A support system needs to decide whether a photograph shows shipping damage. A five-paragraph answer is not the interface the application needs. A short, typed judgment is useful, but the case with the worst photograph still has to go somewhere.
Google’s November 11 Gemini Decisions API release adds a dedicated model for that shape of work. It follows OpenAI’s October 22 Decisions API general availability, which expanded the Luna endpoint covered in our October assessment. Both let applications ask bounded questions against shared evidence. The engineering task is making the answer’s uncertainty usable without assuming the two services mean the same thing by confidence.
Rank, probability, and policy need separate fields
The Gemini guide describes binary questions, fixed choices and ordered rubrics. For a binary answer, a low probability of yes can accompany high confidence in no. For a choice, confidence equals the selected option’s probability; for a rubric the returned score is a probability-weighted level. A midrange rubric score could be a strong consensus on the middle or a mixture of extremes. Save the distribution when that difference matters.
OpenAI’s guide likewise exposes predicates, choices, and rubric scores. Its rubric example reports a separate confidence field alongside the level probabilities, rather than defining the Gemini formula for it. The endpoints also differ in the evidence they accept and how they return missing or refused judgments. A shared adapter should preserve provider identity, question version, raw probabilities and response status. Avoid normalizing away those differences before evaluation.
There is a useful analogy with ranking. The best label in a closed list only beat the choices you provided. It may still be unsuitable. Include explicit uncertainty and out-of-scope routes, and test missing evidence separately from borderline classifications. An obviously unrelated image should not get sent to a specialist just because there was no other plausible label.
Batch questions that see the same authorized evidence
Google documents up to 1,000 questions with up to 256K tokens of shared multimodal evidence. The launch’s reported 150–250 ms end-to-end range is for contexts up to 16K and described across one to 100 questions. Do not extrapolate that number to both maxima at once. Google’s launch rate is $0.10 per million input tokens, with bills for shared input once plus each question’s own tokens, and no output-token charge. The announcement does not establish anyone else’s latency or cost for equivalent validated decisions.
OpenAI’s GA release raises its earlier 200-question beta cap to 400 and adds prompt caching, direct JSON evidence, and images in individual instructions. Its guide supports text, images, and structured inputs, including certain message and tool histories; it does not accept audio files directly. Its separate GPT-Live integration sees a different representation. Google can accept shared audio, video, and PDF inputs directly, with their respective supported forms. Compare both on the same text/image subset first; test each additional modality separately when it matters to the workload.
Batched questions are scored against common input, not a sequential chain where one answer becomes the next question’s evidence. That is productive for independent checks. A decision requiring the output of another needs an explicit next stage or application-side composition. In addition, questions sharing the same blurry photograph can fail together. Do not multiply confidence values as though the underlying errors were independent.
Measure error against useful coverage
For a proposed pilot, use a consented, representative sample of requests labeled independently by two reviewers. Include unreadable images, conflicting text and visual evidence, unseen categories, and cases where both reviewers agree a human should decide. Separate the calibration set used to pick thresholds from the final evaluation set. Keep examples from one source event in the same split.
Compare a rule baseline, an existing generative classifier, and the typed endpoints. First use an identical input representation and question wording as far as possible. Then compare complete provider-specific pipelines, including any necessary transcription or feature extraction. For each approach, plot error among automatically resolved cases against the fraction resolved automatically. Inspect reliability by modality and rare categories. Include false-positive costs, review backlog, and actual p95 latency as concurrency and batch size vary.
The workflow can use these judgments to prioritize a queue. A decision to issue a payment, disclose data, or change state still needs a separate authority check following the execution contract. A high-confidence yes to whether an item looks damaged is evidence about the item; it is not permission to issue a refund.
The research promise is to replace some expensive, hard-to-parse prose with inspectable judgments. The operational promise is narrower and more useful: an application can automate the cases it has evidence for and preserve a clear route for the rest.