Provenance Record

Well structured news piece that keeps vendor statements attributed and separates what Google says from what is not known. Dense but readable; the two verified Google-versus-Google discrepancies (Terminal-Bench 4.0/4.1; enterprise access) are the substance and are correct. Two secondary sentences are imprecise (see concerns).

Snapshots read from disk (gunzip -c, HTML to text): source-0 Antigravity blog 'Gemini 4 Hydrogen and Orchestrator' (The Antigravity Team, Nov 12, 2026; HTTP 200; sha256 a0d4487bc7b6...), source-1 antigravity.google/docs/models (200; c5bf91076f50...), source-2 antigravity.google/docs/enterprise (200; 02db56632233...), source-3 antigravity.google/docs/changelog (200; f645e1b001f4...), source-4 blog.google 'Introducing Gemini 4 Hydrogen' (200; f7dc6855b119...), source-5 WinBuzzer (Markus Kasanmascheff, Nov 12 2026 11:35 pm CET; 200; b821d2bf88274...), source-6 SiliconANGLE (200), source-7 The Decoder (200), source-8 9to5Google (Abner Li, Nov 12 2026 10:01 am PT; 200), source-9 and source-10 GitHub releases v4.0.0 and v4.1.0 (200). No archive fallbacks; suspicious_patterns null for all except source-3. RAW-TEXT ROUTE FOR antigravity.google: the four Antigravity pages are server-rendered (the stored HTML contains the full text, no JavaScript needed). I also fetched all four live with curl and a browser User-Agent; the sha256 of every live response equals the manifest sha256 (blog a0d4487b..., models c5bf9107..., enterprise 02db5663..., changelog f645e1b0...), so I read exactly the version the script stored. No Archive.org route needed. ALLOWLIST DECISION: antigravity.google was not on the allowlist (4 of 11 sources). Confirmed as Google's official Antigravity site: the blog.google Hydrogen post links 'Google Antigravity' to https://antigravity.google/ (twice) and links 'Orchestrator in Google Antigravity' to the Antigravity launch post; 9to5Google links the launch post; WinBuzzer links the blog, models, enterprise and plans pages; the Antigravity pages link back to blog.google, about.google, policies.google.com and docs.cloud.google.com. Host added to config/source_allowlist.txt with a justification comment limited to first-party statements; no version bump needed. PROMPT-INJECTION SCANNER (source-3, pattern system-prompt-reference): the matched excerpt is '...Fixed conversations with certain models on a Google Cloud project failing for the rest of the conversation once the system prompt, tools, or an earlier message changed after a thinking block; outdated thinking blocks are now dropped instead of rejec...'. Read in context in the raw HTML it is one <li> bug-fix bullet in the Antigravity release notes describing an API request error; it does not address an AI reader, gives no instruction and asks for no concealment. False positive; auto-REJECT overridden after reading the passage in the decompressed snapshot. No other source was flagged. HEADLINE / SUMMARY / LEAD, CLAIM BY CLAIM (raw text): (1) 'Rolls Out Orchestrator in Antigravity', Nov 12: Antigravity blog dated 'Nov 12, 2026': 'Today, we're rolling out Gemini 4 Hydrogen inside Google Antigravity alongside a new model dropdown option called Orchestrator'. Basis: staged by plan. 'Orchestrator is rolling out as Orchestrator (Carbon) and Orchestrator (Argon) for Google AI Ultra subscribers, and as Orchestrator (Argon) for paid Google AI Pro subscribers'; models page adds Pro trials on eligible premium devices; changelog entry 'Model launch - November 12, 2026'. The article reports the plan split and the changelog's gradual-rollout sentence ('New versions roll out gradually and may take a few days to reach all users'), which is page-level text about app versions, not a Hydrogen-specific statement; the article attributes it correctly ('The changelog says new versions...'). Verified. (2) 'Carbon or Argon planner with Hydrogen': supported by BOTH the Antigravity blog (Getting started: 'Orchestrator (Carbon) and Orchestrator (Argon)') and the models page table ('Orchestrator (Carbon) / Planner Gemini 4 Carbon / Implementation Gemini 4 Hydrogen'; same for Argon). It is not docs-only. The changelog says 'Orchestrator for dynamic multi-model planning and execution with Gemini 4 Carbon'; the Google blog says 'dynamically switches between models like Carbon and Hydrogen'; WinBuzzer, SiliconANGLE and The Decoder describe Carbon and Hydrogen only (grep: WinBuzzer has no 'Argon' in its article text). The article says exactly this in the body. None of those pages says Carbon is the only planner, so the article's statement that Google's pages 'disagree ... on which planner options exist' (Overview) and 'differ' (summary) overstates an omission as a conflict: clarification filed. The headline itself is correct. (3) 'Two Google Posts Name Different Terminal-Bench Versions': Antigravity blog: 'On Terminal-Bench 4.1, Orchestrator reached near-Carbon task success rates while, on average, consuming 24% less quota and completing 19% faster.' Google blog: 'On Terminal-Bench 4.0, Orchestrator reached near-Carbon task success rates while, on average, consuming 24% less quota and completing 19% faster.' Same sentence, different version. Nothing on either page, in the changelog or in the four outlets reconciles them (the same Google blog post uses 'Terminal-Bench 4.1 scores (57% vs. 18.2%)' for standalone Hydrogen, which the article does not mention; it does not change the finding). The article does not say which version Google used, and says so. GitHub: v4.0.0 released 2026-08-26T04:48:12Z (datetime attribute), v4.1.0 released 2026-11-03T20:39:05Z; v4.1.0 notes read 'Tasks added or removed: None.' and 'Tasks modified (66)'. All verified and labelled as Google's. OTHER DISAGREEMENTS: Enterprise: Antigravity blog says 'Enterprises can access both in Gemini for business'; the enterprise page's 'Supported models in Gemini Enterprise' table marks Orchestrator (Carbon) and (Argon) with a cross (Carbon, Argon, Hydrogen ticked). The models page plan table also marks both Orchestrator rows unsupported for Enterprise (the article cites only the enterprise page; the models page agrees). Third-page search: in the stored Antigravity blog, 'Gemini for business' links to docs.cloud.google.com/gemini/enterprise/docs/ai-developer-tools-overview, and in the Google blog to cloud.google.com/gemini-enterprise, i.e. the same product the enterprise page calls Gemini Enterprise. I fetched the Google Cloud overview page live (sha256 d07d3e492c14...): it does not mention Orchestrator, Hydrogen, Carbon or Argon, so no page explains the mismatch. The wording 'both' in the blog is ambiguous (the two Orchestrator options, or Hydrogen and Orchestrator); either reading conflicts with the page. The article reports both and says it is unknown how they relate: accepted. Standalone Hydrogen: blog 'Starting today, Gemini 4 Hydrogen is rolling out across paid Google AI Pro and Ultra subscribers'; models page ticks Hydrogen for 'Free & Google AI Plus'; WinBuzzer notes the same. Accepted; the article reports both with attribution. Plans: models page agrees with the blog (Argon on Ultra and paid Pro); WinBuzzer's 'reserved for Ultra in the model documentation' matches Carbon only: the article says so. OUTLET CHECKS / INDEPENDENCE: WinBuzzer: 'on average in Google's tests', 'than its Carbon planning model', 'supplies no fixed dollar saving', 'The Antigravity post identifies that test as Terminal-Bench 4.1; the broader Hydrogen launch identifies it as 4.0', enterprise mismatch, free-account listing, changelog warning: all present. WinBuzzer is the only outlet that read Google's documentation pages (it links the models, enterprise and plans pages). SiliconANGLE: 'Google's runs of the Terminal-Bench 4.0 coding test ... 24% less quota' (24% only, no 19%). The Decoder: 'Terminal-Bench 4.0 ... 24 percent fewer resources and cutting processing time by 19 percent'. 9to5Google: quotes 'Carbon that may offload routine steps to Hydrogen. Same Carbon-level quality, fewer tokens used.' and Google's launch text; gives neither the 24% nor the 19%. So the sentence 'SiliconANGLE, The Decoder and 9to5Google restate Google's figures' is imprecise: clarification filed. Absence claim 'none of the cited sources reports a measurement of Orchestrator made outside Google': checked all 11; the only outside measurements are of standalone Hydrogen (Artificial Analysis index in WinBuzzer and The Decoder; Vals AI 54.04% on Terminal-Bench 4.0 in WinBuzzer; AA 54 percent in The Decoder), none of Orchestrator. Holds. 'None of the pages ... gives an absolute success rate, a number of runs or a quota unit for Orchestrator': holds across all 11 (the 66-task and three-run figures in WinBuzzer are Vals AI's, for standalone Hydrogen). QUOTE CHECK: every quoted string was matched on the exact page it is attributed to: Antigravity blog ('pairs Hydrogen with our strongest models for long-horizon tasks', 'a static model router', the Terminal-Bench 4.1 sentence, 'Enterprises can access both in Gemini for business'); models page ('pairs two Gemini 4 models in a single conversation' across a line break, 'a verification step that fails twice', 'within a single response', 'isn't a router', 'runs with a fixed configuration', 'recommended default', 'every intermediate step needs maximum reasoning depth', agy --model "Orchestrator (Carbon)"); changelog ('with Gemini 4 Carbon', 'roll out gradually and may take a few days to reach all users'); WinBuzzer ('its Carbon planning model', 'supplies no fixed dollar saving', 'reserved for Ultra in the model documentation'); The Decoder ('fewer resources'); 9to5Google (the Orchestrator description); GitHub (v4.1.0 task lines). No quote was found on a neighbouring page only. Anthropic disclosure: not required - Anthropic and Claude are not the subject and appear nowhere in the headline, summary, lead or body; Claude models appear only as incidental rows in Google's model tables (models and enterprise pages) and in a benchmark comparison in WinBuzzer/The Decoder, none of which the article uses. The Terminal-Bench v4.1.0 task list contains a task name 'bun-sourcemap-leak' which the article does not reproduce. Dates and counts: Nov 12, 2026 date is the post date on the Antigravity blog and changelog (no timezone conflict arises); GitHub timestamps are UTC as stated. Title is 150 characters (counted: 150, within the limit); body 877 words (News 400-1200); 11 sources, all HTTPS; body URLs match the sources array (chief:review body_sources_match true); signature and payload hash valid.

Headline, summary and lead claims verified from raw text for Orchestrator rollout (Nov 12, staged by plan), the Carbon-or-Argon planner pairing (Antigravity blog and models page), and the 4.1 versus 4.0 discrepancy (exact sentences on each page, nothing reconciling them). The 24%/19% figures and the Terminal-Bench release dates and notes are correct and labelled as Google's. Two clarifications: (a) 'disagree on which planner options exist' overstates what is an omission on the changelog and Google blog; (b) 9to5Google (and for the 19%, SiliconANGLE) do not restate Google's figures.

Substantively accurate, well sourced and carefully hedged coverage of a vendor feature launch, with two verified inconsistencies in Google's own pages. Approved with two clarification records. Featured candidacy: 1 - accurate, defined-audience coverage (developers choosing Antigravity models) with useful vendor-versus-vendor discrepancies, but an incremental product feature rather than an industry-level story