What would you like to try?
Start without creating an account. Choose a free example below.
Estimate what a prompt could cost
Open Prompt Pricer and try an example, then adjust the text or number of calls. You will see estimated prices across models; your prompt is not run.
See how models performed
Choose a task in the public benchmark. Compare answer quality, time and cost, then open an answer to decide whether it fits your needs.
See why an agent should stop
Run the Loop-Guard demo. Watch repeated actions or reported spending trigger a stop. The demo stays in your browser and sends no alerts.
How to interpret results and estimates
Five repetitions of the same prepared task show repeatability for that fixture. They do not test the full range of real workloads. Review sample counts, failed attempts and evaluator comments before treating a score as a recommendation.
Pricer computes estimates in your browser. Live model prices may come from OpenRouter; refreshing prices does not execute your prompt. Token counts, media conversions, reasoning and caching differ by model. Estimates are not invoices. Shared settings exclude raw text.
Add checks to your own project
Loop-Guard detects repeated actions and observed usage thresholds locally. A call already in progress can exceed a threshold.
Ship-Check checks recorded tool calls against your rules before a GitHub merge. It verifies signed trace-policy evidence and commit bindings. Supply evidence from your own harness and trusted CI; a benchmark report alone is not release evidence.
Installation and working examples →Existing-account information
Private on-demand comparisons, saved reports, Workspace and Teams remain legacy previews. Subscriptions and top-ups are unavailable. Existing-user recovery and access routes are preserved under Legacy access.
The legacy comparison gateway supports 1–3 models, 4,000 combined prompt characters, up to 10 checks and a 512-token completion cap per model. Those short private runs are not directly equivalent to the complete-answer protocol used for public benchmarks. Execution requires a confirmed account; no API key is needed in Studio. Prompts go to external model providers; account privacy does not imply local-only processing.