Source Verification: All 17 snapshots were read from sources/2026-11/ (gunzip, tag-stripped text plus raw HTML) and none carries suspicious_patterns. The scanner did not flag anything, so there is nothing to override. Anthropic disclosure: required - the model card is discussed in a passage of its own that counts the rows where Claude Sonnet 5.5, the Herald's contributor model, scores higher than Hydrogen; the closing 'Open points' section carries the one-sentence disclosure ('The Herald's contributor model is built by Anthropic, whose Claude Sonnet 5.5 is one of the comparison columns in Google's model card.'), it is neutral and characterises nothing. The Claude comparison is strictly attributive ('By the Herald's count', figures taken from the card). Reviewer note: the reviewing model is also Claude Sonnet 5.5; the check was done from the model card's raw table. (1) HEADLINE, read as raw text: AA pages were fetched by the pipeline at 2026-11-13 07:48 UTC and again by the reviewer live with a browser User-Agent at 07:49 UTC on 2026-11-13 (sha256 of the live gemini-4-hydrogen-high page ca3abe23f077921390b9d6cb98623c70d0dcf2584d249fd1c5aff9a4882433f8, gemini-4-argon page 564f39a4fa24201851fb29183fa1e5d8dd2e5e56bc3c4b0c0981e177c61422b6). source-6 (page titled 'Gemini 4 Hydrogen (high)'): Intelligence Index 51, #21/78, 7.9 minutes per task, $1.87 cost per task, In $1.50 / Out $7.50. source-4 ('Gemini 4 Argon (high)'): 50, #22/78, 30.0 minutes, $4.97, In $2.00 / Out $10.00. Both rows are the high setting; $1.87 belongs to Hydrogen and $4.97 to Argon (not swapped); the live page text-4xl tiles match (51/7.9/$1.87 and 50/30.0/$4.97). The headline attribution to AA is correct (AA scores, AA cost per Intelligence Index task v5.0.1). 'High' is the default per Google's post and the Google Cloud Hydrogen page. (2) AA TABLE: source-7 low 27/$0.29/1.3; source-8 medium 41/$0.80/3.7; source-6 high 51/$1.87/7.9; source-5 xhigh 54/$3.16/13.7 (#16/78); source-4 Argon high 50/$4.97/30.0; source-3 3.8 Flash medium 30/$6.26/13.0; source-2 3.8 Flash high 31/$7.77/14.7. All 7 rows and the ranks (21st, 22nd, 16th of 78) match. AA token prices: Hydrogen $1.50/$7.50, Argon $2.00/$10.00, 3.8 Flash $0.75/$3.75 - match. AA summaries quoted (xhigh 'slower than average per task and very verbose'; high 'quicker than average per task, however somewhat verbose'; medium 'notably quick per task and fairly concise') are verbatim. AA defines time per task as output tokens divided by output speed - matches the article. (3) GOOGLE POST (source-10, Nov 12, 2026): quoted sentences ('priced flat across its entire 1 million-token context window at $1.50 per million input tokens and $7.50 per million output tokens'; 'surpasses Argon on the Artificial Analysis Intelligence Index at its default high thinking level (51 vs. 50, and reaching 54 at its highest xhigh thinking setting)'; 'matches or outperforms competing models in its class across core developer and enterprise workflows'; 'often finishing tasks faster and using fewer output tokens'; free users 'via auto model selection when it's the appropriate model for the prompt') are verbatim; 2.5 times faster at high default, 57% vs 18.2% on Terminal-Bench 4.1, four thinking levels low/medium/high (default)/xhigh confirmed; 'all tiers of Google AI Studio and the Gemini API' confirmed. Neither the post nor the card defines 'its class' (searched). (4) MODEL CARD (source-13), recomputed: ten rows (SWE-sweep, Terminal-Bench 4.1, Agents' Last Exam v1.1, AutomationBench-AA, Vending-Bench 2, OSWorld-2.1, MKA Climb WSR, MKA Chartography, Terminal-Bench-Science 0.1, HLE-Diamond). Sonnet 5.5 higher than Hydrogen on exactly three: Terminal-Bench 4.1 63.9 vs 57.0, OSWorld-2.1 83.9 vs 77.6, Terminal-Bench-Science 0.1 59.9 vs 51.4 - matches. GPT-6.1 Terra higher on exactly one: Terminal-Bench-Science 0.1 58.3 (its Terminal-Bench 4.1 cell is a dash) - matches. Card quote 'at times the model may use more tokens to maximize performance, especially at higher thinking levels' and 'will be available in the coming days' verbatim. The Terminal-Bench 4.1 row is labelled 'Cheating-adjusted' in the card; the article does not repeat the label (not material, see recommendations). (5) METHODOLOGY (source-12, a PDF at storage.googleapis.com; sha256 e6ef1f92efbce0a02975e0f7fd495b7c04114ba81bf9c12d7be6204f826c58d0 matches the manifest, text extracted with pdftotext): Terminal-Bench 4.1 Hydrogen 'self-computed using a mini-swe-agent vision harness ... using high effort level'; 'All the results for non-Gemini models are sourced from providers' self-reported numbers unless otherwise mentioned below' verbatim; 'AutomationBench results are sourced from the Artificial Analysis leaderboard' and 'HLE Diamond results are reported from the Artificial Analysis leaderboard' confirmed. (6) CARBON POST (source-9, Nov 10, 2026): 'Argon remains the ideal choice for the majority of software engineering and knowledge work' and 'we partnered with Artificial Analysis to launch the new Multimodal Knowledge Agent (MKA) benchmark' confirmed. (7) PRICE DISAGREEMENT: Ars (source-1) says '$1.50 per million input tokens and $7.50 per million output tokens, which is unchanged from 3.8 Flash' - verbatim; AA lists 3.8 Flash at $0.75/$3.75 - confirmed; Google Cloud pricing page (source-11, title 'Agent Platform Pricing | Google Cloud') says 'Gemini 3.8 Flash ... are offered with introductory pricing of $0.75 / $3.75 per 1M tokens input / output through December 31, 2026. Starting January 1, 2027, standard pricing of $1.5 / $7.5 ...' - verbatim, and the article says that page covers Google Cloud's platform, not necessarily the API Ars and AA describe. The same page lists Argon (and Carbon) at $2.00/$10.00 'through November 30, 2026' and $4.00/$20.00 'Starting December 1, 2026'; Ars: 'currently $2/$10 ... go up to $4/$20 starting next month' verbatim. 4.00/1.50 = 20.00/7.50 = 2.67, so 'about 2.7 times' is right. (8) THINKING DEFAULTS: Hydrogen page (source-15): 'thinking levels LOW, MEDIUM, HIGH (default), and XHIGH. Setting thinking_budget or thinking_level="MINIMAL" returns an HTTP 400 validation error'; sampling parameters ignored; implicit caching only; cache control with TTL from January 12, 2027, cache writes 1.25x - all confirmed. 3.8 Flash page (source-14): 'The supported values for 3.8 Flash are LOW, MEDIUM (default), and HIGH' - confirmed (the same page also says MINIMAL is unavailable for 3.8 Flash; the article does not claim otherwise). (9) SPEED ARITHMETIC: 14.7/7.9 = 1.86 (about 1.9) and 13.0/7.9 = 1.65 (about 1.6); labelled the Herald's arithmetic, Google's 2.5x labelled Google's, and the article says no cited page reconciles them. (10) OUTLETS: Ars Technica (source-1, Ryan Whitwam, Nov 12) covers the replacement of 3.8 Flash across Google products, 2.5x, 54 at xhigh vs 50 for Argon, the chart caption 'Hydrogen compared to the mid-tier options from OpenAI and Anthropic.' (credit Google), the free/AI Plus statement and the price lines - all as cited. The Decoder (source-16, Matthias Bastian, Nov 12) writes 'Hydrogen should make that advice moot' - verbatim. 9to5Google (source-0, Abner Li, Nov 12) reports free users get Hydrogen via Auto - as cited. AA is a third-party evaluator, and the article discloses that Google's own table figures (AutomationBench, HLE-Diamond) come from the AA leaderboard and that Google partnered with AA on MKA. Independence for a News piece: Ars, The Decoder, 9to5Google and AA are four non-Google outlets, though 9to5Google largely quotes Google's post. (11) NOT FOUND: the quotation attributed to Google Cloud's Hydrogen page, 'a short-term generally available (GA) release subject to a minimum 45-day retirement notice policy', appears on neither the stored snapshot (source-15) nor a live fetch of that URL on 2026-11-13 07:51 UTC (sha256 4e1b73a9921371325629e32e96cefe66ba145259f996f47400a8d2407cb209df; searched for 'short-term', '45-day', 'retirement', 'notice', 'minimum'); the page gives only 'Launch stage: GA' and 'Release date: November 12, 2026'. The substance exists on a page the article does not cite: Google Cloud's model-versions page (cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions, read live) lists gemini-4-hydrogen under 'Models available for shorter availability periods' and says 'Short-term availability models remain active until a replacement model is released and a retirement date is announced. When we schedule a model for retirement, we post a fixed date ... that gives you at least 45 days to migrate.' Filed as a correction. (12) PROMOTIONAL MECHANISM: the Google Cloud pricing page footnote reads 'Promotional pricing provided through 50% credits back on net spend on select models within a given period', and the 3.8 Flash and Argon through-date rows carry that asterisk; the article reports the listed figures accurately but does not say the promotional prices are delivered as credits. Filed as a clarification. SNAPSHOT NOTES: source-12 is a PDF reached by redirect to storage.googleapis.com (info finding, same Google methodology document, not a concern). Snapshot files read: source-0 to source-16 (.html.gz), manifest.json. No WebFetch fallback was needed. No hosts added to the allowlist.