On this page
Worldstream models and pricing
Pick a model for the experience you want to build. Native detail, scene continuity, response speed and billing rules all affect that choice.
Worldstream model comparison
Scroll the table horizontally to compare all model details.
| World Model | Sequencer price | Provider cost | Video and audio | Response timing | Quality / 5 |
|---|---|---|---|---|---|
| Sequencer VivaAvailable through Sequencer | $40 / active stream hourAbout $0.67 / minute. Viva LLM direction included. | Sequencer managed modelExisting retail rate | 768p, 24 fpsGenerated video and audio | Continuous streamMeasured startup latency not published | Unrated |
| H3 Max Directorfal model. Sequencer route planned. | $331.20 / generated hour$0.092 / second. Estimate. | $288.00 / generated hour$0.08 / second | 480p or 768p, 24 fpsStereo audio | 5–15s segmentsFirst-frame latency not specified | Unrated |
| MiniMax FastH3Reactor model. Sequencer route planned. | $144.90 / session hour$0.04025 / second. Estimate. | $126.00 / session hour$0.035 / second | 1344 × 768, 24 fpsGenerated audio | About 1× real-time generationClips queued before playback | Unrated |
| Visko Orbis StableReactor model. Sequencer route planned. | $40.16 / session hour$0.011155 / second. Estimate. | $34.92 / session hour$0.0097 / second | 832 × 480, 18 fpsUpscaled delivery up to 4K; audio | About 1.8s chunksPrompt changes at chunk boundaries | Unrated |
| Visko Orbis DynamicReactor model. Sequencer route planned. | $86.53 / session hour$0.024035 / second. Estimate. | $75.24 / session hour$0.0209 / second | 640 × 368, 18 fpsUpscaled delivery up to 4K; audio | About 1.8s chunksLive delivery resolution changes | Unrated |
Model setup links lead to Sequencer. Provider specifications and prices were checked October 13, 2026. The provider column shows published costs. The Sequencer column applies 15% markup, calculated as provider cost × 1.15, for planned third-party routes. Viva keeps its existing $40/hour retail rate.
Third-party routes are planned and their prices are estimates. Live quotes determine the actual rate when a connector becomes available. Hourly figures use 3,600 billable seconds and exclude app infrastructure, taxes and separately priced services. Resolution and frame rate describe output format; they do not establish visual quality.
Read the prices correctly
- Sequencer: Advertises $40 per active stream hour. The session quote is authoritative. Viva includes built-in LLM world direction. Custom Genesis LLM requests and extra image/audio generation are separately metered. Worldstreams API, Genesis integration.
- fal H3 Max Director: Bills generated video at $0.08 per second, with a 60-second minimum ($4.80). Public sessions currently have a 15-minute ceiling. Director pricing and session limits.
- Reactor: Bills wall-clock time while a session holds a GPU, including ready-but-idle time and recoverable disconnects. The three Reactor rates in the table also match its public pricing API at the check date. Billing rules, current pricing API.
At the proposed 15% markup, Orbis Stable is about $40.16 per session hour. Viva is $40 per active stream hour, with built-in LLM direction. The billing clocks differ, so compare the session quote and your usage pattern. A complete experience also needs state storage, user access, LLM logic and audience billing. Compare those responsibilities in the setup options.
Read the timing correctly
Startup latency is the time from requesting a session to seeing its first usable frame. Response latency is the time from sending an action to seeing its effect. Generation speed describes how quickly video is produced. Chunk duration describes when a streaming model can accept a new direction.
The timing column reports the measure each provider publishes. It is not a common end-to-end latency benchmark. Queuing, network delivery, buffering and your controller can change the wait a player experiences.
Quality ratings
GENESIS uses a 1–5 rating for each of five criteria. The overall quality score is their mean, shown to one decimal place. Speed and price are reported separately.
| Criterion | What to check |
|---|---|
| Visual detail | Clear subjects, readable environments and visible artifacts. |
| Prompt adherence | Whether the requested action, setting and constraints appear. |
| Character consistency | Whether identities, clothing and objects remain recognizable. |
| Scene continuity | Whether locations, spatial relationships and previous events carry forward. |
| Motion and stability | Natural movement, temporal consistency and recovery after a direction change. |
Use 1 for major failures, 2 for frequent failures, 3 for a usable result, 4 for a consistently strong result and 5 for a reliably excellent result across the test set.
Score at least ten shared scenarios, run each three times for two minutes, and publish the prompts, model/version, resolution, settings, date and recordings. Include the same character, inventory and scene-transition challenges for every model. Have at least two reviewers score recordings without provider labels.
No matched GENESIS benchmark has been completed. Every model is currently unrated, including Sequencer. A provider's demo or a clip-generation leaderboard does not establish its score for a persistent interactive world.
Low-latency LLMs
Viva LLM means the built-in direction service in a Viva Worldstream. It turns accepted world actions into video prompts and is included with the stream. Arbitrary structured calls through tool.llm.generate are a separate service, currently $0.01/request by default.
Scroll the table horizontally to compare all LLM details.
| LLM | Sequencer price | Provider price | Latency / generation speed | Moderated / unmoderated |
|---|---|---|---|---|
| Viva LLMBuilt-in world direction | Included with VivaViva stream: $40 / active hour | Bundled world direction | Runs in the stream loopMeasured response latency not published | Provider safety protectionsHosted world rules also apply |
| Sequencer structured LLMCustom moves, choices and JSON | $0.01 / requestCurrent default. Separate from built-in direction. | Managed request price | Measured response latency not publishedDefault backend: Gemini 3.5 Flash-Lite | Provider safety protectionsValidated structured responses |
| Gemini 3.5 Flash-LiteGoogle | $0.345 / $2.875Planned token route. Includes 15% markup. | $0.30 / $2.50Published token rate | Designed for low latencyNumeric first-token latency not published | Adjustable filtersCore protections remain enabled |
| GPT OSS 20BGroq | $0.08625 / $0.345Planned token route. Includes 15% markup. | $0.075 / $0.30Published token rate | About 1,000 tokens / secondThroughput, not first-token latency | App moderation availableProvider policies and model refusals apply |
| GPT OSS 120BCerebras | $0.4025 / $0.8625Planned token route. Includes 15% markup. | $0.35 / $0.75Published token rate | About 3,000 tokens / secondThroughput, not first-token latency | Provider policies applyHost filtering details not published |
| Your self-hosted LLMChoose your model and hardware | Your infrastructure costNo universal token rate | Your infrastructure cost | Depends on model and hardwareMeasure your own deployment | Moderated or unmoderated by your appThe selected model may still refuse requests |
Token rates are USD per one million input / output tokens for standard text requests. Planned token-priced Sequencer routes use the same 15% markup. The available structured LLM endpoint retains its current per-request rate. Caching, reasoning tokens, tools and provider minimums can affect the final quote.
Latency and moderation
First-token latency measures the wait before the first token arrives. Generation speed measures tokens produced per second after generation begins. Structured JSON usually needs to finish and pass validation before GENESIS can apply it. None of the listed services has a matched GENESIS end-to-end latency benchmark.
The moderation column describes the controls we can verify. Google exposes adjustable filters alongside persistent core protections. Groq documents a separate moderation workflow. Open-weight models can have learned refusals, and hosted providers still apply their policies. A self-hosted app can choose whether to add a moderation layer; that does not remove refusals built into its chosen model.
- Google Gemini pricing and Flash-Lite model details support the token rates and low-latency design. Safety settings describe adjustable filters and core protections.
- Groq model catalog lists GPT OSS 20B at $0.075 input / $0.30 output and about 1,000 tokens/second. Content moderation describes an app moderation workflow.
- Cerebras pricing lists GPT OSS 120B at $0.35 input / $0.75 output. Its model catalog lists about 3,000 tokens/second. Model and capacity availability can change.
Provider speeds are published throughput claims, not measured GENESIS latency guarantees. Connect through Sequencer or bring your own adapter.
Sources and model details
- Sequencer Viva tutorial and Worldstreams API. Viva specifications follow Sequencer's published product material; available deployments and stream quotes can change.
- fal H3 Max Director. This is the streaming endpoint. The H3 Max image-to-video endpoint returns clips and has its own pricing.
- Reactor MiniMax FastH3 specifications. Its published 1× real-time generation rate describes throughput, not time to the first frame.
- Reactor Orbis Stable specifications. It generates 832 × 480 frames and upscales the delivered stream.
- Reactor Orbis Dynamic specifications. Its current model card lists 640 × 368 native frames. The shared API overview lists a different raster, so confirm the selected deployment's output before choosing it for resolution-sensitive work.
Sequencer routes need tested connectors, quotes and lifecycle metering before they become available. The standalone engine also supports your own adapters. Sequencer model routing, connect your own providers.