Simulator and What-if
Price a month of AI inference across any mix of models before you build it, or re-price your own usage on other models, hosts and volumes.
There are two versions of the same simulator. The public simulator prices any workload at list prices, with no account. What-if, in the dashboard, starts from your organization's own usage and applies your custom prices.
Lines
A workload is a list of lines, up to 25. Each line is one kind of call repeated a number of times a month: a model, what one call uses, and the calls a month. The fields follow the model's kind:
| Kind | Per call |
|---|---|
| Text | Input, cached input and output tokens |
| Embeddings | Tokens |
| Realtime voice | Text and audio tokens, in and out |
| Image | Images, plus size and quality |
| Video | Seconds, plus resolution |
| Text to speech | Characters |
| Speech to text | Minutes of audio |
| Rerank | Queries |
| Document OCR | Pages |
Batch rates apply where the provider has them. Volume growth (−20% to +60% a month) projects the total over 12 months.
What each line shows
- The cost a call and a month, and its share of the total.
- The ladder: where this call sits among every current model of the same kind, priced for exactly the same usage. "Cheaper than 72% of 1,264 text models for this call."
- The same model on other hosts: open models such as Llama, Qwen, DeepSeek and gpt-oss are matched across Together, Fireworks, Groq, DeepInfra, Bedrock and the rest, cheapest first. Switch moves the line to that host.
- Cheaper models from the same provider, with the monthly saving. A cheaper model isn't always an equivalent one; test quality first.
Copy link puts the whole workload in the URL, so a shared link opens the same scenario. Download CSV exports the lines.
What-if
The dashboard's What-if starts from the last 30 days: one line per model and host, at its average call, with monthly volume at the same rate. Each line also shows what your ledger actually recorded. Daily totals don't keep image sizes, video resolutions, or batch and priority calls, so set those on the line to match.
If your Veqen administrator has configured a model, members can also describe a workload in plain English, and a language model drafts the lines. Every price still comes from the price book. Drafting is limited per organization each hour and is only available signed in.
API and MCP
POST /v1/simulate takes the same lines and returns the same results, with no key (see the API reference). The public MCP server's simulate_workload tool does the same for AI assistants (see MCP).