A video company, selling judgment
The odd part is who is selling this. HeyGen's other three model families generate video, render avatars and clone voices. Decision 1 is the fourth family, and it produces nothing a viewer would ever see. Its natural use is backstage: routing a support request, ranking a result, deciding whether a person should review something. HeyGen's own example classifies a support ticket, checks whether work is blocked and rates its impact in a single call.
That it accepts video matters for HeyGen's existing customers. A base-model call can include one video alongside the questions, so developers can submit video inputs to the decision endpoint. The Turbo preview raises that to 10 videos and 100 images per request.
The documentation argues against its own numbers
This one spends a section explaining why its scores should not be trusted at face value. Confidence, HeyGen writes, is derived from the returned distribution and is "not a calibrated probability that the answer is correct." A yes/no answer near 0.5 is uncertainty, and 0.5 itself "is not a guarantee that an answer is reliable." Developers are told to test thresholds on labeled examples from their own use case.
Two mechanical details deserve the same attention. A choice question with one option, or a score question with one level, returns a fixed answer with probability and confidence of 1 without running the model at all. That is disclosed, but a pipeline that logs confidence will record a confidence of 1 for a question the model never read — a fixed answer, not a prediction. The context limit applies per model read; reported input usage totals tokens across reads, so it can exceed that limit. The bill is per input token, so the number to watch is the returned usage count, not the size of the material you sent.
What the documentation does not contain is any measure of how often the model is right. No benchmark, no eval set, no error rate. OpenRouter's displayed counters when checked on November 14 showed 746 million tokens processed for Decision 1 and 1.24 billion for the Turbo preview, which shows usage, not correctness. A decision model with no published accuracy results is a tool each buyer has to validate on their own workload, which is exactly what HeyGen's threshold advice quietly assumes.
The practical read
At OpenRouter's listed $0.14 per million input tokens, a token bill for an evaluation is small; estimate the real cost from actual usage, including repeated reads, plus the labeling work. Run that evaluation before the routing does. The teams this suits are the ones already holding labeled examples, a support queue, a moderation backlog, a ranking task, and no patience for parsing free-text answers out of a chat model. The teams it does not suit are the ones who need the probability to mean something on day one. On HeyGen's own wording, it does not yet.
Sources
- [1] HeyGen developer documentation, "HeyGen Decision" (updated November 13, 2026)Read source
- [2] HeyGen developer documentation, "Models" (four v3 model families)Read source
- [3] OpenRouter, HeyGen provider page (pricing and tokens processed, accessed November 14, 2026)Read source
- [4] Inference360, "Changes" (Decision 1 and Decision 1 Turbo listed as new on November 14, 2026)Read source