HeyGen Decision 1 powers multimodal identity checks
HeyGen Decision 1 evaluates video, image, and text inputs to return calibrated probabilities for identity matching and liveness verification. The multimodal classification model replaces multi-step LLM completion chains with deterministic scores, choices, and binary decisions in a single API call.
HeyGen exposes the internal classification engine behind its avatar safety pipeline, proving specialized decision models beat prompt-engineered LLMs on speed, cost, and reliability.
- –**Multimodal identity verification**: Handles side-by-side comparisons like selfie video against photo ID, outputting separate, calibrated probabilities for identity matching and liveness checks in one pass.
- –**Structured-first primitives**: Skips conversational text generation in favor of typed `noul` (yes/no), `choice` (categorical), and `score` (ordinal rubric) responses ready for programmatic logic.
- –**Radical cost efficiency**: Priced at just $0.07 per million input tokens with zero output token fees on OpenRouter, making it vastly cheaper than general-purpose vision models like GPT-4o or Gemini 1.5 Pro.
- –**End-to-end pipeline replacement**: Consolidates complex agentic triage, moderation, and visual QA workflows that previously required chaining prompt extraction, vision parsing, and validation scripts.
DISCOVERED
21m ago
2026-11-20
PUBLISHED
1h ago
2026-11-20
RELEVANCE
AUTHOR
OpenRouter