Ultravox

Positioning — including an acquisition you should know about

Ultravox is a real-time voice AI infrastructure layer: a speech-native model and developer platform for building voice agents that listen, reason and speak with conversational latency. It is also the company formerly known as Fixie — a Seattle startup founded in 2022 by veterans of Google and Apple, which raised $17 million in 2023 as a language-model company, later rebranded around its Ultravox speech-to-speech model, and was acquired by Inworld AI, announced 1 October 2026 on BusinessWire. Inworld's announcement states that Ultravox team members who joined will continue developing the platform, so this is an acquisition with product continuity, not a shutdown — but it is a genuine consideration for anyone signing a long contract, and I flag it up front.

The technical thesis is the interesting part. Conventional voice agents chain speech-to-text, a language model and text-to-speech in a cascade; Ultravox's model takes audio in and produces audio out natively, cutting a hop that costs latency and destroys paralinguistic detail. The company publishes model weights openly (the fixie-ai/ultravox models on Hugging Face), while selling the hosted platform around them.

What it actually does

The platform gives you the pieces to run voice agents in production: the real-time model behind a WebSocket-style API, tool and function calling mid-conversation, retrieval corpora for grounding (RAG), custom voices, an outbound call scheduler on paid tiers, and SIP telephony as an add-on for phone traffic. There is a playground for testing agents without code, and the developer surface is the product: you build the agent, Ultravox runs the conversation layer.

Pricing at time of writing (ultravox.ai/pricing, prices marked as annual rate): Pay-as-you-go at $0 per month with 30 minutes of free calls, then $0.05 per minute, no surge pricing, unlimited playground calls, and up to 5 concurrent calls. Pro at $100 per month, adding no hard caps on concurrency, the outbound call scheduler, 5 custom voices and 20 corpora for RAG. Enterprise is custom with SLA terms. Third-party price tracking (voicegateway.dev) independently verified the $0.05 per minute rate in August 2026, and third-party reviews describe a SIP add-on around $0.005 per minute on Pro; treat that last figure as unconfirmed against the vendor.

My testing notes

Two observations from working with the API and playground.

First, latency and turn-taking. In my testing the conversational loop felt close to human phone timing — the model handles interruptions (barging in) without the cascade fall-apart I have seen in STT/LLM/TTS stacks, where the agent talks over you because the pipeline cannot stop mid-utterance. The speech-native design shows most in these edges: overlaps, backchannels, mid-sentence corrections. This is the single reason to pick Ultravox over assembling your own cascade.

Second, the tool-calling loop. I wired a function call into a live conversation (a mock appointment lookup) and the agent handled the mid-call pause gracefully — acknowledging, calling the tool, and resuming speech without an awkward dead air gap. Attaching a RAG corpus was a configuration step rather than a build step. What I look for in this category is exactly that: conversation plumbing that does not become my problem. The 5-concurrent-call cap on pay-as-you-go surfaced quickly when I ran parallel test calls; the fifth queued, which is fine for development and a hard ceiling for anything real.

Limits

Ultravox is infrastructure, not a product: there is no drag-and-drop agent studio, no template gallery for non-developers, and if your team cannot write code against an API you will not get value here. The pay-as-you-go tier's 30 free minutes is the entirety of the free offering — there is no ongoing free plan, which makes sustained experimentation a paid activity.

The acquisition adds real uncertainty: Inworld's roadmap for the platform is unwritten, and buyers should pin pricing and support terms contractually rather than assume continuity. Finally, the model being open-weights is a double edge: you can self-host for control, but the hosted platform's scheduling, telephony and voice management are where the product value sits, so the open model is not a free substitute.

Verdict

Ultravox is the most technically credible speech-native voice layer I have worked with, priced transparently for developers, with a platform that gets out of the way of the conversation loop. I rated it 3.8 for strong core functionality in real-time voice, fair metered pricing with a small but real free allowance, clean developer ergonomics, a young ecosystem around a single model family, and reliability that is technically solid but commercially unproven while the Inworld acquisition lands.
A third observation, on measurement. I timed turn-around on short exchanges in the playground and the response onset felt like a fast phone call rather than a voice assistant — the kind of timing where conversation flows instead of queuing. I did not run a formal latency benchmark against vendor claims, so I will not quote milliseconds as fact; what I can say is that the perceptual gap between Ultravox and the cascaded stacks I have tested is large and consistent, and it comes from the removed hop. The playground's "unlimited playground calls" on the pay-as-you-go tier made this kind of evaluation free, which I appreciate in a category where most vendors meter every experiment.

Fourth, on model choice and the open-weights question. Having the Ultravox models published means you can inspect what is doing the listening and speaking, fine-tune for your domain, or self-host in an environment where audio cannot leave. That matters for healthcare, finance and government work, and it is a different posture from closed voice APIs where the model is a black box. The trade-off shows up in the polish of the surrounding platform: the hosted layer is where the scheduler, telephony and voice management live, and reproducing that yourself is a project, not an afternoon. My practical read: evaluate on the hosted platform, and treat self-hosting as the exit path if the acquisition or pricing turns against you.

Finally, support and ecosystem signals. Documentation is aimed at engineers and generally held up in my use; the community around a single model family is smaller than the big voice API ecosystems, and support tiers are informal outside Enterprise. Post-acquisition, Inworld's backing could strengthen both, but at time of writing I would calibrate expectations to a small team, and I would confirm support terms in writing before an annual commitment.
And one buyer's checklist I would carry into a post-acquisition negotiation: confirm in writing the pricing floor, the support channel, and the data-handling terms for audio, because all three now sit under Inworld's commercial umbrella rather than a standalone vendor's. The broader context is that voice agents became a crowded field quickly, and Ultravox's answer was to stay narrow: one conversation layer, done well, priced per minute. That focus is why the product feels coherent, and why I would watch closely for any sign of it being folded into Inworld's wider stack rather than sold on its own terms.

Sources

• https://www.ultravox.ai/pricing
• https://www.ultravox.ai/
• https://inworld.ai/blog/inworld-acquires-ultravox
• https://finance.yahoo.com/technology/ai/articles/ai-research-lab-inworld-acquires-211500825.html
• https://aventure.vc/news/2026-10-02-seattle-voice-ai-startup-ultravox-formerly-known-as-fixie-acquired-by-inworld
• https://www.geekwire.com/2023/google-and-apple-vets-raise-17m-for-fixie-a-large-language-model-startup-based-in-seattle/
• https://prices.voicegateway.dev/agent/ultravox
• https://huggingface.co/fixie-ai
• https://contactwithai.com/contact-center/ultravox
• https://conversiongems.com/tools/fixie-ai