPT-4o, GPT-4 Turbo and GPT-3.5 Turbo, while the vendor's own changelog announced the GPT-4.1 family, Deepgram's Aura-2, Gladia's Solaria and several new voices during 2025. You can select providers in the product for which the pricing page publishes no rate at all. The changelog's last entry is from April 2025.

How you actually pay

A prepaid balance in dollars with automatic top-up. The public API exposes a credit balance, a used-credit figure and an auto-refill configuration with a threshold and a refill amount, and the credits are denominated one to one with dollars.

There are no plans, no seats and no monthly platform fee. Signup is email or Google with no sales gate, which makes this among the easiest platforms here to start on and among the hardest to forecast, because nothing published says what happens when the balance reaches zero, whether credits expire, or what concurrency you are allowed.

No free tier is published either. Our own record claimed one and no first-party source supports it, so it has been removed rather than left asserted.

Who ranks for this

Nine results, seven with a commercial stake: two vendor pages, four directories, and a services company that builds voice products and publishes a generic cost calculator. The independent entries are a Reddit thread and a YouTube video whose title is a savings claim.

Three of the eight related searches are people looking for a calculator, a Reddit thread or a review — which is the search stream asking for exactly the arithmetic none of the nine results performs, on a page where the vendor has already published it.

Self-serve, API-first, and you will be writing code. There is no visual builder to fall back on and no sales process to slow you down.

• Choose the voice before you model the cost, because it moves the all-in by four times and every other decision on the page moves it by less.
• Bring your own language model key if you have one. Millis charges nothing for a custom model, which removes one of the three components it otherwise rebills.
• Arrange telephony separately and price it. Twilio, Vonage, Plivo, Exotel or SIP, billed by them and absent from every figure Millis publishes.
• Ask what happens at zero balance — whether live calls drop mid-conversation, whether inbound fails, and whether there is any grace. None of it is published.
• Ask for a rate on any provider added since April 2025, because the pricing page still prices a 2024 menu and the newer options appear only in the changelog.

The most honest pricing page in the category and the one that produces the largest gap, which sounds like a contradiction and is not. Millis publishes the base rate, the component rates and a worked total. Reading all three is what produces 3.3×; reading only the first is what produces the headline.

Model it as four lines and put the synthesis line first, because it is the largest and the most variable. Then add telephony, which appears in none of the vendor's arithmetic and adds about $0.014 a minute at published US carrier rates, taking the example configuration past 4× the advertised price.

The comparison worth running is against Vapi, which advertises the same $0.05 orchestration model and computes to 1.93×, and against Ultravox at 1.28× where synthesis is inside the rate. Millis is cheaper than both at the bottom of its voice menu and dearer than both at the top, on the same platform, with the same headline.