<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>Gemini on Liam Ding</title>
    <link>https://www.liamding.com/en/tags/gemini/</link>
    <description>Recent content in Gemini on Liam Ding</description>
    <image>
      <title>Liam Ding</title>
      <url>https://www.liamding.com/img/og-cover.png</url>
      <link>https://www.liamding.com/img/og-cover.png</link>
    </image>
    <generator>Hugo -- 0.147.7</generator>
    <language>en</language>
    <lastBuildDate>Fri, 13 Nov 2026 00:00:00 +0800</lastBuildDate>
    <atom:link href="https://www.liamding.com/en/tags/gemini/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Gemini Decisions API: one request is cheap, repeating yourself is not</title>
      <link>https://www.liamding.com/en/posts/2026-11-13-gemini-decisions-api-cheap-checklist-full-price-repeats/</link>
      <pubDate>Fri, 13 Nov 2026 00:00:00 +0800</pubDate>
      <guid>https://www.liamding.com/en/posts/2026-11-13-gemini-decisions-api-cheap-checklist-full-price-repeats/</guid>
      <description>Billing the request instead of the question is right for Gemini&amp;#39;s Decisions API. Shipping without caching leaves reference-heavy workloads paying full price every time.</description>
      <content:encoded><![CDATA[<h2 id="the-problem">The problem</h2>
<p>On 11 November 2026 Google released the Gemini Decisions API and its model, Gemini Decision 1. It does not write text. You send one shared input, which can mix text, images, audio, video and PDFs, plus a map of questions: yes/no checks, multiple choice with 2 to 255 options, or a score on a rubric you define. It returns a probability and a confidence score for each answer. Google&rsquo;s docs allow 1 to 1,000 questions per request, bounded by a 262,144-token context window, and the launch post quotes 150–250 ms end to end for inputs up to 16K tokens.</p>
<p>Google is not first. OpenAI made its own Decisions API generally available on 22 October, on GPT-6 Luna. TypeSafe&rsquo;s Jev already sells the same three question types. Simon Willison notes that Google even reuses Jev&rsquo;s word &ldquo;noul&rdquo; for a yes/no question. So the model is not the interesting decision. The price sheet is.</p>
<p>Google priced it at $0.10 per million input tokens and nothing for output. More importantly, it picked the unit. The shared input is billed once per request, however many questions ride on it; each question adds only its own instructions and criteria. The same pricing page also says what is missing: no context caching on either tier, no Batch API. The Decisions API docs repeat it in plain words: repeated input across requests counts in full.</p>
<p>That is one product decision with two faces. Asking many questions about one thing in one call is very cheap. Asking the same reference material again and again, across calls, gets no discount at all.</p>
<h2 id="what-i-concluded">What I concluded</h2>
<p>Billing the request instead of the question is the right call. Leaving out caching is the part that is not finished.</p>
<p>The first half fits what Google is actually selling. The pitch is one messy item arriving in several formats, such as a scanned claim with damage photos and a call recording, checked against a whole checklist at once. Per-request billing makes that the cheapest way to use the product. VentureBeat worked the example: a 4,000-token record checked by 20 questions of 50 tokens each costs about $0.50 per 1,000 records as one request, and $8.10 if each check is its own call that resends the record. My arithmetic agrees: 4,000 + 20 × 50 = 5,000 tokens, against 20 × 4,050 = 81,000 tokens, a factor of about 16. The price list teaches developers to batch independent checks, which is also the design Google&rsquo;s docs recommend. Pricing and API shape point the same way.</p>
<p>The second half is a gap. A lot of decision traffic is the opposite shape: one long reference, like a policy manual, a product catalog or a reference image, judged against a stream of different items. On Gemini that reference is billed in full every time. OpenAI already shipped the fix: since its beta, the Decisions API reuses a matching prefix of shared input at cache-read prices. On OpenAI&rsquo;s pricing page, cached GPT-6 Luna input is $0.01 per million tokens against $0.10 uncached, and the 30-minute cache write is $0.125 per million. Take a 10,000-token reference sent with 1,000 requests, counting only that reference and not the tokens each question adds. That is 10 million reference tokens. On Gemini that is $1.00. The OpenAI side also assumes all 1,000 requests land while the cache entry is still valid, and that the identical prefix is at least 1,024 tokens. This reference is 10,000 tokens, so it clears that bar. The caching guide sets 30 minutes as the default minimum lifetime, at a write rate of 1.25× uncached input, and a cache read refreshes that coverage for at least another 30 minutes. The write rate is the whole charge for those tokens, not a fee on top of the uncached rate. If every request after the first hits the cache, it is about $0.10 of cache reads plus one cache write of about $0.0013. That is my arithmetic, on assumed token counts and a perfect hit rate inside the cache lifetime, but the order of magnitude is the point. At the same headline price, the cheaper vendor depends on whether your traffic is &ldquo;one item, many questions&rdquo; or &ldquo;one reference, many items.&rdquo;</p>
<p>There is a third detail that will bite people switching vendors. Google&rsquo;s docs say yes/no probabilities always fall between 0.01 and 0.99, so a threshold above 0.99 never matches. Simon Willison writes, as a general point, that &ldquo;Jev and OpenAI both return 1.0 or 0.0 for yes/no questions.&rdquo; That sentence is not a video result. The video test on the same page is Gemini only: he sent his Yonkers stomping-robot clip with the questions, and Gemini returned about 0.92 for a robot and about 0.05 for a dragon. TypeSafe&rsquo;s docs say Jev does not natively accept video. A rule written as &ldquo;auto-approve at 1.0&rdquo; silently becomes &ldquo;never auto-approve&rdquo; on Gemini. Google does say this in its docs. It belongs on the migration checklist too.</p>
<h2 id="constraints">Constraints</h2>
<ul>
<li><strong>The model scores options; it does not generate.</strong> That is why output is free and why latency is low. It also means anything open-ended, like extracting an invoice number or writing an explanation, still needs a separate generation call.</li>
<li><strong>Questions are independent.</strong> Only OpenAI&rsquo;s docs require a separate request when a decision depends on an earlier answer. Google&rsquo;s docs give both options: &ldquo;Either send a second request after you read the first answer, or ask both questions in one request and ignore the answer you don&rsquo;t need.&rdquo; On Gemini, you pay for that evidence again only when the follow-up is split into a second request.</li>
<li><strong>Multimodal input is the differentiator.</strong> Google charges $0.10 per million tokens for text, image, video, audio and PDF alike, with no separate transcription or OCR charge. TypeSafe&rsquo;s docs say Jev reads text and JSON, plus PNG or JPEG images anywhere in the request, and does not natively accept audio or video. OpenAI takes text, images and JSON. If your input is a recording or a video, Google is one hosted option. VentureBeat also names Cloudflare&rsquo;s Clef-Omni, hosted on Workers AI at $0.15 per million input tokens with a 64,000-token window, and says to compare Gemini and Clef-Omni on the same clips.</li>
<li><strong>The category already has a price.</strong> OpenAI is at $0.10 uncached with free output. TypeSafe lists Jev at $0.042 per million input tokens. Among Google, OpenAI and Jev, that is the lowest price. Google matched the incumbent&rsquo;s headline instead of undercutting it.</li>
</ul>
<h2 id="the-options-and-what-i-traded-away">The options, and what I traded away</h2>
<p><strong>Bill per question.</strong> Simple to explain, and it would charge a 1,000-question request more than a one-question request. It would also kill the core pitch, because the whole point of one forward pass over shared input is that extra questions cost almost nothing to run. Rightly rejected.</p>
<p><strong>Bill the request, with no caching (what Google shipped).</strong> Clean, predictable, and it rewards the batched multimodal workload Google wants to own. The cost lands on reference-heavy workloads, and on a dependent step only when that step is split into a second request.</p>
<p><strong>Bill the request, with caching (what OpenAI ships).</strong> Covers both traffic shapes at the same headline price. The cost is complexity: cache writes, prefix rules and a hit rate the buyer has to manage. OpenAI&rsquo;s docs note that the reusable prefix ends at the shared input, and anything inside a question sits outside it.</p>
<p>Google took the second option. I would have taken the third, or at least published a date for it.</p>
<h2 id="evidence">Evidence</h2>
<ul>
<li>Google&rsquo;s launch post and pricing page: $0.10 per million input tokens, $0 output, shared input billed once per request, each question adds only its own tokens, no thinking-token, transcription or OCR charges.</li>
<li>Google&rsquo;s pricing page and Decisions API docs: context caching &ldquo;Not available&rdquo;, no Batch API, no Flex or Priority inference, repeated input across requests counts in full.</li>
<li>Google&rsquo;s launch post: on its shared benchmark suite, Decision 1 is nearly 5x cheaper than Gemini 3.5 Flash-Lite through the Interactions API. VentureBeat gives the run costs: $0.88 for Decision 1, $0.68 for Jev 1.14 and $4.14 for Flash-Lite.</li>
<li>Google&rsquo;s launch post: Gemini Notebook uses the API to route queries to a model tier and thinking level, with up to 40% lower latency for everyday queries.</li>
<li>VentureBeat: Decision 1 had the lowest calibration error in Google&rsquo;s comparison, 0.0250, about half of Jev&rsquo;s 0.0494.</li>
<li>OpenAI&rsquo;s Decisions docs, pricing page and prompt caching guide: GPT-6 Luna only, 1 to 400 questions per request, prompt caching on a shared prefix of at least 1,024 tokens, $0.10 uncached and $0.01 cached input per million tokens, 30-minute cache write $0.125 per million, output not charged. A decision that depends on an earlier answer goes in a separate request.</li>
<li>VentureBeat: Cloudflare&rsquo;s Clef-Omni, on Workers AI, takes audio and video at $0.15 per million input tokens with a 64,000-token window. For recordings and video, compare it with Gemini on the same clips.</li>
<li>Google&rsquo;s Decisions API docs: yes/no probabilities are bounded to 0.01–0.99.</li>
</ul>
<h2 id="where-this-stops-being-true">Where this stops being true</h2>
<ul>
<li><strong>If your reference material is short,</strong> caching barely matters and Google&rsquo;s pricing is simply the same price as OpenAI&rsquo;s, with more input types and a higher question limit (1,000 against 400).</li>
<li><strong>If your traffic really is one item with many checks,</strong> per-request billing is close to ideal and the missing cache costs you nothing.</li>
<li><strong>If your input is plain text,</strong> the price question shifts. TypeSafe lists Jev at $0.042 per million input tokens. Among Google, OpenAI and Jev, that is the lowest price. Artificial Analysis lists Drex Max v1.1&rsquo;s input price at $0.03 per million tokens, and says the cost it shows on the index reflects a temporary launch price. At the standard price, that cost would be double. VentureBeat reports that the independent Artificial Analysis Decision Index v2.1 put Gemini ninth of 18 at launch, with a score of 65.98 against Jev&rsquo;s 66.23 and an estimated $345 per million tasks against Jev&rsquo;s $156. Checked on 12 Nov 2026, the Decision Index page&rsquo;s charts are labeled &ldquo;18 of 19 models,&rdquo; and its decision-model list has 19 names in score order, with Gemini&rsquo;s 65.98 in tenth place. The Gemini model page on the same site still shows a summary of #9 of 18, with the score rounded to 66. Google&rsquo;s own benchmark lead depends on which test you look at.</li>
<li><strong>If Google adds caching soon,</strong> this critique shrinks to a launch-day gap. The launch post already says Gemini in Google Cloud support is &ldquo;coming soon&rdquo;; caching may follow, but Google has not said so.</li>
</ul>
<h2 id="what-i-would-do-differently">What I would do differently</h2>
<ol>
<li>Ship context caching for shared input, or put a date on it. &ldquo;Not available&rdquo; on the pricing page is honest; a date would let buyers plan.</li>
<li>Add a pricing example for the reference-heavy case next to the batched one, so buyers can see which shape they are in before they build.</li>
<li>Keep the 0.01–0.99 bound prominent wherever thresholds come up, and suggest a default threshold for teams migrating from vendors that return 1.0.</li>
<li>Convert media into money on the pricing page. The Gemini Embedding 2 entry on the same page shows per-image and per-second costs; the Decision 1 entry shows only the per-token rate. One minute of audio or one PDF page, priced out, would let buyers estimate cost per case.</li>
<li>For buyers today: measure your own traffic shape first. If one item carries many checks, especially with media, Gemini is a strong first pilot. If one long reference meets many items, price OpenAI&rsquo;s cached path next to it before you choose. Either way, set thresholds per model, not per vendor-neutral constant.</li>
</ol>
<p>Google made the checklist cheap and fast, and the pricing page tells developers exactly how to use it. The reference manual that every request has to repeat still has no discount, and that is the next line Google needs to add to the page.</p>
]]></content:encoded>
    </item>
  </channel>
</rss>
