Inworld AI review

Inworld AI sells models and inference infrastructure for consumer applications, with a particular emphasis on voice. It is not just a text-to-speech tool: its offerings cover hosted LLM inference, transcription, text-to-speech and speech-to-speech.

For LLM inference, the listed open models include Gemma, GLM, MiMo, DeepSeek and Clef. Realtime TTS-2 Flash is advertised as starting audio in as little as 50 ms. Realtime STT-1 returns streaming transcripts with emotion, style, age and accent traits. Speech-to-speech handles listening, reasoning and replies over a single WebSocket or WebRTC connection.

Voice direction uses plain-language instructions, and Inworld says its voices work in almost 500 languages. A fast inference mode is available on request through the service_tier setting. In short, this is infrastructure for teams building AI interactions into an app or game, rather than a finished consumer chatbot.

The main draw is having speech generation, transcription and conversational audio in one service. A speech-heavy app can use streaming transcripts with traits such as emotion and accent, then generate a spoken reply over WebSocket or WebRTC. That is a more direct fit for live voice interaction than stitching together separate text and audio tools, though the published details do not establish how the parts compare in practice.

There is also a useful range of entry points. The On-Demand plan is aimed at evaluation and prototyping, while paid tiers add monthly credits, more concurrent requests and larger custom-voice allowances. Creator is listed at $25 per month with $25 in credits and 10 concurrent requests; Builder is $100 per month with $100 in credits and 50 concurrent requests. Developer costs $300 monthly and lists 150 concurrent requests plus priority email support.

For teams that need more room, Growth lists 500 concurrent requests and HIPAA and BAA as add-ons. Enterprise offers custom limits and EU and India data residency. Fast inference is available on request, rather than presented as an automatic default. Those options make Inworld worth evaluating when real-time voice is central to the product.

The headline monthly price is not the full cost. Paid plans include credits, but the published pricing details do not publish the rates for additional credits or auto-reload. Overage is charged at the same rate as the plan, yet the specific usage rates are not listed here. We would price expected traffic before treating a tier’s monthly fee as a production budget.

The free On-Demand option is capped: it lists up to 70 minutes of TTS, 100 custom voices, two professional voice clones and five concurrent requests. Creator raises concurrency to 10; even Developer lists 150, so teams expecting heavier simultaneous traffic may need Growth or custom Enterprise terms. Growth is $1,500 per month, and Enterprise pricing is custom.

Some controls also carry unclear costs. HIPAA and BAA and zero data retention are listed as Growth add-ons, but their prices are not published. Annual billing is described as two months free, while annual plan prices are not published. One Reddit thread about game use raised concerns about third-party inference costs, but it did not confirm actual deployments or provide Inworld usage bills. That leaves the core question—what sustained voice traffic costs—unanswered by the listed plan fees alone.

Inworld AI has a pay-as-you-go On-Demand option and four fixed monthly plans before custom Enterprise pricing. On-Demand includes up to 70 minutes of TTS, 100 custom voices, two professional voice clones, and five concurrent requests; usage rates are not published here.

Creator is $25 per month with $25 in monthly credits. Builder is $100 with $100 in credits, and Developer is $300 with $300 in credits. The jump from Creator to Builder is $75 a month; Developer costs another $200 monthly. Growth rises to $1,500 a month and includes $1,500 in credits. Enterprise pricing is custom.

The pricing lists annual-billing equivalents of $20.83, $83.33, $250, and $1,250 per month for those paid tiers, while noting annual billing gives two months free. Confirm the actual billing terms before signing: annual plan prices are not published. Additional credits, auto-reload, HIPAA/BAA, and zero data retention cost extra or have unlisted prices. Overage is charged at the plan rate, but those usage rates are not published.

Switching effort depends on which parts you adopt. Inworld offers hosted open-model inference and speech tools; its speech-to-speech service uses one WebSocket or WebRTC connection. If your app already relies on a different provider’s APIs, plan to adapt the integration and test the connection flow and outputs. The published materials don’t document migration tools or a data-export process.

Before moving production traffic, check what happens to custom voices and professional voice clones if you leave. Inworld’s plans set limits for both, but portability terms are not published in the information available to us. The same caution applies to usage: overages use your plan’s rate, but rates are not listed here, so forecast costs with the current pricing page before committing.

We don’t have Reddit posts or comments in the published material, so we can’t substantiate why users leave or what migration problems they report. Ask Inworld directly about export, voice ownership, and deletion terms before switching.