The report
Google released Gemini Decision 1 through a dedicated Decisions API on November 11. Instead of producing open-ended text, the model returns scores for bounded yes/no, choice and rubric questions, including batches sharing a multimodal context. This is a useful component for routing and evaluation, but a high probability for an option is still a model judgment, not an independently enforced permission to execute it.
What is included in the claim of speed
The first-party product release advertises up to 1,000 questions sharing context of as many as 256,000 tokens, with candidate choices ranging from two to 255. Google quotes a price of $0.10 per million input tokens, with no billable output tokens. Its reported 50 to 100 milliseconds of model compute and 150 to 250 milliseconds end to end were measured on smaller inputs, up to 16,000 context tokens and one to 100 questions. Those latencies should not be projected onto the advertised maximum batch or full context window.
A typed answer has several operational advantages over parsing a free-form response: it has an enumerated option set, predictable output and a score that can be calibrated against examples. The same structure creates a new failure mode if the caller's labels are wrong, policy information is incomplete or the model is allowed to choose a label meaning approval without a separate authorization check.
What should be allowed to follow a model decision?
For a low-stakes routing decision, an erroneous choice may mean extra compute or a poor search result. For a security decision, the identical response format could govern whether an agent reads a customer file or transfers funds. These applications need different acceptance criteria. The structured-decision boundary analysis discusses how label and rubric semantics can produce a correct-looking typed response with the wrong operational consequence. The Auth0 retrieval case shows a separate failure in an enforcement adapter, even when the model itself is not issuing a malicious instruction.
A meaningful evaluation should pair known authorized and denied actions with misleading choice descriptions, missing policy facts and contradictory contextual evidence. Report both false approvals and false refusals, then require a deterministic executor to check the effective principal, target and effect. The product launch does not document a demonstrated authorization bypass in Gemini Decision 1. It establishes a useful, testable decision interface whose safety depends on what downstream code trusts it to decide.