GPT-Live-1 vs Inworld Realtime TTS-2: A Price War on Voices, a Premium on Listening

Inworld Realtime TTS-2 reached general availability with something the voice market had been missing: a published rate card. The flagship model is $25 per million characters on demand, its faster sibling Inworld Realtime TTS-2 Flash is $15, and both step down through volume tiers to $12.50 and $7. Sitting at Elo 1,244 and second place on Artificial Analysis's speech arena, that makes the flagship roughly half the price of the current arena leader and one of the cheapest ways to buy a top-five voice. GPT-Live-1 is the other model in this comparison and the opposite proposition — $0.05 per minute of voice, no characters involved, and no way to buy it without also buying the listening, the turn-taking and the interruption handling that come attached.

The interesting thing about comparing them is that the price gap looks enormous and then mostly evaporates. Synthesis is in a price war. Conversation is not.

What Inworld actually shipped

The TTS-2 line began as a research preview in May 2026 with a specific claim: that this is a model which does not generate speech in isolation. It takes the prior turns of a conversation as audio input, reasons about tone and pacing, and adjusts how it responds. That positioning — conversational awareness rather than cleaner audio — is what separated it from the narration engines that dominated text-to-speech through 2025.

General availability turned the preview into a family and attached a price list. Flash is the throughput play, with Inworld quoting around 20 milliseconds of server-side time-to-first-byte at the 90th percentile against 100 milliseconds for the flagship and a sub-200-millisecond median to first audio. Those are vendor figures from Inworld's own documentation; the one third-party measurement in circulation, from Coval, puts Flash at roughly 25 milliseconds and the flagship in the low hundreds. Neither has been independently audited end to end.

The feature most worth noting has nothing to do with latency. Inworld added a verbatim mode so that order numbers, confirmation codes, addresses and serials are read back character-for-character instead of being normalised into something that sounds natural but is wrong. For a voice agent that has just taken a booking, that single tag is the difference between a working product and a demo that gets escalated to a human.

Dimension by dimension

• Kind — GPT-Live-1: full-duplex speech-to-speech in one model vs Inworld Realtime TTS-2: a text-to-speech model, with duplex available separately through Inworld's Realtime API

• Price unit — GPT-Live-1: $0.05 per voice minute, billed per second, reasoning backend metered separately vs Inworld Realtime TTS-2: $25 per million characters on demand, $20 on the Creator plan and $12.50 at the top tier, with Flash at $15, $10 and $7

• Independent quality — GPT-Live-1: 70.3% on the Speech to Speech Index, joint sixth of fourteen vs Inworld Realtime TTS-2: Elo 1,244 from 1,091 samples and second on the Artificial Analysis Provider Voice arena, with Flash at Elo 1,211 from 1,626 samples and sixth

• Interruption — GPT-Live-1: native, decided inside the model many times a second vs Inworld Realtime TTS-2: not a property of the TTS model; Inworld's Realtime API supports barge-in with roughly 100 milliseconds of interruption latency, over a single socket that speaks the Open​AI Realtime protocol

• Latency — GPT-Live-1: 0.798 seconds turn-taking latency in Open​AI's own testing, no time-to-first-audio published vs Inworld Realtime TTS-2: 100 milliseconds server-side at the 90th percentile, Flash at 20, with a sub-one-second median to the first audio chunk across the full Realtime API pipeline

• Languages — GPT-Live-1: major languages well served, uneven beyond them vs Inworld Realtime TTS-2: 200-plus locales, with 15 at Tier 1 quality and 79 more at Tier 2, and one voice identity preserved across all of them

• Voices and cloning — GPT-Live-1: twelve presets, no cloning vs Inworld Realtime TTS-2: 282 built-in voices, instant zero-shot cloning from 5 to 15 seconds of authorised audio, professional cloning and full delivery steering on the flagship only

• Deployment — GPT-Live-1: hosted only, one endpoint, no free tier, concurrency capped at 25 to 500 sessions vs Inworld Realtime TTS-2: hosted API plus a gated on-premises container requiring an H100, and a free on-demand tier that includes up to about 70 minutes of synthesis

The duplex question, answered precisely

It would be easy to write this comparison as "one listens, the other only speaks," and it would be wrong in a way that matters. Inworld's text-to-speech models are text-in, audio-out. But Inworld also sells a Realtime API that runs over a single WebSocket, WebRTC or SIP connection and is full-duplex in the sense a builder cares about: it takes semantic voice-activity detection with an adjustable eagerness setting, supports manual turn control, and interrupts on barge-in in roughly 100 milliseconds. It is protocol-compatible with the Open​AI Realtime interface, which makes migration a configuration exercise rather than a rewrite.

What Inworld's Realtime API is not is a native speech-to-speech model. It is a cascade — a swappable speech recognition model, a language model of your choosing, and a synthesiser — orchestrated inside one connection. That is a legitimate architectural choice and it is the same one most production voice agents make. It is simply a different one from GPT-Live-1, where a single model decides when to yield the turn.

The practical difference shows up when the caller interrupts mid-clause. In a cascade, the interruption is detected, the generation is cancelled, and a new turn begins — a sequence that works well and costs you a round trip. In the end-to-end model, the decision was already being made continuously. The former is a system responding quickly. The latter is a system that never stopped listening.

Running the numbers, because the units hide the answer

Speech runs at roughly 150 words a minute, and English averages around five and a half characters per word, so a minute of spoken output is about 800 to 900 characters. At $25 per million, Inworld Realtime TTS-2 produces a minute of speech for roughly two cents. At Flash's $15, it is under a cent and a half.

Against GPT-Live-1's $0.05 per voice minute, that looks like a rout — until you price the rest of what GPT-Live-1 is doing. Inworld's Realtime API still needs speech recognition and a language model, both billed separately, and both carrying their own latency. Fold those in and the cascade's per-minute cost climbs past a nickel for any call that involves a real question. The end-to-end model's premium is not for the voice. It is for the turn-taking, and it is roughly the cost of the speech recognition stage you no longer buy.

Where the cascade wins decisively is volume without conversation. Notification calls, delivery confirmations, appointment reminders, scripted IVR — anything where the text is known before the call starts and nobody is going to interrupt. There, character pricing at $7 to $15 per million is simply cheaper than per-minute duplex, and paying for turn-taking you never use is waste.

There is also a middle path that the two-model framing hides. If you are assembling a cascade anyway, the speech layer does not have to come from a dedicated voice vendor: Open​AI's own TTS models and Goog​le's Gem​ini TTS preview models are ordinary API endpoints, and they are routed. Around 190 models from eleven upstream providers sit behind one OrcaRouter key at provider list price with zero markup — so a synthesis model can sit in the same catalogue, on the same key and the same bill, as the reasoning model it feeds, with automatic failover if one endpoint degrades mid-deployment. It is the least glamorous stage of a voice stack and the easiest one to consolidate. Neither Inworld's Realtime TTS-2 nor GPT-Live-1's Live Sessions endpoint is available that way; those two belong to their vendors.

Which one to pick

Choose Inworld Realtime TTS-2 if the volume is the point and the conversation is scripted. High-throughput agents, notifications, multilingual products where 200 locales and one preserved voice identity matter, or any deployment that needs the on-premises container. You get a top-two arena voice at roughly half the leader's price, a free tier to prototype on, cloning that GPT-Live-1 does not offer at all, and a Realtime API that handles interruption competently in the cascade pattern you probably already run.

Choose GPT-Live-1 if the interruption is the product. Calls that go off-script, callers who talk over the agent, workflows where the turn-taking is the difference between a conversation and a menu. You pay a per-minute rate that does not fall when speech gets cheap, you give up cloning and 200 languages, and you buy back the seams.

The price war in synthesis is real and it is good news for anyone building a voice agent — a top-five voice now costs a fifth of what it did eighteen months ago. It just does not resolve this comparison, because the thing GPT-Live-1 charges for is not the thing Inworld is discounting.