Find the September language, image, voice, and music additions, understand transparent-background and typed-input options, and review the live quote instead of relying on stale price lists.
Where: Use the model picker in Flowgen, Designer, or another supported AI surface; see the current quote before each paid run.
Before you start
- Use the model picker for the operation you are performing: language-model choices, image generation/editing, voice, and music use different catalogs and inputs.
Key ideas
- New language models: The September 22 refresh added Grok 4.7, GPT-6 Sol, GPT-6 Luna, and Claude Opus 5.5 to the selectable language-model catalog. DeepSeek V4.1 Flash’s automatic effort behavior was also tuned to avoid unnecessary default reasoning latency.
- MiMo V2.6 Pro and Flash: The selectable language-model catalog now includes Xiaomi MiMo V2.6 Pro and MiMo V2.6 Flash through OpenRouter. They support text, image, video, and audio inputs where the active surface accepts those attachments. Pro is suited to stronger orchestration/planning; Flash is the lower-cost, faster route. Use the live Flowgen assistant/model picker and quote to confirm the available role, thinking level, and current price.
- Qwen Image 2.1 is an image route: Qwen Image 2.1 supports image generation and reference-based editing through separate model IDs. The current catalog exposes 1K/2K output choices, multiple reference images, and transparent backgrounds where compatible.
- GPT Image 2.5 can return real transparency: The Flare/Sunburst generation and edit routes expose a Transparent background control. It maps to the provider’s published background enum; it is real alpha output, not a request to draw a checkerboard. Check output format compatibility and the selected route before relying on transparency.
- Gemini 3.8 TTS adds speech styles: Gemini 3.8 Flash TTS and Flash Lite TTS provide speech generation with structured style and multi-speaker support, broad language coverage, and WAV output. They use the Gemini Developer API route in this implementation; Vertex TTS is not the active route for these IDs.
- Lyria 3.5 is a music model: Lyria 3.5 is a music-generation choice for longer song output with vocals and timed lyrics; it is distinct from language models and text-to-speech voices.
- Picker and quote are authoritative: A model’s base price, thinking multiplier, accepted input types, available controls, and provider availability can differ. The live picker and operation quote reflect current project configuration better than a saved screenshot or old price table.
Steps
- Identify the job first: use a language model for assistant reasoning, GPT Image 2.5 or Qwen Image 2.1 for still-image generation/editing, Gemini 3.8 TTS for spoken audio, or Lyria 3.5 for music.
- Open the matching model picker in the intended product surface. Search by the visible current label rather than copying an old internal model ID from an earlier result.
- For GPT-6 Sol/Luna, Grok 4.7, Claude Opus 5.5, or MiMo V2.6 Pro/Flash, review the selected thinking/effort level and the task’s supported attachments before sending a long or paid request.
- For GPT Image 2.5 or Qwen Image 2.1, choose generation versus editing, connect the intended reference images through typed inputs, and select the supported size/background settings shown for that model.
- For Gemini TTS, enter/select the spoken text and voice/style controls supported by the voice surface. Use Flash for the richer acting/multi-speaker path or Lite for a lower-latency, lower-cost choice when the picker offers both.
- For Lyria 3.5, describe the musical intent and inspect the duration/lyrics controls shown by the current node before generating.
- Review the live Spark quote and any usage/effort estimate before a paid run. Start with one representative result, review quality, then scale to more variations or a longer batch.
- When reopening an older canvas, preserve historical result labels but confirm the current picker resolves the saved selection to an available route before spending.
Model families in this refresh
- Language and assistant models: MiMo V2.6 Pro/Flash, Grok 4.7, GPT-6 Sol, GPT-6 Luna, Claude Opus 5.5, and DeepSeek V4.1 Flash.
- Image generation/editing: GPT Image 2.5 Flare/Sunburst and their edit routes; Qwen Image 2.1 and Qwen Image 2.1 Edit. The edit routes consume typed reference images.
- Transparent background: GPT Image 2.5 and Qwen Image 2.1 routes expose the provider-supported alpha/background choice; verify the output format can carry alpha.
- Voice: Gemini 3.8 Flash TTS and Gemini 3.8 Flash Lite TTS. Structured speech style belongs to the model’s voice controls, not to words prepended to the spoken transcript.
- Music: Lyria 3.5 for song generation with vocals/timed lyrics.
- DeepSeek V4.1 Flash Auto maps to Off by default to avoid unnecessary reasoning latency/token use.
Quote and compatibility checks
- Language-model base Sparks are not a maximum for every prompt: thinking effort, context size, measured usage, and cache behavior can affect the final amount.
- Image, voice, and music generation use their own operation pricing; check the quote in the node or media picker.
- Image-model controls and reference ceilings are model-specific. A transparent output option may be unavailable with incompatible formats.
- Provider routes and introductory rates can change. Use the current quote rather than copying a dated rate from this article.
Tips
- Use Qwen Image 2.1 Edit only when you intend the connected references to guide an image edit; use the generation route for a new scene.
- For TTS, review the generated audio rather than judging from the text preview alone.
- Do not interpret “model added” as proof that every interface exposes every modality; choose the operation-specific picker.
Limitations and important notes
- Models, routes, reference limits, context windows, voice catalogs, and prices may change.
- Vertex does not serve the Gemini 3.8 TTS IDs in the verified implementation; the working route is Gemini Developer API.
- A capability listed in model metadata does not guarantee every provider path is currently healthy.
Troubleshooting
I cannot find a model in this picker.
Check that you are using the operation’s correct catalog: language models are not image, voice, or music models. Some models are available only on their compatible operation.
The final Spark charge differs from the base label.
Review the quote, thinking/effort, measured usage, and model-specific size or duration. The base Spark value is not always a flat cap.
An older saved model does not run.
Reopen the project in the updated client and inspect the selected current model. If an external MCP caller sends a retired ID directly, update that caller to a current picker/catalog ID.