Choose from the expanded model roster while capability-aware controls and live Flowie prices stay in sync.
Where: Select a generation node or open the Flowgen assistant → Model
Key ideas
- Capability-driven UI: The selected model determines which reference/start/end/video inputs and duration/resolution/audio controls are valid.
- Expanded roster: The shared catalogs include the current Google, Anthropic, OpenAI/OpenRouter, image, and video options registered by the app.
- Flowie badge: The node/chat shows the price derived from the same model id and settings used at charge time.
- One row per image model family: Supported text-to-image/edit pairs appear as one model choice. At run time the image endpoint is selected from the number of connected reference images, instead of asking users to predict an internal Edit row.
- Model-owned video parameters: Every video picker/card reads duration, aspect, and resolution from the selected endpoint. Unsupported controls hide, one-choice controls become static, and stale values are coerced before request and price calculation.
- August model additions: The current shared catalog includes Gemini 3.8 Flash and Grok 4.6 for supported reasoning/orchestration roles; LTX 2.5 Pro/Fast and new Grok video/image routes; Grok Imagine 2/Quality; MiniMax Music 3; Seed Audio 1.0; and Hi3D/Hi3D Multi-View.
- Reference limits and real price: The selected model defines accepted reference-image count, input roles, duration/resolution/audio controls, and price. Extra references are reported instead of silently sent, and model rows quote the model round actually charged.
- Current text and video routes: GLM 5.3 (2 Sparks), DeepSeek V4 Flash (1), and DeepSeek V4 Pro (2) join text/orchestrator selection. Seedance 2.5 and PixVerse V6 add sound-capable video routes with exact per-second resolution pricing.
Steps
- Select the node or chat.
- Open Model and compare provider/cost/capabilities.
- Choose the model before wiring model-specific inputs.
- Review the Flowie estimate and run.
- After choosing a model, re-check the visible input handles, reference count, output controls, and quoted Sparks before running.
Why settings change by model
The shared catalog carries capability metadata for input handles, duration, resolution, aspect ratio, audio, and price. A model choice therefore changes the valid node contract as well as the output engine.
Selection checklist
- Required inputs/reference roles.
- Output duration/resolution/aspect.
- Audio support.
- Latency/quality trade-off.
- Estimated Flowies and batch size.
Video coercion
- Duration snaps to the nearest supported length.
- Resolution snaps down rather than silently choosing a more expensive tier.
- Aspect ratios match equivalent values, such as 16:9 and 1280:720.
- Model family order is client-owned and consistent across card, sidebar, edit/extend/avatar, and agent planning pickers.
Image resolution note
Nano Banana 2 exposes 512, 1K, 2K, and 4K in the current UI/pricing ladder; the Lite variant exposes 512 and 1K. The external Vertex adapter must preserve imageSize for those selections to reach the provider.
Examples of changed contracts
- LTX 2.5 Pro: 720p/1080p, native audio, 6-10 seconds.
- LTX 2.5 Fast: up to 4K/20 seconds with native audio where offered.
- Grok Imagine 2 supports edit/reference-image pricing and a Quality tier.
- MiniMax Music 3 supports full songs with lyrics up to the live limit.
- Seed Audio supports short prompt/audio-reference voice and sound workflows.
- Hi3D offers single-view and named multi-view 3D reconstruction.
Model-specific facts
- GLM 5.3 is text-only and selectable as an orchestrator; one provider route does not support structured output.
- DeepSeek V4 Flash is text-only; its floating alias can move as the provider updates it.
- DeepSeek V4 Pro is text-only and suited to planner/orchestrator work.
- Seedance 2.5: 480p 14, 720p 30, 1080p 72 Sparks/second.
- PixVerse V6: 360p 3, 540p 4, 720p 5, 1080p 9 Sparks/second; 1-15 seconds, sound, five shapes.
August 31 model additions and compatibility
GLM 5.3 Flash is a distinct multimodal language-model route with native image/video input and mandatory reasoning. Gemini 3.5/3.6/3.7 Flash are retired from pickers; updated clients resolve saved selections to Gemini 3.8 Flash. Historical catalog labels remain available for old runs, and the base flash round tier is unchanged.
- Wan 3.0 and Prime support frame, reference, and extend workflows with source-video billing and explicit duration limits.
- Omni 1.1 Flash adds 360p/720p/1080p/4K, up to ten images and three videos of at most ten seconds each, last-frame support, and continuation.
- Grok Imagine 2 Standard costs 3/5 Sparks at 1K/2K; Quality costs 5/6.
- GPT Image 2 supported output sizes are flat-priced: generation 20 Sparks, edit route 22.
- OpenRouter attachments follow catalog capabilities. PDFs/Markdown are prepared as document content instead of being reduced to unreadable URLs.
- OpenRouter agent streaming and increased output ceilings support longer answers; output remains bounded and is included in usage-based billing.
Current language models and measured pricing
The shared catalog adds Gemini 3.8 Flash, Qwen3.8 Flash, GPT-6 Astra, Claude Haiku 4.5, Claude Sonnet 5, and Claude Opus 5. Saved Gemini 3.5/3.6/3.7 Flash choices now resolve to 3.8. Historical labels remain unchanged, and the visible base round may increase with effort, measured usage, or Astra’s large-context tier.
- Qwen image/video capability is catalogued, but live vision was not verified during the refresh because the upstream route was unavailable.
- Astra/Claude default to OpenRouter for reliable prefix caching; optional Kie-first routing is a server setting.
- The current action quote is authoritative for a paid run.
GPT Image 2.5
- Flare is the fast tier and Sunburst is the precise tier.
- Connecting references automatically selects the matching edit twin.
- Both support 1K, 2K, and 4K plus up to sixteen references across the supported aspect-ratio list.
Tips
- Treat the live picker as the authoritative roster; model availability and pricing can change.
- Run a representative low-cost test before a large loop.
Limitations and important notes
- The ignored/vendor-owned Vertex sanitizer is outside this repository’s tracked source; verify Nano Banana 2 resolution end-to-end in the deployed backend before relying on a 2K/4K delivery.
- Some image dispatch/pricing paths were committed separately from the unified picker and should be verified together before release.
Related guides
- Model pricing in Flowies
- Choose the language model in any chat
- Generation Providers & Models
- GLM 5.3 Flash: multimodal chat, reasoning, and pricing
- Wan 3.0 and Prime: video modes, references, and controls
- Gemini Omni 1.1 Flash: generate, reference, edit, and extend video
- Grok Imagine 2: Standard, Quality, resolution, and image edits
- September 2026 language-model refresh: Gemini 3.8, Qwen 3.8, GPT-6 Astra, and Claude 5
- Generate and edit images with GPT Image 2.5 Flare and Sunburst