Provenance Record

Compact, well-structured News item (416 words, within the 400-1200 range) with Overview, What We Know, Competitive Context and What We Don't Know sections. Neutral tone, no AI self-reference. Reviewer is a Claude model reviewing an Anthropic story; Anthropic's self-reported figures were not credited and every figure was checked against the snapshots.

Read all four snapshots from disk (gunzip, tag-stripped text); manifest suspicious_patterns is null for all four. source-0.html.gz (anthropic.com/claude-sonnet-5-5, dated September 28, 2026): confirms 70.6% Terminal-Bench 4.0 vs Sonnet 5's 10.3% (Opus 5.5 66.4% in footnoted table, not in article), 46.2% FrontierCode 1.1 at Max effort, 55.5% CursorBench 4.0, GDPval-AA 1844 vs 1846, OSWorld 2.1 80.1% vs 81.8%, $2/$10 pricing (Opus 5.5 $4/$20), 'costs up to 30% less per task' ('In our testing'), 30%+ faster, first Sonnet with cyber safeguards 'similar to those on Opus 5.5', claude-sonnet-5-5 identifier, AWS/Google Cloud/Azure, zero data retention, Haiku 5.5 'in the coming weeks'. The Anthropic page does not itself state Claude Code availability. source-1.html.gz (VentureBeat): confirms 'up to 30%' cost reduction attributed to 'Anthropic says', primarily fewer tokens and tool calls; $2/$10 vs Opus 5.5 $4/$20; GDPval-AA and OSWorld 2.1 figures ('Anthropic reports'); Box quote '2.4x faster and used 12% fewer total tokens' verbatim; Zendesk 20% faster; Lovable one third fewer tool calls; GPT-6 Sol at identical $2/$10; Gemini 3.8 Flash $0.75 input (an introductory price through end of 2026, which the article does not mention). VentureBeat states Anthropic supplied the customer testing and that it had asked Anthropic how the speed figure was measured with no answer yet. source-2.html.gz (9to5Mac, Sep 28 2026): confirms Sonnet 5.5 replaces Sonnet 5 'from June', follows Opus 5.5 'last week', 'up to 30% less for most work' (via an embedded @claudeai post), and 'now live in Claude Code' via an embedded post from Lydia Hallie, i.e. an embedded social post rather than 9to5Mac's own reporting. source-3.html.gz (SiliconANGLE, Sep 28 2026): confirms 70.6% vs 10.3% ('the company said'), Haiku 5.5 in the coming weeks. Launch date of Sept 28, 2026 confirmed by the Anthropic page, VentureBeat, 9to5Mac and SiliconANGLE. Exclusions held: no Opus 5.5 Terminal-Bench score, no default effort level, and no mention of invisible text watermarking (which appears only in the SiliconANGLE snapshot) in the article body. Internal link /article/2026-09/23-anthropic-launches-claude-opus-55-cutting-coding-agent-costs-40-while-arriving-same-day-in-github-copilot resolves to an existing published article. All body URLs are in article.sources and vice versa.

Every number in the article traces to the cited snapshots. Two attribution issues, both recoverable: (1) The Opus 5.5 sentence says it 'became Claude Code's default model'; the linked Machine Herald article and the Claude Code changelog it cites say it became the default Opus model, and none of the four cited sources speak to Claude Code's overall default. (2) Several figures are worded as outlet findings rather than Anthropic's own: SiliconANGLE 'reports' 70.6%, VentureBeat 'says' GDPval-AA and OSWorld scores, VentureBeat 'relays results from early users' that Anthropic supplied. The 'What We Don't Know' section does say the benchmark figures are reported by Anthropic, which limits the harm; no figure is presented as independently measured, but the in-line wording is looser than the sources. Neither issue touches the headline (which says 'Reporting'), the summary ('Anthropic reports') or the lead.

Substantively accurate and well sourced; all specific claims verified against snapshots. Published with a corrections record covering two attribution imprecisions that a short public note can honestly cover.