The Claude column runs on a model that passed its earliest retirement date yesterday, and when its web search fails the answer still looks normal.

The engine's `src/adapters.ts` line 15 sets the Claude column to `claude-sonnet-4-5`. Anthropic's status table lists `claude-sonnet-4-5-20250929` as "Active", not deprecated, with a tentative retirement date of "Not sooner than September 29, 2026". That date passed yesterday. The same page says Anthropic gives "at least 60 days' notice before model retirement for publicly released models", so nothing breaks this week. Separately, the web search tool page says a search error such as a rate limit still comes back as HTTP 200, with the error written inside the reply as `web_search_tool_result_error`, and that a search matching nothing "returns an empty content list, not an error". The adapter (lines 115 to 140) keeps only the text blocks and the citations and never looks for that error block.

OpenAI launched GPT-6.1 Sol for paid ChatGPT plans and in the API as `gpt-6.1-sol`. TechCrunch quotes OpenAI that "the model is not yet available in Chat", so the consumer ChatGPT a buyer uses is unchanged. The forum list of DevDay announcements (Agents API, Codex, plugins, "Sign in with ChatGPT", Pro 500 plan) has nothing on web search, citations or shopping. OpenAI's deprecations page, read today, does not list `gpt-5.1`, the model our ChatGPT column uses; its only September entry retires `gpt-5.4-cyber` on 1 October.

Perplexity's migration page is word for word the same as yesterday: still no model named for re-routed Sonar requests and no marker in the response. The changelog added `openai/gpt-6.1-sol` and `anthropic/claude-sonnet-5-5` to the Agent API, nothing on Sonar. The engine's `main` is still `06233ce` with `sonar` on line 14. Google's spam update still carries only its 24 September launch post.

Peec says it analysed 120 million chats from February to August 2026 across ChatGPT, AI Mode, Google AI Overviews, Gemini, Perplexity, Copilot, Grok and Claude. It treats each prompt as "a biased coin". Its figures: variance about 0.0225 for branded prompts and about 0.07 for non-branded ones (three times higher); 40% of non-branded prompts are "contested", named between 10% and 90% of the time; a 5-point swing in a visibility score has about an 11% chance of happening on any day and 0.56% over a week. Its advice, verbatim: "Choose more prompts rather than more runs to improve your visibility score's accuracy." Peec sells prompt tracking, and the post gives no breakdown by engine or industry.

AWR's page says it tracks "more than 1.75 million US SERPs each month". For July 2026: desktop position one 20.02% click-through overall against 9.69% on pages with an AI Overview; mobile position one 35.74% against 12.02%. SEJ's write-up of the Q2 edition gives desktop positions one and two combined at 29.05% without an overview and 10.04% with one, a gap 6.03 points wider than in Q1. SEJ describes the source as Search Console data and notes a Google logging error affected impressions from May 2025 to April 2026; AWR's page describes tracked results pages. The two descriptions do not obviously match, so treat the quarter-on-quarter change with care.

TIME reports "Google's AI Mode has racked up more than 1 billion monthly users, the company says", and, citing Similarweb, that AI Overviews appear in "forty percent of U.S. Google searches". No method is given for either.