Inworld AI

Inworld AI has changed shape more than almost any vendor in this catalog, and if you knew it as the AI NPC company, you need to update that picture before evaluating it. Founded in 2021 — Kylan Gibbs is co-founder and CEO, with Dialogflow co-founder Ilya Gelfenbeyn among the founders — Inworld raised a $50M Series A in 2022 for AI-driven game characters and roughly $50M more from Lightspeed in 2023 for its character engine. The company then pivoted its center of gravity to realtime voice infrastructure: text-to-speech, speech-to-text, speech-to-speech and LLM inference for consumer-facing applications. The legacy Character Studio — the no-code tool where you authored NPCs with knowledge, triggers, variables and voice — has been retired outright, and its workflows have moved into the Inworld Portal and the Inworld Agent Runtime with Unity and Unreal SDKs (docs.inworld.ai/character-studio-deprecation). In 2026 Inworld acquired Ultravox to accelerate its speech-to-speech research. What is being sold now is a voice and model API stack, and in my testing of the pricing, docs and rate cards, it is a serious one.

What it actually does

The product surface is a credit-based API platform with five self-serve tiers plus enterprise, published at inworld.ai/pricing. On-Demand is $0 pay-as-you-go with up to 70 minutes of TTS included, 100 custom voices, 2 professional voice clones, voice cloning and voice design, speech-to-speech access, 100-plus LLM models via Inworld Inference, a commercial license and community support. Creator is $25 a month in matching credits, Builder $100 (the tier the site marks popular), Developer $300, and Growth $1,500; each tier adds volume discounts, more custom voices, more professional clones, higher concurrency and, at Growth, add-ons for zero data retention, HIPAA and a BAA. Enterprise is custom with price matching, custom limits, 100-plus professional voice clones, an SLA, EU and India data residency, and a named account team. Credits equal your plan price and roll over for three months; overage is charged at your plan's rates.

The rate card itself is the pitch. Realtime TTS-2 runs $25 per million characters on On-Demand and falls to $20, $17.50, $15 and $12.50 across the paid tiers, reaching as low as $5 at enterprise scale; TTS-2 Flash runs $15 down to sub-$5. STT 1 is $0.15 per hour, falling to $0.10. LLM inference starts at $0.07 per million input tokens on open models. On its homepage Inworld frames the comparison directly against named rivals — realtime TTS at Growth rates versus ElevenLabs, realtime STT versus Deepgram — at roughly an order of magnitude per unit. Those are vendor comparisons at vendor-selected tiers, but the published per-unit numbers behind them are checkable, which is more than most voice vendors offer.

The voice features are current-generation: 200-plus languages, instant and professional voice cloning, voice design, steering on Realtime TTS-2, speaking-rate and temperature control, custom pronunciation and timestamps. Concurrency scales from 5 concurrent requests on On-Demand to 500 on Growth, with estimated concurrent user sessions from 20 to 2,000. For character work, the Runtime SDKs for Unity and Unreal expose composable building blocks — knowledge retrieval, safety checks, long-term memory, expressive voice, emotion and relationship modeling — with character templates that recreate what Studio did, but as code and graph data you control rather than a hosted authoring tool.

My testing notes

First observation: the pricing page is the most honest artifact in the voice-API market. The site runs your usage through an estimator — my test case quoted the $25 Creator credit covering roughly 2,500 minutes of TTS-2 Flash audio at about 1,000 characters per minute — and shows the recommended plan against on-demand overage. In my testing of that estimator with a few volume assumptions, the recommendation logic tracked the math instead of steering me upward, and rollover credits soften the use-it-or-lose-it squeeze most credit plans impose. The cost-forecasting caveat is structural rather than dishonest: a single credit bucket spans TTS, STT and LLM tokens at different rates, so a mixed workload needs arithmetic before it needs a subscription.

Second observation: the migration from Studio to Runtime is a real developer lift, and the docs say so. The deprecation guide maps every retired Studio workflow — knowledge, safety, triggers, variables, voice studio, player profiles — to Runtime components and engine templates. Reading it as someone who has watched vendors sunset products silently, this is a well-run deprecation: explicit mapping, runnable Unity and Unreal demos, and a support path. But it is still a rewrite for anyone who authored in Studio, and the no-code authoring audience the original product courted is now outside the target market entirely.

Third: the numbers to verify claims are mostly present. Per-unit rates, concurrency limits, clone counts and compliance scope are published per tier, including the specifics that usually hide — GDPR and SOC 2 Type II on every plan, credit rollover duration, exactly which tier includes HIPAA. When a vendor shows the table, I can hold it to the table. What I cannot verify from outside is audio quality and latency at your workload; the sub-200-millisecond latency and speech-arena rankings circulating in third-party write-ups are plausible but vendor-sourced, so budget a real pilot with your content.

What it does not do

There is no app here — Inworld sells infrastructure. If you want a finished voice agent or a no-code character builder, the retirement of Character Studio means that product is gone; you need engineers with the Runtime SDKs or another vendor entirely. The NPC heritage remains visible in the emotional and relationship modeling, but game developers must now compose characters from primitives rather than configuring them in a studio, and that is a materially different project. On cost: the credit-bucket model makes mixed-workload forecasting harder than a flat per-unit bill, LLM inference is billed at provider cost but through Inworld's stack, and community support on the free tier means production users should expect to pay for priority paths. Enterprise compliance features (ZDR, HIPAA, BAA) are add-ons or tier-gated, not defaults.

There is also roadmap risk worth naming. Inworld has now pivoted from VR characters to NPCs to voice infrastructure, with an acquisition folding Ultravox into the speech stack. Each move has been executed cleanly, but a buyer embedding this in a five-year product should ask directly about the agent runtime's commitment and the pricing history before signing.

Verdict

As it stands today, Inworld is one of the better-value realtime voice and inference APIs on the market, with published unit pricing, meaningful free access and compliance scope that grows with tier. I rate it 3.8: core functionality depth is high across TTS, STT, speech-to-speech and LLM inference with a composable character runtime (40%), value is strong on the published rate card against named rivals (20%), usability is good for a developer platform — estimator, clear tiers, solid docs — but there is no low-code path (20%), the ecosystem is rebuilding around Runtime with Unity, Unreal and Discord support (10%), and reliability signals mix SOC 2 Type II and fresh momentum with the plain fact of two major product pivots (10%). Voice-app developers should put it on the shortlist and pilot it; character-tool seekers need to know the studio they may have read about no longer exists. On price and documentation, Inworld is an easy recommendation to test.

Sources

• https://inworld.ai
• https://inworld.ai/pricing
• https://inworld.ai/resources/what-is-inworld-ai
• https://docs.inworld.ai/character-studio-deprecation
• https://docs.inworld.ai/models
• https://matchfast.co/tool/inworld-ai
• https://themoderntechstack.com/ai-tool/inworld
• https://spatius.ai/blog/inworld-ai-review-2026
• https://aibuildermarketplace.com/b2b/inworld-ai-review
• https://www.bizjournals.com/sanjose/inno/stories/news/2022/08/23/mountain-view-based-startup-raises-50m-series-a.html
• https://www.forbes.com/sites/charliefink/2023/08/02/inworld-adds-another-30-million-for-ai-character-engine
• https://platform.tracxn.com/a/d/company/614c278f49647313ed0d97de/inworld%20ai