Prompt Workbench
Write the messages, fill the variables and inspect the export. Save useful versions in this browser. No model call happens here.
Model data. Verified-price snapshots from . Details
35 of 39 tracked models have verified token rates in the saved catalogue. Rates come from the OpenRouter public API. Models missing from that catalogue have unknown pricing and are excluded from cost estimates and chat selection.
Rates are USD per million text tokens. Caching, reasoning tokens, images, tool charges, provider routing and long-context tiers can change the bill. Check the provider before budgeting.
Context limits have their own dated review. The Context Budget Planner shows the catalogue and top-provider quotes separately, using the smaller known limit. Output caps are a separate constraint. Weight links identify repository revisions whose file lists were checked. Licence labels come from the model cards and are not an open-source certification. Check the actual terms and access conditions. Published weights do not make a hosted API or self-hosting free.
01 / WRITE THE MESSAGES
Use {{variable}} to add a reusable input.
Use {{variable}} to add a reusable input.
Assistant prefix, for models that support it
This becomes a final assistant message in the export. Prefilling depends on the model and provider. Chat Playground does not import a prefix. Earlier saved prefixes are preserved here.
02 / INSPECT AND TAKE IT WITH YOU
See exactly what you are copying.
About 0 input tokens · approximately $0 at the saved input rate.
This is a rough text estimate. It excludes message overhead, output, reasoning and caching. Tokenisation varies by model. Request exports set a 2,048-token output limit, which you can edit.
The playground receives the filled system prompt, user draft and model choice in this tab. Nothing is sent until you choose Send.
Each message and filled result is limited to 200,000 characters. Libraries hold up to 500 versions within 5 MB of JSON. Prompts and exports stay in your browser. Check model output with the JSON validator or compare prompt versions in Prompt Diff.