Choose a model for your workflow
Check capabilities, cost and data handling before connecting. You choose the model used for each request.
Start with the task
| Your workflow | What to check |
|---|---|
| Review source files | What to checkEnough input space for your code, instructions and conversation, with room left for the answer. |
| Check a draft finding | What to checkInclude the relevant code, evidence and scope. Compare whether the answer supports its claims, not just whether it agrees. |
| Run an agent with tools | What to checkChoose the tools capability and a client that supports OpenAI-compatible Chat Completions. Your client executes tools; the Model API supplies text and tool calls. |
Check streaming and structured-output support separately. A model’s name or “low refusal” label does not establish tool compatibility or finding quality. Use the exact model ID from the catalog.
Compare cost and readiness
Check the model’s current availability before starting. In the catalog, switch workload view to compare estimated token costs, including the minimum request charge. An agent can make several requests for one task, so set an output limit and a spending cap on its API key.
Estimates are not quotes for a completed investigation. Credit is reserved before a request, and the unused portion is released after settlement. A self-hosted model may take several minutes to start after idle. Allow for that startup time when choosing a model for interactive work. See charging and timeout details.
Choose where your content goes
Read the hosting provider and source details on the model card. Our self-hosted RunPod deployments disable prompt and response logging and cross-request prefix caching. Managed models use Infron or OrcaRouter, including any downstream providers, under their own terms and retention policies. They do not come with a zero-retention guarantee.
Send only material you are authorized to share with the selected providers. Review the data handling details before connecting confidential work.
Read the evidence in context
Our Paydirt benchmark, published , measures small-file Solidity and TypeScript review and checking draft reports. Compare its errors, costs and evaluation method alongside its scores. It does not measure whole-repository audits or live web and cloud assessments.
Wiz Cyber Model Arena is an external reference for broader tasks. Its results evaluate a model together with its agent harness. They do not score our hosted variants: a base model’s result does not transfer to a fine-tune, quantized version or different harness. If the exact setup has no matching evaluation, its performance is unmeasured here.
Start with a small authorized task whose outcome you can check. Keep the model, harness and settings recorded so comparisons mean something.