Choose the refreshed language models, understand base Sparks and measured usage, and know how retired Gemini and Claude transport selections migrate.
Where: Any supported AI model picker, the action cost quote, or Pricing → AI model rounds.
Key ideas
- A shared current catalog: The dashboard, Flowgen, Designer, pricing page, and MCP pricing catalog use the same selectable model definitions. Gemini 3.8 Flash, Qwen3.8 Flash, GPT-6 Astra, Claude Haiku 4.5, Claude Sonnet 5, and Claude Opus 5 are part of this refresh.
- Base price plus measured usage: Base rounds are 1, 1, 20, 2, 4, and 10 Sparks respectively before thinking effort and measured input/output adjustments. Sonnet and Opus at Auto use the existing high-effort price. Large Astra prompts above 272,000 tokens use the documented large-context rates.
- Compatibility without falsifying history: Saved Gemini 3.5, 3.6, and 3.7 Flash selections resolve to Gemini 3.8 Flash when run. Historical result metadata keeps its original model label. Saved Claude transport values move from vertex-anthropic to anthropic without rewriting old results.
Steps
- Open the model picker and choose from the currently selectable list; do not copy a historical model id from an old result into a new request.
- Match the model to the job: use a fast lower-cost route for routine work and a higher-reasoning route only when complexity warrants it.
- Set thinking/effort deliberately and inspect the displayed estimate. The base round is not always the final charge for long context or measured output.
- Attach supported files or URLs through the product’s attachment controls so the selected provider receives typed context.
- Run one representative round, review quality and actual usage, then scale to batches or long conversations.
- If opening an older canvas, confirm the picker displays the replacement model before a costly run. Keep old result labels as provenance.
Price reference before thinking and measured usage
- Gemini 3.8 Flash — $0.75 input / $3.75 output per million tokens; 1 base Spark.
- Qwen3.8 Flash — $0.15 / $0.47; 1 base Spark.
- GPT-6 Astra — $10 / $50; 20 base Sparks.
- Claude Haiku 4.5 — $1 / $5; 2 base Sparks.
- Claude Sonnet 5 — $2 / $10; 4 base Sparks.
- Claude Opus 5 — $5 / $25; 10 base Sparks.
Routing and caching
- Gemini 3.8 uses the Google Vertex/Firebase route, starting in global where the verified route is available.
- Astra and Claude default through OpenRouter because repeated-prefix caching was reliable there; optional Kie-first routing may be enabled server-side with fallback.
- Claude uses OpenRouter’s native Messages interface and accepts supported HTTPS image/PDF source blocks.
- Prompt-prefix cache savings are provider observations, not guaranteed discounts. Full usage/cached-token accounting survives streamed provider responses.
Known boundaries
- Qwen vision is advertised by model metadata but was not verified live during the refresh because the only upstream provider was unavailable/rate-limited.
- Browser delivery still receives accumulated final JSON through the existing bridge; provider streaming does not mean token-by-token UI rendering.
- Application output ceilings remain 64K for Claude/OpenRouter and 65,535 for Gemini.
- Gemini introductory pricing is time-bounded through 2026-12-31 and should not be read as a permanent-rate guarantee.
Tips
- Use measured usage from a representative run to budget a long batch; the base price is designed for comparison, not a maximum.
- Preserve the exact old model name when discussing an old artifact, but select the current runnable replacement for new work.
Limitations and important notes
- Provider availability, context limits, and prices can change; trust the current in-product quote at the moment of a paid run.
- A model advertising image input does not prove every upstream route is healthy.
- Optional server routing switches are deployment configuration and are not user-facing model-picker controls.
Troubleshooting
An older canvas still tries to run Gemini 3.5/3.6/3.7 Flash.
Reopen it in the updated client so the retired id resolves to gemini-3.8-flash. If an external caller sends the raw retired id, update that caller too.
The final charge is above the base Sparks label.
Check thinking effort, measured input/output, cache usage, and Astra’s large-context threshold. The base value is not a flat cap.