Provenance Record

Well-structured 1001-word News article covering a significant AI model launch. The benchmark table is clear and appropriately nuanced — the article correctly presents wins and losses rather than only positive results. The 'What We Don't Know' section shows good editorial judgment by flagging the MCP Atlas vs. MMMU-Pro labeling inconsistency. Writing is professional and technically precise throughout.

Read source-0 (TechCrunch, 200 OK), source-1 (BigGo Finance, 200 OK), source-2 (FelloAI, 200 OK), source-3 (LLM-Stats, 200 OK), source-4 (DataCamp, 403 BLOCKED — not verified from snapshot), source-5 (Google official blog, 200 OK). Key findings: (1) Benchmark figures confirmed — Terminal-Bench 2.1 (76.2%/70.3%) confirmed in LLM-Stats (source-3) and FelloAI (source-2); GDPval-AA Elo (1,656/1,314) confirmed in FelloAI, LLM-Stats, and Google blog; MCP Atlas (83.6%/78.2%) confirmed in LLM-Stats and Google blog; Finance Agent v2 (57.9%/43.0%) confirmed in LLM-Stats; ARC-AGI-2 (72.1%/77.1%), HLE (40.2%/44.4%) confirmed in FelloAI. (2) Pricing ($1.50/$9.00 input/output, $0.15 cached) confirmed in FelloAI and LLM-Stats; Gemini 3.1 Pro pricing ($2.50/$15.00) confirmed in FelloAI. (3) Official Google quote 'Our strongest agentic and coding model yet...' confirmed as body text in Google blog (source-5), not a labeled pull quote. (4) Koray Kavukcuoglu title ('CTO of Google DeepMind and the company's Chief AI Architect') confirmed in Google blog. (5) Tulsee Doshi 'orchestrator, your planner' quote confirmed in TechCrunch. (6) Enterprise customers (Shopify, Macquarie, Salesforce, Xero, Databricks, Ramp) attributed to DataCamp but DataCamp (source-4) was blocked; however, all six enterprises are confirmed in the Google official blog (source-5). (7) Gemini 3.5 Pro 'in internal testing' claim: LLM-Stats says 'in internal use and slated for next month'; FelloAI says 'pushed to next month' with audience reaction at I/O; the article's phrasing is a reasonable paraphrase. (8) TWO UNVERIFIED CLAIMS: (a) '289 tokens per second' — no snapshot contains this specific figure; sources say '4x faster' in tokens/sec but do not give a numeric baseline of ~72 tokens/sec to derive 289. (b) '230 countries and 70 languages, according to TechCrunch' — the TechCrunch snapshot does not contain these specific numbers; the Google blog says 'billions of people globally' without specific figures. These claims require corrections.

All major benchmark numbers are confirmed by multiple sources and Google's own blog. Pricing figures are confirmed. The sensationalist-language flag ('incredible') is a false positive — it appears inside a direct verbatim quote from Koray Kavukcuoglu, not as editorial language. The BigGo Finance source attributes 83.6% to MMMU-Pro multimodal while Google's blog attributes it to MCP Atlas (tool-use); the article correctly flags this discrepancy in the 'What We Don't Know' section, which is good editorial transparency. Two specific figures lack source support: '289 tokens per second' and '230 countries and 70 languages attributed to TechCrunch'.

Substantively strong, well-researched article with accurate benchmark reporting confirmed across multiple independent sources and the official Google announcement. Two minor specific claims are not traceable to their cited sources: a token-throughput figure (289 t/s) and an availability statistic (230 countries, 70 languages) attributed to TechCrunch. Both are subordinate claims in body sections — not in the headline, summary, or lead. A two-item corrections record is sufficient to publish this article with an honest trail. APPROVE_WITH_CORRECTIONS.