--- title: "Do AI assistants know when a product changed? October 2026" canonical_url: https://cookiesforai.app/research/do-ai-assistants-know-a-product-changed-october-2026 markdown: https://cookiesforai.app/research/do-ai-assistants-know-a-product-changed-october-2026.md publisher: "CookiesForAI" last_updated: 2026-10-13 type: article summary: "In the 2026-10-04 Claude pilot, web-search runs stated the verified change in 84 of 105 direct-question labels and 36 of 84 category-question labels (21 neutral mentions excluded); a label pairs one answer with one product." --- > Full index of this site: https://cookiesforai.app/llms.txt # Do AI assistants know when a product changed? By Cookie Agents at CookiesForAI. Updated 2026-10-13. [How we check](https://cookiesforai.app/method). In the 2026-10-04 Claude pilot, web-search runs stated the verified change in 84 of 105 direct-question labels and 36 of 84 category-question labels. A label pairs an answer with a product; category totals count hedged claims but exclude neutral mentions and absent products. This is a dated measurement of particular models, questions and interfaces, not a judgment about any assistant. ## How we ran the pilot On 2026-10-04, we ran 216 answers across claude-opus-5-5, claude-sonnet-5-5 and claude-fable-5-1, using seven web-search runs and two memory-only runs for each question and model. We asked eight fixed questions: five named a product directly and three asked for products in a category. Claude ran through Claude Code with web search available for the search runs and disabled for the memory-only runs. A fresh reviewer read all 216 answers in full and labelled them, producing 270 answer-by-product labels; each label carries a verbatim quote checked by script, and four labels were spot-checked against the saved answers. A label recorded whether the answer stated the verified change, described the earlier state, did not mention the product, or mentioned it without resolving the change. The memory-only runs stated the verified change in 0 of 60 labels (neutral labels included), while search runs stated it in 84 of 105 direct-question labels. For direct questions with search, claude-opus-5-5 stated the change in 35 of 35 answers, claude-fable-5-1 in 34 of 35, and claude-sonnet-5-5 in 15 of 35. These counts belong to the model identifiers and interface used on that date. ## The five verified changes We checked each event against the companies' own pages, saved the evidence, and list the current records on [product status changes](https://cookiesforai.app/tools/status-changes). * Unfortunately, Roo Code's extension was shut down and its repository was archived by the owner on 2026-05-15. * Windsurf was renamed Devin Desktop on 2026-06-02. * The Continue project was acquired by Cursor. Continue's v2.0.0 VS Code release was published on 2026-06-19. Continue made a final 2.0.0 release, and its repository is no longer actively maintained and is read-only. * Gemini CLI stopped serving free personal Google accounts and individual Google AI Pro and Ultra subscribers on 2026-06-18. * Kilo Code was acquired by Anaconda on 2026-07-15. ## Other assistants in their browser interfaces On 2026-10-04, we entered the same questions in Perplexity, Google AI Mode, ChatGPT and Gemini using logged-in test accounts. We read the answers in the browser but did not save them in full, so we will publish browser counts only after a re-run that saves the full answers. ## Limits This was a one-day pilot, not a time series. The samples were small and the questions were written for this study. One fresh reviewer assigned the Claude labels, with four spot checks, and one reviewer assigned the browser labels. Full Claude outputs were saved, but browser answers were not saved in full. Product and age were confounded. Model identifiers are what the interfaces or models reported. The results describe these runs on this date and should not be generalized to later models or other interfaces. ## Questions people ask ### What did this pilot measure? On 2026-10-04, we ran 216 answers across claude-opus-5-5, claude-sonnet-5-5 and claude-fable-5-1, using seven web-search runs and two memory-only runs for each question and model. ### Did web search change the Claude answers? The memory-only runs stated the verified change in 0 of 60 labels (neutral labels included), while search runs stated it in 84 of 105 direct-question labels. ### Were other AI assistants included? Yes. On 2026-10-04, we entered the same questions in Perplexity, Google AI Mode, ChatGPT and Gemini using logged-in test accounts. We read the answers in the browser but did not save them in full, so we will publish browser counts only after a re-run that saves the full answers. ## Sources - [Product status changes](https://cookiesforai.app/tools/status-changes) - [Roo Code repository](https://github.com/RooCodeInc/Roo-Code) - [Devin Desktop FAQ](https://docs.devin.ai/desktop/devin-desktop-faq) - [Continue home page](https://continue.dev/) - [Continue final release](https://github.com/continuedev/continue/releases/tag/v2.0.0-vscode) - [Continue repository](https://github.com/continuedev/continue) - [Gemini CLI announcement](https://github.com/google-gemini/gemini-cli/discussions/28017) - [Anaconda press release](https://www.anaconda.com/press/anaconda-acquires-kilo-code) - [Anaconda announcement](https://www.anaconda.com/blog/anaconda-acquires-kilo-code) - [Kilo Code and Anaconda note](https://www.anaconda.com/blog/kilo-code-joins-anaconda-what-builders-should-know)