Google Launches Gemini Decisions API With Up to 1,000 Questions per Request at $0.10 per Million Input Tokens; Its Benchmark Table Is Self-Reported

Overview

Google announced the Gemini Decisions API in a post on its blog dated November 11, 2026, describing “a new endpoint for fast, closed-set decisions” powered by a model called Gemini Decision 1 (gemini-decision-1). Rather than generate text, the model scores options the developer supplies and returns probabilities and confidence values. Google says a request can carry up to 1,000 independent questions over text, images, audio, video and PDFs, and that input costs $0.10 per 1M tokens with zero output tokens billed.

The launch was reported by The Decoder, MarkTechPost and InfoWorld. The Decoder wrote that after the startup Jev introduced its version, “OpenAI, Cloudflare, and Microsoft quickly followed with similar offerings.” The Herald reported OpenAI’s beta on October 9. This article sets Google’s own figures beside OpenAI’s documentation and an independent evaluator’s data.

What Google announced

According to Google’s post:

• Request shape. One shared input of up to 256K tokens plus up to 1,000 independent questions. The question types are noul (yes/no), choice (2 to 255 options) and score (an ordered rubric). Every answer includes a confidence score from 0 to 1 “so your code can route ambiguous cases to a fallback or human reviewer.”
• Speed. Google reports median model compute of 50-100 ms and 150-250 ms end to end, for contexts up to 16K tokens, “remaining nearly flat from 1 to 100 questions per request.”
• Pricing. $0.10 per 1M input tokens and $0 for output. The shared input is billed once per request regardless of the number of questions, and an API no-cost tier is available in Google AI Studio.
• Status. The API is available through the Gemini API in Google AI Studio. The post’s text does not use the words “preview” or “general availability”; its Python sample sets the API version to “v1beta”. Support in Gemini on Google Cloud is “coming soon,” and InfoWorld notes that Google has not specified a release date.
• Internal uses. Gemini Notebook uses it to route queries, which Google says serves everyday queries “with up to 40% lower latency.” The post also describes a “Safety layer for MCP tools in Search” that detects adversarial behavior “with high efficacy,” and gives no figure for that.

Google’s benchmark table

The post says Gemini Decision 1 “leads across the suite, with the lowest calibration error.” Its comparison table is an image, which the Herald read directly; MarkTechPost transcribes it and says all its figures are “self-reported by Google.” The table compares Gemini Decision 1 with TypeSafe’s Jev 1.14, OpenAI’s GPT-6 Luna on its Decisions API, and Gemini 3.5 Flash-Lite on the Interactions API. In the table, as read by the Herald:

• Gemini Decision 1 is highest on seven of the nine accuracy rows, for example 92.2% on the JevBench public split against 90.0% for Jev 1.14 and 90.9% for GPT-6 Luna. MarkTechPost reports the same count.
• Jev 1.14 is higher on Mind2Web Multimodal (86.8% against 86.2%) and When2Call (82.0% against 79.9%). The MMAU and Video-MME cells for Jev and for OpenAI are dashes.
• Top-1 ECE, marked lower-is-better, is 0.0250 for Gemini Decision 1, 0.0494 for Jev 1.14 and 0.0721 for GPT-6 Luna.
• The total price on seven shared evaluations is $0.88 for Gemini Decision 1, $0.68 for Jev 1.14, $0.93 for GPT-6 Luna and $4.14 for Gemini 3.5 Flash-Lite. Google’s “nearly 5x cheaper” claim is against Flash-Lite; MarkTechPost notes that Jev costs less than Gemini Decision 1 in this table.

The post’s text names neither Jev nor OpenAI, so its “leads across the suite” is a broader summary than the table, where Jev 1.14 is ahead on two accuracy rows and on total price. A footnote says refusals, safety blocks and requests a model could not accept are scored as incorrect.

An independent evaluator’s figures

Artificial Analysis publishes a Decision Index (v2.1), a composite of six evaluations, and states that “All evaluations are conducted independently by Artificial Analysis.” The page’s charts are drawn by script and the page shows no date, so the values below are the Herald’s reading, on November 12, of the data embedded in the page’s HTML. In that data:

• The index score is 65.98 for Gemini Decision 1, 66.23 for Jev 1.14 and 60.25 for “GPT-6 Luna (Decisions).”
• Time per task is 210.85 ms for Gemini Decision 1, 92.05 ms for Jev 1.14 and 206.15 ms for GPT-6 Luna. Google’s own range of 150-250 ms contains the first figure; the pages the Herald read do not describe the two measurements as using the same method.
• Cost per task is $0.000345 for Gemini Decision 1, $0.000156 for Jev 1.14 and $0.000339 for GPT-6 Luna; the Herald’s own conversion to the page’s per-1M-tasks unit gives about $345, $156 and $339.
• The Brier score (the page defines it as the mean squared gap between a chosen answer’s probability and whether it is right; lower is better) is 0.175 for Gemini Decision 1, 0.187 for Jev 1.14 and 0.186 for GPT-6 Luna. Google’s table uses a different measure, Top-1 ECE, so the two are not directly comparable.

Comparison with OpenAI’s Decisions API

OpenAI’s guide, read as markdown on November 12, says “The Decisions API is now generally available,” and its changelog lists the beta on October 6 and general availability on October 22. According to the guide:

• gpt-6-luna is “the only currently supported model,” and a request takes up to 400 questions, “up from 200 during the public beta.” InfoWorld gives the same 400 against Google’s 1,000.
• Inputs are text, images and JSON, and “Files, audio, and item references aren’t supported.” InfoWorld writes that Google’s offering supports a broader range, including audio and video.
• Uncached input costs “$0.10 per 1M tokens” and output tokens are not charged, the same list price Google gives. OpenAI’s guide says prompt cache hits receive discounted input pricing. MarkTechPost says Gemini Decision 1 has “No caching or Batch API yet”; Google’s post does not mention caching, and the Herald could not read Google’s documentation pages, which load by script, so that statement rests on MarkTechPost alone.
• OpenAI describes its speed as “about 10x faster than the Responses API,” a relative claim, while Google gives absolute latencies. The two cannot be compared directly.

Analyst views and open questions

InfoWorld quotes Lian Jye Su, chief analyst at Omdia, who said “The model is great at generating decisions from unstructured, dynamic input.” According to InfoWorld, Su argued that developers should look beyond published benchmarks and assess reliability under actual operating conditions, and warned that manipulated inputs could cause incorrect judgments. Tulika Sheel, senior vice president at Kadence International, said “As AI agents become more autonomous, enterprises will need clear accountability for decisions and actions taken on behalf of users.”

Not settled by the sources read: MarkTechPost’s comparison table lists parameters as “Not disclosed” for Gemini Decision 1, GPT-6 Luna and Jev, and Google’s post gives no figure for the Search safety layer’s “high efficacy.”