Ultravox Review (2026)

This site had Ultravox wrong, so start there.

Our cost model recorded the advertised 5 cents a minute as excluding synthesis, which put the computed all-in at $0.078 and the multiple at 1.56 times the headline. The vendor's homepage says otherwise, in nine words: five cents per minute, including TTS, for up to five concurrent calls.

Corrected, the only excluded component is telephony, the computed floor is $0.064 a minute, and the multiple is 1.28. That moves Ultravox from the fifth-widest gap in the category to the fourth-narrowest, and it moves the number in the direction that flatters the vendor, which is exactly the direction this site has to be most careful about.

Why there is no transcription line to bill

The architecture explains the price. Ultravox is a speech-native model: audio goes in and text comes out, with no automatic speech recognition stage in between. There is no transcription step, so there is nothing to charge for one.

That is one hop removed from the latency budget, which is the entire pitch and a real engineering advantage. It also means synthesis is genuinely a separate step the vendor buys and folds into the rate, with ElevenLabs listed among its subprocessors. You can bring your own synthesis key instead, but that adds a bill rather than removing one.

Telephony is excluded twice, and only one of them is obvious

Ultravox sells no phone numbers. It describes itself as a bring-your-own-telephony platform, with integrations for Twilio, Telnyx, Plivo and Exotel, so the carrier bill is yours as it is on most platforms here.

The part that is easy to miss is that connecting by SIP incurs an Ultravox surcharge as well: half a cent a minute on pay-as-you-go, in addition to the regular call minute. So the honest stack has three lines rather than two.

Two quirks in that SIP line are worth knowing before you model it. Ultravox states that SIP minutes can accrue while your agent is not connected to the user, on transfer legs that outlive the agent. And billing rounds up to six-second blocks, which the vendor calls rounding to the nearest deciminute while every example in its own table rounds up.

What the $100 plan actually buys

Pro is $100 a month and the per-minute rate on it is the same 5 cents. The only rate that moves is SIP, from half a cent to $0.0048.

Two hundredths of a cent. Recovering the fee on rate alone would take half a million SIP minutes a month, which is about eleven calls running continuously, day and night, for a month.

So Pro is not a volume discount and should not be evaluated as one. It is a concurrency unlock: pay-as-you-go caps you at five simultaneous calls, and Pro removes the cap. That is worth $100 to anyone running production traffic and worth nothing to anyone else, which is a cleaner proposition than the pricing table makes it look.

The pricing page carries a banner announcing that prices are marked as annual rates. Nothing on the page is marked. The homepage supplies the missing asterisk, noting the $100 applies when billed yearly, and the monthly price appears nowhere. That is the reverse of the usual trick: the discount badge is published and the undiscounted price is not.

The transcript is free, and that is the problem

Our own record used to say you do not get a clean transcript for free. That was wrong in the other direction. You do get one, at no charge.

What you get is the thing worth understanding. Because the model never reads text, transcripts are produced by a conventional ASR pass running alongside, and the vendor's documentation says they exist solely to help you understand what is happening in your calls. When the transcript and the agent's behaviour disagree, the vendor's own guidance is to listen to the recording, because the transcript is usually the inaccurate one.

Read that as a compliance statement rather than a support note. A speech-native platform is documenting that its transcript is not a record of what the model heard. It is a parallel best-effort reconstruction. For debugging that is fine and honest. For a regulated call where the transcript is the evidence of what was said and agreed, it is a different proposition entirely, and the vendor's trust page lists SOC 2 Type II and HIPAA as in progress rather than complete.

This is not a reason to avoid the product. It is a reason to know which of your calls can run on it.

Open weights, and what the vendor does not say about them

The models are published on Hugging Face under Fixie's account, MIT-tagged and ungated, from v0.2 through v0.7. That is genuinely open in a category where the word is usually decorative, and the smaller checkpoints have been downloaded over a million times in a month.

Two caveats the vendor leaves to you. The MIT tag sits on Fixie's repository, while the base models underneath carry their own licences, which that tag does not override. And nothing published compares the economics of self-hosting against the hosted 5 cents, so the question the open weights raise is the one question the pricing page does not answer.

A plan that exists in the docs and not on the pricing page

The concurrency documentation describes three plans: pay-as-you-go, Pro, and a Scale plan offering priority access to as many as a hundred concurrent calls, recommended for sustained high load. The pricing page has no Scale tier at all, and its priority-concurrency row reads no, no and custom.

A buyer reading the pricing page cannot find the plan the documentation tells them to buy. Ask for it by name.

Of nine results on this product's pricing query, five are Ultravox's own pages and two more are written by companies with a direct commercial stake, one of them a competitor and one a reseller wrapping the product. There is no independent review on the first page, which is the pattern on every query this site has audited.

Free tier, no card, thirty minutes of calls and unlimited playground calls that are never charged. Genuinely self-serve up to the point where concurrency matters.

• Model the SIP line separately from the call line. They are two clocks, they round up independently, and one of them can run while your agent is not on the call.
• Budget Pro as a concurrency unlock, not a discount. Five simultaneous calls is the free ceiling; the $100 removes it and changes nothing about your rate.
• Decide early whether your calls need an evidentiary transcript. If they do, this architecture does not produce one and no configuration will change that.
• Ask for the Scale plan by name if you need priority concurrency, because it is in the documentation and not on the pricing page.
• Ask what Pro's $100 actually is. The pricing table implies a platform fee and the FAQ describes it as covering usage in advance. Those are different things.

Unusually honest at the headline and incomplete underneath it. The rate is real, it has held for two years, and synthesis really is inside it. Three things are missing.

The comparison worth running is not against the bundled platforms. It is against a stack you assemble on Vapi, where the same components are billed separately at 1.93× the advertised rate. Ultravox lands cheaper all-in and gives up the ability to swap the model, because here the model is the product.