RightNow AI takes an unusual angle on the model API: instead of asking an LLM to write an answer, you give it everything you have — customer text, JSON payloads, robot state, a chunk of codebase, a PDF, images, a voice clip — and it returns a decision with probabilities attached to each option. The published demo shows a support triage scenario where two text fragments and then an image shift the verdict from 'support 92%' to 'refunds 97%' as the inputs stream in.
The streaming behaviour is the technical core. Because the decision updates as each part arrives, a robotics or agentic system can act on a partial input rather than blocking until the last byte lands. A single call can also carry multiple questions, and RightNow supports a 38,000-token standard context with access to a 1M-token window for large payloads. Output is not a lone label but per-option probabilities, which makes threshold-based routing and audit trails straightforward.
RightNow is backed by Y Combinator, and the site publishes a decision benchmark that plots its index score against USD per million tasks alongside other decision models, with a clear footnote that the comparison draws on different evaluation sets. Pricing is documented in the open: $1 in free credit at signup, then metered rates for uncached and cached text, answered questions and answered calls, with an on-page estimator for modelling monthly cost.
