Source Verification: Snapshots read via gunzip/HTML-to-text from sources/2026-11/<slug>/. source-0 (blog.google, sha256 ddcdb667...): confirms 1,000 questions, 256K shared input, $0.10 per 1M input tokens and $0 output, 'Your shared input is billed once per request', noul/choice (2 to 255)/score types, confidence 0 to 1 with the 'fallback or human reviewer' sentence, 50-100 ms compute and 150-250 ms end to end 'nearly flat from 1 to 100 questions' for contexts up to 16K, no-cost tier in Google AI Studio, 'with support coming soon to Gemini in Google Cloud', Gemini Notebook 'up to 40% lower latency', 'Safety layer for MCP tools in Search ... with high efficacy', 'leads across the suite, with the lowest calibration error', 'nearly 5x cheaper ... than Gemini 3.5 Flash-Lite with the Interactions API', Python sample with api_version v1beta. Page metadata: datePublished 2026-11-11T18:00:00+00:00 (displayed 'Nov 11, 2026'), so November 11 holds in both UTC and the page's displayed date. Absence claims verified by grep of the post text: no occurrence of Jev, OpenAI, TypeSafe, caching, batch, 'preview' or 'general availability' (the only 'beta' hit is v1beta). THE BENCHMARK TABLE: the table is an image (gemini-decision__benchmarks__tab.width-2000.format-webp, sha256 89b834db...5d); I downloaded it, converted it and read it directly. Every figure the article states matches: JevBench public split 92.2/90.0/90.9 (Gemini/Jev/Luna), Mind2Web Multimodal 86.2 vs Jev 86.8, When2Call 79.9 vs Jev 82.0, Gemini highlighted best on 7 of 9 accuracy rows, MMAU and Video-MME dashes for Jev and OpenAI, Top-1 ECE 0.0250/0.0494/0.0721 (0.2104 Flash-Lite, not cited), total price on 7 shared evals $0.88/$0.68/$0.93/$4.14, column headers 'TypeSafe Jev 1.14', 'GPT-6 Luna / OpenAI Decisions API', 'Gemini 3.5 Flash-Lite / Interactions API (minimal)', and the refusals/safety-blocks-scored-as-incorrect footnote. 4.14/0.88 = 4.7x, consistent with 'nearly 5x'. The headline's 'self-reported' is supported: the table is Google's own and MarkTechPost says 'All figures below are self-reported by Google'. The article's contrast between 'leads across the suite' and Jev 1.14 ahead on two accuracy rows and total price is accurate. source-1 (The Decoder, Matthias Bastian, Nov 11): confirms 1,000 questions, 150 to 250 ms, $0.10 per million input tokens, ~5x cheaper than Flash-Lite per Google, internal uses, and the verbatim sentence 'After the startup Jev introduced its version, OpenAI, Cloudflare, and Microsoft quickly followed with similar offerings.' (the article quotes the part from 'OpenAI' on, correctly). MARKTECHPOST (source-2): the automated snapshot failed (HTTP 403, Cloudflare 'Just a moment' page; browser-User-Agent curl also 403; Archive.org has no capture). Route used: the publisher's own RSS feed https://www.marktechpost.com/feed/ (HTTP 200, UA Mozilla/5.0 Chrome/124), which carries the full item with content:encoded; item sha256 e7a0a76f11c76c5ffadbca42daf132f7cbaf44a9d2e1b1153d4029929abae645 (pubDate Wed, 11 Nov 2026 21:21:42 +0000, link matches the cited URL). Raw text confirms: 'All figures below are self-reported by Google', a transcribed table matching the image, 'Leads 7 of 9 accuracy benchmarks', 'Jev 1.14 costs $0.68 vs $0.88 on the shared suite', the comparison table 'Parameters: Not disclosed' for all three, 'Prompt caching: Not supported' and 'No caching or Batch API yet, so repeated context is billed in full' for Gemini Decision 1. The article attributes the caching/Batch statement to MarkTechPost alone and says Google's docs could not be read, which is the honest framing (docs load by script; aistudio.google.com is not allowlisted; Google's post is silent on caching, confirmed by grep). source-3 (InfoWorld, Prasanth Aby Thomas, Nov 12): confirms 1,000 vs OpenAI's 400, same $0.10 list rate with no output charge, 'Google's offering supports a broader range of multimodal inputs than OpenAI's Decisions API, including audio and video', 'Google has not specified a release date' for Google Cloud, and both analyst quotes verbatim: Lian Jye Su, chief analyst at Omdia: 'The model is great at generating decisions from unstructured, dynamic input'; Tulika Sheel, senior vice president at Kadence International: 'As AI agents become more autonomous, enterprises will need clear accountability for decisions and actions taken on behalf of users'. The paraphrases of Su (look beyond published benchmarks, assess reliability under actual operating conditions, manipulated inputs could cause incorrect judgments) match the text and are labelled 'According to InfoWorld'. source-4 (Artificial Analysis Decision Index v2.1): the page text states six evaluations and 'All evaluations are conducted independently by Artificial Analysis'; no publication or update date appears in the HTML (no datePublished/dateModified), so 'undated' is correct. I parsed the embedded model data (Next.js payload) and confirmed: indexScore Gemini Decision 1 65.977, Jev 1.14 66.234, GPT-6 Luna (Decisions) 60.246; timePerTaskMs 210.85 / 92.05 / 206.15; costPerTaskUsd 0.0003454 / 0.0001562 / 0.0003387 (x1M = about $345 / $156 / $339, the Herald's conversion is right); Brier (index-level) 0.1752 / 0.1868 / 0.1859, i.e. 0.175 / 0.187 / 0.186. The Brier definition quoted matches the page. The article's own caveats (different measures from Google's ECE; Google's 150-250 ms vs AA's 210.85 ms not shown to use one method) are accurate. One omission found in the same data: the page lists 18 decision models by default (19 in the data, one older version hidden) and Gemini Decision 1 ranks 9th of 18 on the index; eight score higher (Drex Max v1.1 72.62, SubQ Snap 1.2 72.58, GLiDE v1.2 72.52, Mammoth 27B 71.36, d1-ultra 70.58, Decider v1.2 69.77, Elephant 9B 66.39, Jev 1.14 66.23). The article's three-model comparison is literally correct but never says other entries outscore both Gemini Decision 1 and Jev 1.14; this is filed as a clarification. source-5 (OpenAI Decisions guide): 'The Decisions API is now generally available'; 'gpt-6-luna, the only currently supported model' (the article quotes 'the only currently supported model'); 'Ask up to 400 questions per request, up from 200 during the public beta'; 'evaluates text, images, and JSON'; 'Files, audio, and item references aren't supported'; 'uncached input costs $0.10 per 1M tokens. Prompt cache hits receive discounted input pricing ... Output tokens are not charged'; 'about 10x faster than the Responses API'. The article states these no more strongly than the page (it also says 'list price', and the guide's note that regional-processing premiums and long-context multipliers apply is not contradicted). source-6 (OpenAI API changelog): Oct 6 'Released the Decisions API in beta with gpt-6-luna'; Oct 22 'now generally available with gpt-6-luna ... up to 400 questions, increased from 200 in public beta'. Dates match the article. Independence: The Decoder, InfoWorld, MarkTechPost and Artificial Analysis are four independent outlets/evaluators (MarkTechPost's table is a transcription of Google's, flagged as such); the two OpenAI pages are the comparator's own documentation, used for OpenAI's claims only. No suspicious_patterns in any manifest entry. Allowlist: all six distinct hosts (blog.google, the-decoder.com, marktechpost.com, infoworld.com, artificialanalysis.ai, developers.openai.com) are already on config/source_allowlist.txt; no additions. Anthropic disclosure: not required - the subject is Google's Gemini Decisions API; Anthropic and Anthropic products are not the subject, not in the headline, summary or lead, and there is no Anthropic credit passage.
Factual Accuracy: No wrong statement found in the headline, summary or lead. Headline facts verified against Google's post, The Decoder, InfoWorld and MarkTechPost: up to 1,000 questions per request and $0.10 per 1M input tokens, with scope stated in the body (Gemini Decision 1, input tokens, $0 output, shared input billed once per request, no-cost tier in AI Studio, v1beta, Google Cloud support 'coming soon'). The caching/Batch statement is rightly limited to MarkTechPost. Date: the post is dated November 11 in both its displayed date and UTC (18:00 UTC), so no time-zone discrepancy. Quantifier/absence tests (the grep list): 'only', 'neither', 'every', 'each', 'both' sentences all check out against every snapshot (the post names neither Jev nor OpenAI in its text; 'the only currently supported model' is OpenAI's wording). Derived figures recomputed (conversion to per-1M-tasks, 7-of-9 count, cost-per-task figures) and correct. The one subordinate gap: the summary's AA sentence is true for the three entries it names but omits that eight other decision models on the same page outscore Gemini Decision 1 (see source_verification); filed as a clarification, not a correction, because nothing stated is wrong.