Google 不是第一家。OpenAI 自己的 Decisions API 在 10 月 22 日正式开放,跑在 GPT-6 Luna 上;TypeSafe 的 Jev 早就在卖同样三种题型。Simon Willison 还指出,Google 连是/否题的叫法「noul」都沿用了 Jev 的。所以有意思的决定不在模型,在价目表。
Google 定价每百万输入 token 0.10 美元,输出不收钱。更关键的是它选了计费单位:共享输入每个请求只算一次,不管上面挂多少个问题;每个问题只加它自己的说明和选项描述。同一张价目表也写明了缺什么:两档都没有上下文缓存,也不支持 Batch API。Decisions API 文档说得更直白:跨请求重复发送的输入,全额计费。
问题彼此独立。 只有 OpenAI 的文档要求:依赖前一个答案的判断,要另发一次请求。Google 的文档给了两条路,原文是:「Either send a second request after you read the first answer, or ask both questions in one request and ignore the answer you don’t need.」也就是读完第一个答案后再发一次请求,或者两个问题放在同一次请求里、把不需要的答案丢掉。在 Gemini 上,只有把追问拆成第二个请求时,同一份证据才会再付一次。
Gemini Decisions API: Fast, calibrated multimodal decisions (Google blog) — 11 Nov 2026, Min Choi and Felix Yu (Google DeepMind). gemini-decision-1; up to 1,000 yes/no, choice and rubric questions over one multimodal input up to 256K tokens; scores options in a single forward pass; 150-250 ms end-to-end, 50-100 ms median model compute for inputs up to 16K tokens, nearly flat from 1 to 100 questions; $0.10 per 1M input tokens, $0 output; shared input billed once per request, each question adds only its own instruction and criteria tokens; no thinking, transcription or OCR charges; nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API on Google's shared benchmark suite; Gemini Notebook routing with up to 40% lower latency; Google Cloud support coming soon.
Gemini Developer API pricing: Gemini Decision 1 (Google AI Studio docs) — Checked 12 Nov 2026. Each request bills shared input tokens once regardless of question count; each question adds tokens for its own instructions, criteria and formatting; zero output tokens; Batch API not supported; paid input $0.10 per 1M tokens for text, image, video, audio and PDF alike; context caching price 'Not available' on both tiers; free tier used to improve Google's products, paid tier not.
Decisions API (Google AI Studio docs) — Checked 12 Nov 2026. 1 to 1,000 questions per request, bounded by the 262,144-token context window; 2 to 255 options per choice question; put independent questions in one request; doesn't support context caching, Batch API, or Flex or Priority inference, and repeated input across requests counts in full; yes/no probabilities always fall between 0.01 and 0.99, so a threshold above 0.99 never matches. Dependent decisions: 'Either send a second request after you read the first answer, or ask both questions in one request and ignore the answer you don't need.'
Decisions (OpenAI API docs) — Checked 12 Nov 2026. Generally available; gpt-6-luna is the only supported model; 1 to 400 questions per request; text, images and JSON input; decisions that depend on an earlier answer go in separate requests; prompt caching added since beta, reusing a matching prefix of shared input with discounted cache reads, cache writes charged; the reusable prefix must contain at least 1,024 tokens, and question content does not count toward that minimum; uncached input $0.10 per 1M tokens, output not charged.
API pricing (OpenAI API docs) — Checked 12 Nov 2026. gpt-6-luna standard short context: input $0.10, cached input $0.01, 30-minute cache write $0.125 per 1M tokens. The 30m write column is the rate used for the default 30-minute cache lifetime.
Google's Gemini Decision 1 model scores up to 1,000 questions in one API call (VentureBeat) — 11 Nov 2026, Carl Franzen. Worked example: 4,000-token record with 20 checks of 50 tokens costs about $0.50 per 1,000 records in one request vs $8.10 as 20 separate calls; Google's shared tests cost $0.88 with Decision 1, $0.68 with Jev 1.14, $4.14 with Gemini 3.5 Flash-Lite; Decision 1 calibration error 0.0250 vs Jev 0.0494; Artificial Analysis Decision Index v2.1 puts Gemini ninth of 18 at launch (65.98, $345 per million tasks; Jev 1.14 66.23, $156); OpenAI's Decisions API GA on Oct 22. Cloudflare Clef-Omni takes audio and video, with a 64,000-token window on Workers AI at $0.15 per 1M input tokens; for recordings and video, compare Gemini and Clef-Omni on the same clips. Nace Drex Max v1.1 at $0.03 per 1M input tokens.
Gemini Decisions API (Simon Willison) — 11 Nov 2026. Same three question types as Jev, reuses the 'Noul' term; priced the same as GPT-6 Luna decisions; 'Jev remains the cheapest of the three, at 4.2 cents per million.' General line, not a video result: 'Jev and OpenAI both return 1.0 or 0.0 for yes/no questions.' He also writes that Jev is text and JSON only. The video test on the page is Gemini only, on his Yonkers stomping-robot clip: robot probability 0.9234055387792485, dragon 0.04976021841520312.
Prompt caching (OpenAI API docs) — Checked 12 Nov 2026. For GPT-5.6 and later, the minimum cacheable prefix is 1,024 tokens. The default lifetime is 30 minutes, and the write rate for that lifetime is 1.25× uncached input, a total rate rather than a surcharge on top of ordinary input. A 30-minute read refreshes coverage to at least 30 minutes after that read.
Models (TypeSafe docs) — Checked 12 Nov 2026. TypeSafe lists Jev 1.14 at $0.042 per million input tokens ($42 per billion); output is free. Jev reads text as a string, a JSON object, or an array of text values, and PNG or JPEG images anywhere in the request. It does not natively accept audio or video.
Artificial Analysis Decision Index v2.1 — Checked 12 Nov 2026. Charts are labeled '18 of 19 models.' The page's decision-model list has 19 names in score order; Gemini Decision 1 at 65.98 is tenth, behind Jev 1.14 at 66.23. Launch ranking cited from VentureBeat stays ninth of 18.
Gemini Decision 1 (Artificial Analysis) — Checked 12 Nov 2026. The model-page summary still reads #9 / 18, with the Decision Index score shown as 66. Charts on this page say '18 of 18 models.'
Drex Max v1.1 (Artificial Analysis) — Checked 12 Nov 2026. Input price $0.03 per 1M tokens. The page says this cost reflects a temporary launch price, and that at the standard price the cost would be double.