Inworld AI, Convai and NVIDIA ACE still get filed together as “AI NPC platforms” for video game characters, but by October 2026 they sell three different things. Inworld AI now describes itself as a research lab and inference provider for realtime, consumer-facing AI: voice models, speech recognition and hosted language models sold as APIs, with games as one market among several. Convai sells a complete platform for building conversational 3D characters. NVIDIA ACE is a set of small AI models, Unreal Engine plugins and SDKs that studios use to run characters on a player’s own GeForce RTX GPU or in the cloud.
That difference decides which one fits your studio. Inworld has retired its legacy Character Studio, the hosted authoring tool many developers remember, and now points game teams to its Agent Runtime SDKs for Unity and Unreal Engine. So any comparison that treats Inworld as a hosted character editor is describing a workflow the company has closed.
This guide lines up all three as of October 20, 2026, using each company’s pricing pages, documentation and announcements, plus statements from the studios building with them. It covers what each product is today, what it costs, which games actually use it, and which kind of team should pick which.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Changed at Inworld AI in 2025 and 2026
Inworld was founded in 2021 by three people. Kylan Gibbs is co-founder and CEO, and previously led product for large language models at Google DeepMind. Ilya Gelfenbeyn is co-founder and executive chairman, and earlier co-founded API.AI, the conversational AI platform Google acquired in 2016 and renamed Dialogflow. Michael Ermolenko is co-founder and CTO. The company is based in Mountain View, California.
For game developers, the biggest change is the end of Character Studio. Inworld’s documentation says the legacy Studio, including its characters, scenes, knowledge, triggers, variables, voice studio and player profiles, “is no longer available.” Those workflows moved to the Inworld Portal, which handles workspaces, API keys, billing and observability, and to the Inworld Agent Runtime, where developers build characters inside their own engine or backend.
The Runtime reached game engines a couple of months after it first shipped as a Node.js SDK. Inworld introduced its Unreal AI Runtime SDK on October 29, 2025, putting speech-to-text, text-to-speech and more than 100 language models behind a single API key, alongside a visual graph editor, observability tools and ready-made templates for AI NPCs. A Unity AI Runtime SDK opened in early access in the same announcement. Inworld’s migration guide now sends former Studio users to Character templates in Unity and Unreal, or to a Node.js Voice Agent template for web and backend projects.
Through 2026, Inworld rebuilt its product line around voice and inference for consumer apps. The milestones below come from Inworld’s own announcements, checked against the press release for each one and outside coverage where it exists.
| Date | What Inworld announced | Why it matters to developers |
|---|---|---|
| January 21, 2026 | Inworld TTS-1.5 | An upgrade to its text-to-speech model focused on latency, expression and stability |
| June 10, 2026 | New pricing across the whole stack | Inworld said API prices would “drop by half or more for most developers,” and its Router passes third-party models through at no markup |
| August 31, 2026 | Realtime TTS-2 | Launched at No. 1 on Artificial Analysis’ Controlled Voice board, with pacing, tone and emotion steered by plain-language instructions; it ranked sixth there by October 20 |
| September 30, 2026 | Acquisition of Ultravox, a Seattle voice-agent platform | Brings the Ultravox team’s speech-to-speech research to Inworld; Ultravox customers’ built-in Inworld voices moved to Realtime TTS-2 |
| October 14, 2026 | Inworld Inference | Inworld-hosted open models through an OpenAI-compatible API, with a Realtime Router that can also reach more than 100 third-party models |
Gibbs framed the October launch around cost. “Cost is the first thing standing in these developers’ way, so it is the first thing we are taking down,” he said in the October 14 Business Wire release, which announced the service under the name Realtime Inference. “But a consumer app does not succeed on cheap inference alone.”
Games haven’t disappeared from the pitch. “Games and media” is one of five solution areas on Inworld’s site, described as “voices and live dialogue for games and narrated media,” and the company’s FAQ answers the question “Is Inworld only for gaming?” with a plain no. What changed is the layer Inworld sells: the voice, transcription, model serving and routing underneath a character, rather than a finished character product.
AI Character Platforms in 2026: What a Studio Actually Buys
Every AI character runs on the same basic loop. Speech recognition turns what the player says into text, a language model decides what the character says or does, text-to-speech turns the reply into audio, and an animation layer can sync the face to that audio. Around the loop sit memory, knowledge retrieval, safety filters and the game logic that decides what the character is allowed to do.
In 2026 there are three ways to buy that loop:
- Buy the parts and build the character yourself. Inworld sells the voice, transcription and model-serving layers, plus engine SDKs and templates that wire them together. Your team owns the character logic.
- Buy a complete character platform. Convai bundles the character’s backstory, voice, lip-sync, knowledge, perception and actions into one hosted service, with plugins for common engines.
- Run the character on the player’s machine. NVIDIA ACE supplies small speech, language and animation models built to run on GeForce RTX GPUs. There’s no server bill per conversation, but the character has to fit in the GPU memory the game leaves free.
None of these is strictly better. The right pick depends on how much character logic you want to own, where your players run the game, and whether a per-conversation cloud bill or a per-player hardware requirement is easier for you to live with.
How Each Platform Works Today
Inworld AI: Voice and Inference APIs With Engine SDKs
Inworld’s site now sorts its products into two groups. Voice covers Realtime TTS (the Realtime TTS-2 and TTS-2 Flash voice models), Realtime STT for streaming speech-to-text and a speech-to-speech service. Inference, launched on October 14, 2026, serves hosted open models through one API, routes across them with automatic failover through the Realtime Router, and offers a dedicated deployment for a customer’s own model.
The speech-to-speech service runs a whole voice conversation over one WebSocket or WebRTC connection, lets developers choose the language model for each session, and is compatible with OpenAI’s Realtime API, so existing clients can move over with few changes. Realtime STT-1 adds voice traits to transcripts: emotion, style, age and accent, and since October 1, 2026 it can label who is speaking as a $0.02-an-hour add-on.
On the model side, Inworld Inference serves open models on a fleet Inworld operates, including GLM 5.3 Flash, Gemma 4 31B, Gemma 4 26B, MiMo 2.6 Pro and DeepSeek V4.1 Flash, through an OpenAI-compatible API. The Router reaches more than 100 language models from providers including OpenAI, Anthropic, Google, xAI, Meta and Mistral, with automatic failover and A/B testing built in. Since October 9, 2026 the same API also serves video generation and decision models, which takes Inworld’s catalog to 127 models.
For games, the entry points are the Unity AI Runtime SDK and the Unreal Agent Runtime SDK. Inworld’s Unreal Character template covers character and player components, profiles, dialogue and emotion graphs, voice and lip-sync, and the Unreal launch also shipped MetaHuman and Lipsync templates. Inworld’s voices are also available inside voice-agent frameworks such as LiveKit, Vapi and Pipecat.
Convai: A Complete Conversational Character Platform
Convai was founded by Purnendu Mukherjee, its CEO, who was previously a senior deep learning engineer in NVIDIA’s conversational AI organization. The San Jose company builds what it calls spatially aware 3D characters: a developer defines a character’s backstory, personality, knowledge and voice, then deploys it into a game, simulation, website or XR experience.
Its character tools cover backstories, more than 60 languages, more than 500 voices, a Knowledge Bank for uploaded documents, narrative design and in-world actions. Characters can perceive their surroundings through vision, either from a webcam or from a virtual game environment. Convai also supplies lip-sync, facial animation and a library of animations, and developers can record custom animations in its VR animation app.
Two hosted tools sit on top: Avatar Studio for cloud-rendered avatars and Convai Sim for simulations, both metered in monthly minutes. Convai’s site organizes the business around three use cases, learning and training, brand representatives, and social worlds and gaming, and lists ISO 27001 compliance and on-premises deployment for enterprise buyers.
Convai’s 2026 releases have gone into developer tooling rather than new plans. MCP servers for Unreal Engine (July 2026) and Unity (August 2026) let coding agents such as Codex, Claude Code and Cursor set up a Convai character from a typed prompt, and a Scene Tagger tool added in September 2026 describes the objects in an Unreal level so characters know what’s around them.
NVIDIA ACE: Models, Plugins and SDKs for On-Device Characters
NVIDIA’s ACE for Games page describes ACE as “a suite of AI technologies designed to help middleware and game developers build knowledgeable, actionable, and conversational in-game characters,” with cloud and on-device models for speech, intelligence and animation. It isn’t one product. A studio picks the pieces it needs.
The current pieces include Nemotron ASR for speech recognition, Resemble AI’s Chatterbox TTS and Alibaba’s Qwen 3 TTS for speech, NVIDIA Nemotron small language models in 4B and 9B sizes plus Qwen3.5 4B for dialogue and decisions, and Audio2Face-3D and Audio2Emotion for facial animation. In September 2026, NVIDIA highlighted two speech additions: Nemotron Speech 3.5 Streaming, a 600M-parameter speech recognition model built for low latency, and Qwen3 TTS with support for custom fine-tuning. On October 13, 2026, NVIDIA said its next ACE language model, Nemotron 3.5 Bolt E4B, a 4GB model it calls the third generation of its game-character models, is due on October 23.
Two SDKs tie the models into a game. The Game Agent SDK, still in beta, is an open-source C/C++ framework with Agent, Chat and RAG APIs for connecting characters to game state, knowledge and actions. The In-Game Inferencing SDK runs local models inside the game process and schedules AI work alongside graphics; its September 2026 update added Gemma 4 support, a Stable Diffusion plugin and early support for NVIDIA’s RTX Spark platform.
Unreal Engine 5 teams get ready-made plugins for speech recognition, a local language model with function calling, text-to-speech and Audio2Face-3D, under the MIT license. Studios on other engines work with the SDKs directly.
Full Specs Comparison: Inworld AI vs Convai vs NVIDIA ACE
The table lines up what each company documents as of October 20, 2026. Where a company doesn’t publish a figure, the table says so rather than estimating.
| Specification | Inworld AI | Convai | NVIDIA ACE |
|---|---|---|---|
| What it is | Realtime voice (TTS, STT, speech to speech) and LLM inference APIs for consumer apps, including games | Conversational 3D character platform for games, XR, training and brand avatars | Suite of speech, language and animation models, Unreal plugins and SDKs for in-game characters |
| Where it runs | Inworld’s cloud, with voice served from US, Europe, India and Asia Pacific regions and LLM serving in the US; Enterprise adds EU and India data residency | Convai’s cloud; Enterprise adds an on-premises option | On the player’s GeForce RTX GPU, or as cloud-hosted models |
| Speech | Realtime TTS-2, Realtime TTS-2 Flash, Realtime STT-1 | Built-in speech and 500+ voices; ElevenLabs voices on paid plans | Nemotron ASR (including Nemotron Speech 3.5 Streaming), Chatterbox TTS, Qwen3 TTS |
| Language models | 100+ third-party LLMs through the Router; hosted open models such as GLM 5.3 Flash, Gemma 4 and MiMo 2.6 Pro; video and decision models added in October 2026 | Switchable core LLMs; flagship-model caps list Gemini 1.5 Pro, GPT-4o and Claude 3.5 Sonnet | Small on-device models: Nemotron 4B and 9B, Qwen3.5 4B; Nemotron 3.5 Bolt E4B due October 23, 2026 |
| Facial animation | Word and character timestamps for lip-sync; Unreal Character template with voice and lip-sync built in | Built-in lip-sync and facial animation | Audio2Face-3D and Audio2Emotion |
| Engines | Unity AI Runtime SDK, Unreal Agent Runtime SDK, Node.js templates | Unity, Unreal Engine, three.js, PlayCanvas | Unreal Engine 5 plugins (5.4 to 5.7); other engines through the Game Agent SDK |
| Player hardware | None (server-side) | None (server-side) | A recent GPU with spare VRAM for on-device mode: the Game Agent SDK needs an NVIDIA Ampere or newer card, while NVIDIA lists the ACE models as compatible with other GPUs and CPUs |
| Named game use | Narrator and NPC voices in Latitude’s Voyage (Inworld case study); Streamlabs assistant built with NVIDIA | Site lists use cases, not shipped game customers | inZOI (early access) and NARAKA: BLADEPOINT MOBILE PC VERSION shipped; PUBG Ally public beta; Aniimo due early 2027; others announced |
| Entry price | On-Demand, $0 pay as you go; Creator $25 a month | Free plan; Indie Dev $29 a month | No subscription; costs are engineering time and player GPU memory |
| Compliance | SOC 2 Type II and GDPR on every plan; HIPAA, BAA and zero data retention add-ons from Growth | ISO 27001 listed; Enterprise data ownership, SLAs and isolated deployment | On-device mode keeps player speech on the player’s machine |
| Company | Founded 2021, Mountain View; more than $125 million raised (company figure) | Founded 2022, San Jose; about $5 million in seed funding (startup databases) | Part of NVIDIA; ACE debuted in 2023 |
Two splits stand out. Inworld and Convai run in the cloud, so the player needs no special hardware, but every conversation is a metered call. NVIDIA ACE moves inference onto the player’s GPU, which removes the per-conversation bill but adds a hardware floor: KRAFTON’s PUBG Ally runs on NVIDIA cards with at least 8GB of VRAM.
The other split is ownership. With Inworld you assemble the character from SDK templates and APIs and keep the logic in your engine. With Convai the character lives on Convai’s platform. With ACE the models ship inside your game build, each under its own license.
Pricing Breakdown: What Each Platform Costs in October 2026
Inworld and Convai both publish full price lists. NVIDIA ACE’s on-device stack has no subscription at all, so its costs show up elsewhere; only its cloud microservices carry a paid tier, through NVIDIA AI Enterprise.
Inworld AI Pricing
Inworld sells monthly plans that come back as usage credits. The $300 Developer plan, for example, grants $300 of credit a month, spent across text-to-speech, speech-to-text, speech-to-speech and LLM inference at that tier’s rates. Unused credits roll over for up to three months, and yearly billing gives two months free.
| Plan | Monthly price | Realtime TTS-2 (per 1M characters) | Realtime TTS-2 Flash (per 1M characters) | Realtime STT-1 | Concurrent requests |
|---|---|---|---|---|---|
| On-Demand | $0, pay as you go | $25 | $15 | $0.15 an hour | 5 |
| Creator | $25 | $20 | $10 | $0.10 an hour | 10 |
| Builder | $100 | $17.50 | $9 | $0.10 an hour | 50 |
| Developer | $300 | $15 | $8 | $0.10 an hour | 150 |
| Growth | $1,500 | $12.50 | $7 | $0.10 an hour | 500 |
| Enterprise | Custom | As low as $5 (pricing page); $10 (home page) | Under $5 (pricing page); $5 (home page) | Custom | Custom |
Open models on Inworld Inference bill per token from the same credits, starting at $0.07 per million input tokens for Gemma 4 26B, and cached input costs up to 99% less. MiMo 2.6 Pro, for example, lists $0.43 per million input tokens and $0.0036 for cached input. The free On-Demand tier includes up to 70 minutes of TTS or 400 minutes of STT. Inworld’s own worked example: at about 1,000 characters a minute, an hour of Realtime TTS-2 Flash speech costs about $0.42 on the Growth plan. Inworld’s pages differ on Enterprise voice rates: the pricing page says TTS-2 goes as low as $5 per million characters, while the home page quotes $10 for TTS-2 and $5 for TTS-2 Flash.
Convai Pricing
Convai’s pricing page charges by “interactions,” counting one each time a user sends input and the character responds. Every plan includes the character creation tools and vision; paid plans add long-term memory, character versioning and more monthly active end users.
| Plan | Monthly | Billed yearly (per month) | Interactions a month | Avatar Studio and Convai Sim minutes | Concurrent character sessions | Monthly active end users |
|---|---|---|---|---|---|---|
| Free | $0 | $0 | 100 | 60 | 1 | 1 |
| Indie Dev | $29 | $22 | 3,000 | 600 | 1 | 75 |
| Professional | $99 | $69 | 10,000 | 2,400 | 3 | 250 |
| Scale | $499 | $299 | 50,000 | 10,200 | 15 | 1,250 |
| Business | $1,199 | $499 (limited-time offer) | 125,000 | 21,000 | 30 | 2,500 |
| Enterprise | Custom | Custom | Custom | Custom | Custom | Custom |
Two limits matter in practice. Interaction quotas reset each billing cycle and don’t roll over, and each plan caps how many interactions can use flagship language models: 1,500 of the Indie Dev plan’s 3,000, for example. Enterprise adds data ownership, SLAs, a separate deployment isolated from the developer tiers, and an option to run Convai’s entire backend in your own data center.
NVIDIA ACE Pricing
ACE’s on-device stack has no subscription and no per-conversation fee; the only NVIDIA charge sits on the cloud side, where its production-branch NIM containers and enterprise support need an NVIDIA AI Enterprise subscription. The SDKs, plugins and models are downloads from NVIDIA’s developer site, each under its own license: the Unreal Engine plugins and the Audio2Face-3D SDK under MIT, the Audio2Face-3D models under the NVIDIA Open Model License, and the Audio2Face-3D training framework under Apache. NVIDIA describes the Game Agent SDK as open source.
The real costs are engineering time and GPU memory, because local models compete with the game for VRAM. KRAFTON quantized PUBG Ally’s 2B-parameter language model so it fits in the headroom PUBG leaves on 8GB cards, and an experimental build of Creative Assembly’s Total War: PHARAOH advisor used about 6GB of VRAM on its own, according to OC3D.net figures reported by heise online.
| Tier | Inworld AI | Convai | NVIDIA ACE |
|---|---|---|---|
| Free | On-Demand: $0, pay as you go, up to 70 minutes of TTS | Free: 100 interactions a month | SDKs, plugins and models to download |
| Entry paid | Creator: $25 a month | Indie Dev: $29 a month | No paid tier |
| Mid tier | Builder: $100 a month; Developer: $300 a month | Professional: $99 a month; Scale: $499 a month | No paid tier |
| Top listed plan | Growth: $1,500 a month | Business: $1,199 a month | No paid tier |
| Enterprise | Custom; TTS-2 from $5 to $10 per 1M characters, depending on which Inworld page you read | Custom quote, with an on-premises option | On-device: engineering time and player GPU memory; cloud NIM in production: NVIDIA AI Enterprise |
Latency and Benchmark Claims: What’s Measured, and by Whom
We couldn’t find an independent benchmark that tests Inworld, Convai and NVIDIA ACE head to head on the same character workload. Every figure below is either the company’s own number, measured under conditions it chose, or a reading from Coval’s public TTS board, and none of them was taken on the same character workload.
Inworld publishes the most numbers. It says Realtime TTS-2 reaches first audio in 100 milliseconds at P90, measured on its own servers before network time, and that Realtime TTS-2 Flash posted a 58 ms median time to first audio on Coval’s public TTS board, against 183 ms for ElevenLabs Flash v2.5, though both are each model’s best day from October 6 to 12, 2026. Coval’s rolling 30-day averages on October 20 were 181 ms for TTS-2, 84 ms for TTS-2 Flash and 194 ms for ElevenLabs Flash v2.5. On reliability, Inworld says GLM 5.3 Flash completed more than 20 million requests over one 30-day period with a 99.9995 percent completion rate.
KRAFTON’s numbers for PUBG Ally describe design choices rather than one benchmark. Its engineers told NVIDIA that cloud-hosted language models felt too slow for live squad chat once network and inference latency were added together. So they moved to a 2B-parameter Mistral-NeMo-Minitron model running on the player’s GPU and split the work in two: a behavior tree handles movement, aiming and combat at the game’s tick rate, while the language model handles intent, coordination and speech.
Convai doesn’t publish latency figures on its home or pricing pages. Its home page lists low-latency responses as a feature without a number, but its 2026 blog posts cite response latency under 200 ms over WebRTC for its Unreal plugin and Web SDK, with real-world numbers depending on the connection.
For a studio, the practical difference is model routing versus owning. Inworld’s Router lets a team switch the model behind a character, or split traffic between models for A/B tests, without changing game code. NVIDIA’s on-device models are small enough to share a gaming GPU, which limits how much reasoning each reply gets but removes network latency and the per-call bill. Convai sits in between: a managed backend where you choose among its supported core models.
Supported Game Engines and SDKs
Convai has the broadest documented engine reach. It offers Unity and Unreal Engine plugins, supports web builds through three.js and PlayCanvas, and publishes tutorials for adding its characters to MetaHuman projects in Unreal. Its site describes the platform as “platform, device, and OS agnostic.”
NVIDIA’s support is deepest in Unreal Engine 5. Its Audio2Face-3D plugin is available for Unreal Engine 5.4, 5.5 and 5.6, and its speech recognition, language model and text-to-speech plugins for 5.5, 5.6 and 5.7, with Blueprint and C++ support. There’s also an Autodesk Maya plugin for offline facial animation. Unity and custom-engine studios work through the C/C++ Game Agent SDK and the In-Game Inferencing SDK instead of drop-in plugins.
Inworld now covers both major engines with its Runtime SDKs, each with a Character template, plus a Node.js Voice Agent template for backend and web projects. Its voice and model services are network APIs, reached over WebSocket, WebRTC or an OpenAI-compatible endpoint, so engines without an official SDK can still call them. If you’re starting from scratch, our guide to building an AI NPC in Unity walks through the basic loop.
Who Is Actually Shipping: Named Games and Customers
NVIDIA’s ACE for Games page names five partner games and the Streamlabs assistant, and NVIDIA has announced two more games separately: NARAKA: BLADEPOINT MOBILE PC VERSION, which shipped its AI teammate in March 2025, and Aniimo, a creature-collecting RPG due to add ACE in early 2027. They’re at very different stages. Here’s where each stood when we checked on October 20, 2026.
| Game | Developer | ACE feature | Status |
|---|---|---|---|
| inZOI | KRAFTON | “Smart Zoi” NPCs that plan, act and reflect, driven by an on-device Mistral NeMo Minitron small language model | Experimental setting since early access began on March 28, 2025, first on GeForce RTX cards; the game is still in early access, with 1.0 targeted by March 2027 |
| NARAKA: BLADEPOINT MOBILE PC VERSION | NetEase | AI teammate that fights alongside you, finds items, swaps gear and suggests skills | Added March 27, 2025, in the China-exclusive version |
| PUBG: BATTLEGROUNDS | KRAFTON | PUBG Ally, a “Co-Playable Character” named Ella, in the Ally Duo mode | Two-week public beta in Arcade mode from June 17, 2026; we found no announcement of a permanent release |
| Total War: PHARAOH | Creative Assembly | AI advisor that answers strategy questions using more than 1,200 game data tables | Shown at CES 2026; experimental, with a playtest program planned for 2026 |
| MIR5 | Wemade Next | An AI boss that learns from players’ past tactics | Announced in January 2025; Wemade’s site still lists MIR5 as upcoming, and we couldn’t confirm a release |
| Dead Meat | Meaning Machine | A murder suspect players can question in their own words | Unreleased; its Steam listing gives a 2026 release window |
So two ACE character features reached ordinary players in March 2025: Smart Zoi in inZOI, which is still in early access, and the NARAKA teammate. PUBG Ally reached players as a time-limited beta, and the rest are demos or unreleased games. KRAFTON says feedback from more than a thousand players shaped Ally, and its beta announcement quotes Chief AI Officer Kangwook Lee calling the mode “an early step toward exploring how AI companions can create new ways to play.”
NVIDIA’s page also lists a Streamlabs streaming assistant built with Inworld AI, described as a producer, technical assistant and 3D sidekick for streamers.
Inworld’s clearest game customer is Latitude, the studio behind AI Dungeon. In Voyage, Latitude’s AI-native RPG, which is now in open beta on the App Store, Google Play and the web with no waitlist, Inworld runs all eight narrator voices and the dynamic NPC voice cast, about 97% of Voyage’s text-to-speech volume, with more than 15 million TTS calls in six weeks. Inworld-hosted models also run Voyage’s World Engine tasks, the skill checks, HP, inventory and memory rules that sit under the story. Inworld puts the cost of narrating every story beat at about 6% of Voyage’s total AI spend. Those figures come from Inworld’s own case study, so treat them as vendor-reported.
Outside games, Inworld’s named customers are consumer apps: Status by Wishroll, which says it scaled from prototype to one million users in 19 days, plus Bible Chat, Talkpal, Luvu and Particle.
Convai’s home and pricing pages describe use cases rather than naming game customers. Its best-known public demo came at NVIDIA’s CES 2024 press conference, where Fast Company reported unscripted conversations with its characters, though with “a minor delay” before they answered.
What the People Building These Tools Are Saying
Inworld’s Kylan Gibbs now talks about voice and consumer economics rather than NPC dialogue trees. When Inworld bought Ultravox, he said: “Ultravox has built a platform that developers rely on for real-time voice agents. We’re excited to welcome members of the Ultravox team and continue building the platform. Bringing Realtime TTS-2 to the Inworld voices customers already use is the first benefit we’re delivering.”
Status CEO Fai Nur gave the customer’s side in the same release: “We scaled from prototype to 1 million users in 19 days with over 20x cost reduction.”
At Ubisoft, the writers behind the company’s generative AI experiments describe their job as setting limits rather than writing every line. “I still write the story and character personalities, but instead of fixed lines, we create these kinds of fences that let NPCs improvise within the world but stay within the boundaries of the lore and motivations we have given them,” narrative director Virginie Mosser said in November 2025.
Convai’s Purnendu Mukherjee described the goal in a January 2024 interview with Fast Company as “giving these NPCs a mind,” with characters that “can perceive the environment and interact with it” in 3D scenes. That perception-first pitch still shapes Convai’s product: vision and in-world actions are included on every plan, free tier included.
Ubisoft’s NEO NPC and Teammates: Where AAA Experiments Stand
Ubisoft has been the most visible big publisher testing generative characters in public. At GDC in March 2024, a small R&D team at Ubisoft’s Paris studio showed NEO NPC, built with NVIDIA’s Audio2Face and Inworld’s large language model. There were three voice-driven demos: building a relationship with a resistance fighter named Bloom, Bloom reacting live to a drone mission, and planning a stealth infiltration with a character named Iron.
“We didn’t want to rush anything, and today is just the first experiment,” project director Xavier Manzanares said at the time, according to GamesIndustry.biz.
On November 21, 2025, Ubisoft followed up with Teammates, which it calls its first playable generative AI research project. Players take the role of a resistance member searching an enemy base for five missing teammates, helped by an AI voice assistant named Jaspar and two AI squadmates, Pablo and Sofia, who take voice orders on when, how and whom to attack. Jaspar can also highlight enemies, explain lore, change settings and pause the game.
Teammates is still research, but Ubisoft is putting more money behind it. Its engineers showed the project at NVIDIA’s GTC conference in March 2026, running on a high-end GeForce RTX card with a small language model rather than in the cloud, and the company’s May 20, 2026 annual results said it is “accelerating investments behind Teammates, its first playable Generative AI experience.” Ubisoft says a few hundred players have tried it in a closed playtest, and it hasn’t tied the technology to a released game. That’s a useful reality check: more than two years after NEO NPC, the most visible AAA project in this space is still a playable prototype.
Use Cases: Which Platform Fits Which Studio
Solo developers and small teams that want a talking character this week. Convai is the shortest path. The free plan includes 100 interactions a month, Indie Dev costs $29 a month for 3,000, and the Unity and Unreal plugins ship the voice, lip-sync and knowledge pieces together.
Studios that want to own their character logic. Inworld fits teams that would rather design the character in their own engine and buy the parts: Realtime TTS-2 voices, streaming transcription, hosted models and a router across more than 100 LLMs, wired together with the Unity or Unreal Runtime templates.
Voice-heavy games and AI-native RPGs. When a game narrates constantly, text-to-speech becomes a major cost line. Inworld’s per-character pricing and Latitude’s experience with Voyage, where Inworld’s case study puts narration at about 6% of total AI spend, make it the first quote to get.
PC games with a GeForce RTX audience. NVIDIA ACE is the only one of the three built to run the whole loop on the player’s machine, with no per-conversation server cost. Budget for VRAM, and plan a fallback for players without a capable card: the Game Agent SDK needs an NVIDIA Ampere or newer GPU.
Unreal Engine 5 studios. All three support Unreal. ACE’s Unreal plugins bundle ready-to-use local models, including an English speech recognition model, Qwen 3.5 4B and Chatterbox Turbo voices, while Inworld’s Unreal Character template builds in voice and lip-sync.
Web, XR and training projects. Convai’s three.js and PlayCanvas support, Avatar Studio and Convai Sim make it the most direct route to a browser-based or simulation character.
Regulated or compliance-heavy teams. Inworld includes SOC 2 Type II and GDPR on every plan and offers HIPAA with a BAA, plus zero data retention, from the Growth plan. Convai lists ISO 27001 and on-premises deployment. ACE’s on-device mode keeps player speech on the player’s own machine.
Migration Guide: Moving Off Legacy Inworld Studio, or Between Platforms
Teams that built characters in Inworld’s old Studio have a documented path. Inworld’s guide maps each Studio concept to its Runtime equivalent:
| Studio concept | Where it lives now |
|---|---|
| Characters and scenes | Character templates in Unity, Unreal or Node.js |
| Knowledge | Knowledge primitives |
| Voice studio and pronunciations | Inworld TTS, voice cloning and custom pronunciation |
| Triggers, variables and integrations | The Runtime graph plus your own application logic |
| Player profiles | Player components in the Unity and Unreal Character templates |
The same steps work when you move between any of these platforms, because they all wrap the same loop of speech in, reasoning and speech out.
Step 1: Separate what’s platform-specific. Split your integration into character state and memory, the conversation runtime, and presentation (voice, lip-sync and animation). Most migration work sits in the last two.
Step 2: Export character definitions in a neutral format. Save prompts, backstories, knowledge documents and behavior rules as plain text or JSON before you touch the new SDK, so you aren’t rebuilding a character’s voice from old build logs later.
Step 3: Rebuild one character in a blank test scene. Inworld and Convai expect a network call per turn, while ACE’s on-device path expects local inference. That’s a bigger architectural change than swapping one cloud vendor for another, so prove it on a single character first.
Step 4: Rebuild voice and animation separately from dialogue. Moving to ACE means wiring Audio2Face-3D and a local voice model yourself. Moving from ACE to a hosted platform trades some animation control for simplicity.
Step 5: Model the costs before you commit. Inworld bills usage credits, Convai bills interactions and avatar minutes, and ACE costs engineering time and player VRAM. Run your expected daily players and conversations through each model before you lock in a build.
Step 6: Roll out behind a flag. Move a subset of characters or players first, and compare latency, dialogue quality and cost in production before you retire the old integration.
Pros and Cons of Each Platform
Inworld AI
- Pros: Published price list with discounts that grow with volume; Realtime TTS-2 launched at No. 1 on Artificial Analysis’ Controlled Voice board, though by October 20 it had slipped to sixth there; one Router across more than 100 LLMs; SOC 2 Type II and GDPR on every plan; Unity and Unreal Runtime SDKs with Character templates.
- Cons: Games are one market among several rather than the company’s focus; the hosted Studio workflow is gone, so Studio-era projects need rebuilding in the Runtime; cloud usage costs scale with player count.
Convai
- Pros: Complete character platform with perception, actions and lip-sync; the widest documented engine and web reach of the three; a free plan and a $29 entry plan; an on-premises option for enterprise.
- Cons: Interaction caps, plus per-plan limits on flagship models and monthly active users; the flagship models its pricing page names (Gemini 1.5 Pro, GPT-4o, Claude 3.5 Sonnet) are older generations; its site lists use cases, not shipped game customers.
NVIDIA ACE
- Pros: Runs on the player’s GPU with no per-conversation fee; ACE characters have shipped in two games players can buy today, inZOI (still in early access) and NARAKA: BLADEPOINT MOBILE PC VERSION; MIT-licensed Unreal plugins and an open-source Game Agent SDK; the deepest Unreal Engine 5 tooling of the three.
- Cons: Built for NVIDIA RTX cards with spare VRAM, since its Game Agent SDK requires an Ampere or newer NVIDIA GPU, though NVIDIA now lists the ACE models themselves as compatible with other GPUs and CPUs; several listed partner features are still demos or tests; more engineering to assemble; its ready-made game-engine plugins are Unreal-only.
Security, Licensing and Data
Inworld and Convai are hosted services: you license access, not model weights. Inworld documents the most compliance detail. Every plan carries SOC 2 Type II and GDPR, HIPAA with a BAA and zero data retention are add-ons from the Growth plan, and Enterprise adds an SLA, a DPA and EU and India data residency.
Convai’s Enterprise plan says customers own and control their data, including all input and output data, and runs Enterprise traffic on a separate deployment with SLAs. Its site also lists ISO 27001 compliance, AI safety guardrails and on-premises deployment.
NVIDIA’s model is different because the software ships to you. The Unreal plugins and the Audio2Face-3D SDK are MIT-licensed, the Audio2Face-3D models use the NVIDIA Open Model License, and third-party models such as Qwen or Chatterbox carry their own terms, so legal teams need to check each model a game redistributes. On-device inference also changes the privacy picture. In PUBG Ally, speech recognition, the language model and text-to-speech all run locally on the player’s RTX GPU, so the conversation doesn’t need a round trip to a server.
The Ethics Debate: Voice Actors, Unions and Player Trust
Generative characters arrived in the middle of a labor fight over AI in the game industry, and that fight now has a contract. SAG-AFTRA’s video game performers went on strike in July 2024 after talks broke down over consent and pay for AI replicas. In July 2025, 95.04% of voting members approved the 2025 Interactive Media Agreement, ending a strike that lasted about 11 months.
The union and the game companies said the deal includes “consent and disclosure requirements for A.I. digital replica use and the ability for performers to suspend consent for the generation of new material during a strike.” For studios, any character voice built from a real performer needs that performer’s documented consent, and the voice pipeline should be able to switch a replica off.
Player trust is the other half. A character that can say anything can also say the wrong thing, which is why Ubisoft’s writers talk about building “fences” around improvisation, and why KRAFTON limited PUBG Ally to a single map and a single mode, with a fixed list of items, so its answers stay grounded in the match. Budget time for moderation and stress-testing, not just integration.
Funding and Where the Market Is Heading
Inworld says it has raised more than $125 million from investors including Lightspeed Venture Partners, Kleiner Perkins, Founders Fund, CRV, Stanford University, Intel Capital, Microsoft M12, Meta, Samsung NEXT, LG Technology Ventures and Bitkraft. PitchBook figures cited by GeekWire in October 2026 value the company at $515 million, with $119.8 million raised in total.
Convai’s funding is much smaller. Startup databases list a single $5 million seed round in 2022, and we found no larger announced round. NVIDIA funds ACE itself, and since ACE debuted in 2023 it has kept adding models and tooling: new speech models in September 2026, then a third-generation small language model for ACE characters, Nemotron 3.5 Bolt E4B, announced on October 13 for release on October 23.
Each company’s 2026 moves point in a clear direction. Inworld is betting on voice and low-cost inference for consumer apps, games included. NVIDIA is betting on small models that share the player’s GPU, with partners like KRAFTON, which said in October 2025 that it was becoming an “AI First” company and in April 2026 released four open multimodal models of its own under the Raon name, which it plans to fine-tune for individual games. Convai is betting on embodied characters for training, brands and virtual worlds as well as games.
The consumer end of the market got a warning this month. Character.AI, one of the largest AI companion apps, filed for Chapter 11 on October 12, 2026, with Disney offering $20 million for a license to some of its intellectual property.
The Verdict: Which Platform Should Your Studio Pick in 2026
There’s no single winner, because these are no longer three versions of the same product. If you want a finished, talking 3D character with the least engineering, start with Convai. If you want to own the character design and buy a voice model that launched at No. 1 on Artificial Analysis’ Controlled Voice board and still sits in its top ten, plus low-cost model serving underneath it, Inworld is the stronger choice, and its Unity and Unreal Runtime templates give you a head start. If your players are on GeForce RTX PCs and you want no per-conversation cost, NVIDIA ACE is the only one of the three built for that, and its autonomous characters are already in two games players can buy, one of them still in early access.
The bigger shift is architectural. Inworld has moved down the stack to voice and inference, Convai has stayed at the character layer, and NVIDIA is pushing the whole loop onto the player’s hardware. Pick the layer you want to own, then pick the vendor.
Frequently Asked Questions
Does Inworld AI still offer Character Studio?
No. Inworld’s documentation says the legacy Character Studio has been retired. Its workflows moved to the Inworld Portal and the Inworld Agent Runtime, with Character templates for Unity and Unreal and a Node.js Voice Agent template.
Is Inworld AI still used for games?
Yes, though games are now one market among several. Inworld lists games and media as a solution area, ships Unity and Unreal Runtime SDKs, and, according to its case study, runs the narrator and NPC voices in Latitude’s Voyage, which is now in open beta with no waitlist.
How much does Inworld AI cost?
Plans run from the free, pay-as-you-go On-Demand tier to Creator at $25 a month, Builder at $100, Developer at $300 and Growth at $1,500, with custom Enterprise pricing. On demand, Realtime TTS-2 costs $25 per million characters and TTS-2 Flash $15; the Growth plan lowers those to $12.50 and $7.
How much does Convai cost?
Convai has a free plan with 100 interactions a month, then Indie Dev at $29 a month, Professional at $99, Scale at $499 and Business at $1,199, with lower prices for yearly billing and custom Enterprise pricing.
Is NVIDIA ACE free?
There’s no subscription or per-conversation fee for the on-device path: the Unreal plugins, Game Agent SDK and models are free downloads under licenses such as MIT, Apache 2.0 and the NVIDIA Open Model License. The cloud models ship as NIM microservices, free to prototype with on build.nvidia.com under rate limits, while the enterprise production branch needs an NVIDIA AI Enterprise subscription, which NVIDIA’s NIM FAQ prices at $4,500 per GPU per year. You pay in engineering time, and on-device characters need a recent GPU with spare video memory: NVIDIA’s Game Agent SDK requires an Ampere or newer NVIDIA card, though NVIDIA now lists the ACE models themselves as compatible with other GPUs and CPUs.
Which games use NVIDIA ACE today?
inZOI’s Smart Zoi and the AI teammate in NARAKA: BLADEPOINT MOBILE PC VERSION reached players in March 2025. PUBG: BATTLEGROUNDS ran PUBG Ally as a two-week public beta from June 17, 2026. The Total War: PHARAOH advisor, MIR5’s AI boss and Dead Meat were still experimental, unconfirmed or unreleased when we checked, and Aniimo is due to add ACE in early 2027.
Which platform has the lowest latency?
We couldn’t find an independent head-to-head test. Inworld cites 100 milliseconds to first audio at P90 for Realtime TTS-2 on its own servers, and a best-day 58 ms median for TTS-2 Flash on Coval’s public board; NVIDIA’s argument is that on-device inference removes the network round trip; Convai’s 2026 blog posts cite sub-200 ms response latency over WebRTC for its Web SDK and Unreal plugin, though its home and pricing pages give no number.
Can AI characters replace voice actors?
Not without consent. SAG-AFTRA’s 2025 Interactive Media Agreement requires consent and disclosure for AI digital replicas, and performers can suspend consent for new material during a strike.
Which AI character platform should an indie developer pick?
Convai’s free plan and $29 Indie Dev tier, with Unity and Unreal plugins, are the fastest way to a working prototype. If you mainly need strong voices for a narration-heavy game, get an Inworld quote too.
Related Coverage
- Battlefield 6 vs Call of Duty MW4: $42 Price Gap [2026]
- NoxPlayer vs MEmu vs Genymotion: 15GB RAM Gap [2026]
- Xbox Game Pass vs PS Plus vs Switch Online: $9.66 Gap [2026]
- BlueStacks vs LDPlayer vs GameLoop: Android 14 Gap [2026]
- Epic Games Store Mobile vs App Store: 18-Pt Fee Gap [2026]
