# An AI release tracker — new models and features gathered on a schedule, checked against the source Recipe No. 35, Work and teams. From The know.sh Cookbook: https://know.sh/cookbook/ai-release-tracker - For: an engineering lead who has to know which new models and features are worth trying this month - You bring: the vendors and tools your team uses, the questions you care about (price, context length, tool calling, licence), and an assistant that can search the web - You get: a library with a finding for every release that matters to you, each linked to its official announcement, a monthly digest your team reads by link, and a short list of what to try - Time: an hour to set up, then ten minutes after each weekly run - Keep it: private while you work; share one document by a read-only link Every week brings new models, new context lengths, new prices and quiet deprecations, announced in blog posts, changelogs and release notes scattered across a dozen vendors. Most teams keep up by accident: someone sees a post, someone else half-remembers a price, and the model your product depends on is retired before anyone reads the notice. This recipe turns the flood into a library that updates itself. Once a week a scheduled task in your assistant — Claude and ChatGPT both run them — searches the vendors' own release notes for anything new against your watch list. Your assistant files each release as a finding: what shipped, when, a link to the official announcement, and what it means for the questions you track. You read the week in ten minutes, correct what it got wrong in the editor, mark the few that matter, and the month's digest writes itself from the findings. know.sh does not watch the web itself; your assistant does, on its schedule. The official pages remain the source; the findings are your team's notes on them. ## What you will use - **Shelf**: One shelf, *AI releases*, with a line naming the vendors and tools you follow. - **Research document**: A standing *Watch list* document for the questions you track, and one document per month, *AI releases — October 2026*, whose overview becomes that month’s digest. - **Finding**: One finding per release, filed as *Evidence*, with the official announcement as its source; your team’s verdicts filed as *Decision*; unconfirmed reports filed as *Question*. - **Scheduled runs**: A weekly scheduled task in Claude or ChatGPT that searches official release notes and changelogs and prepares the week’s entries. - **Your AI assistant**: Files the week’s releases into the month’s document, checks each against the watch list, flags duplicates and drafts the digest. - **The editor**: Where you correct dates and prices, merge duplicates, mark the releases that matter as **Key**, and write the verdict yourself. - **The A–Z index**: Gathers the vendor and product names that recur across months, so every release from one vendor is a single line. - **Public link**: A read-only link to each month’s document, posted in the team channel. ## Method ### 1. Write down what you are watching for Before any automation, decide what counts. Make a *Watch list* document with a finding for each thing you track: the vendors whose models you use or might use; the capabilities that change your plans (tool calling, context length, structured output, vision); the numbers that matter (price per million tokens, rate limits); licences for open-weight models; and deprecation notices for anything in production. This document is what keeps the weekly run from filling the library with noise. Your assistant reads it every time, and you can change it whenever your needs change. ### 2. Let your assistant build the shelf Connect the assistant you already use to know.sh, then give it your notes and let it do the setup: "Using know.sh, create a shelf called *AI releases*, a *Watch list* document from the notes below with one finding per item, and a document called *AI releases — October 2026*." Open the result and fix it in the editor. The watch list is short, so read every finding: the wording you settle on here is what the weekly run will match against. ### 3. Schedule the weekly run Create a scheduled task in your assistant. In Claude, scheduled tasks are on the Pro, Max, Team and Enterprise plans and run in the cloud even while your computer is asleep; in ChatGPT, scheduled tasks are on every plan and deliver their results by notification or email. Describe it in plain words: "Every Monday at 8 a.m., search the official release notes, changelogs and blogs of the vendors on my watch list for anything released in the past seven days." When the result arrives, reply *file this in know.sh*, and your assistant adds each release to the month's document. Some setups can file straight into know.sh during the scheduled run itself; that is newer and not every plan or connector supports it yet, so check **Revisions** after each run either way. ### 4. File each release against its official source Each release becomes one finding, titled in the same pattern so the month reads in order: *Released by Northwind, 29 September: Lumen 2 Pro*. The body says what shipped, what it costs, how it answers the watch list, and anything the announcement leaves vague. The source is the vendor's own page. Coverage from newsletters and social media is a lead, not a source. If your assistant finds a release only in second-hand reports, it files it as a *Question* ("Reported: Northwind to retire Lumen 1 in January") until an official page confirms it. ### 5. Read the week and fix it up Ten minutes each Monday: open the month's document and read the new findings. Merge the two findings that describe the same launch, correct a price your assistant misread, delete what turned out to be irrelevant, and mark the one or two releases your team should try as **Key**. Write the verdict yourself, as a *Decision*: "Try Lumen 2 Pro for the summarisation service; it is cheaper at our volume and supports tool calling." Highlight the lines that affect production, such as a deprecation date. Highlighted findings show their numbers in bold wherever they appear in the index. ### 6. Let the digest write itself, then share it At the end of the month, ask your assistant to draft the document's overview as a digest: three sections, *Worth trying*, *Worth knowing* and *Deprecations*, each line citing its finding number. Edit it in your own words, then press **Share** and post the read-only link in the team channel. Each month is its own document, so each month has its own link; the team channel becomes the archive of links, and the shelf keeps the whole history in one place. ### 7. Ask the library before you plan Before a planning meeting, ask your assistant a question the library can answer: "Using know.sh, which releases on the AI releases shelf this quarter support tool calling with a context of at least 200,000 tokens, and what do they cost? Cite finding numbers." Because every finding links to its official source, the answer is something you can check in minutes rather than take on trust. Look up (⌘K) finds a model name or a word such as *deprecation* across every month at once. ## Prompts to try Your assistant, connected to know.sh (Claude, ChatGPT or a local model): > Every Monday at 8 a.m., search the official release notes, changelogs and announcement blogs of these vendors for anything released in the past seven days: [your vendors]. For each release give the date, what shipped, the price if stated, and whether it supports tool calling, long context or structured output. Link only the vendor’s own pages; list anything you saw only second-hand separately, as unconfirmed. Your assistant, connected to know.sh (Claude, ChatGPT or a local model): > Using know.sh, read my Watch list document and every finding in AI releases — October 2026 on the AI releases shelf, and draft the month’s digest as that document’s overview: three sections, Worth trying, Worth knowing and Deprecations, each line citing its finding number. Do not change the findings. Your assistant, connected to know.sh (Claude, ChatGPT or a local model): > Using know.sh, list every finding on the AI releases shelf from this quarter whose source is not the vendor’s own website, and put the list in a new finding called Sources to replace, with a link to the finding each one comes from. ## Variations - Tracking one product rather than a market? Point the weekly run at a single vendor’s API changelog, and file each change your integration depends on. - In a launch week, switch the task to daily; paid plans in both Claude and ChatGPT allow it. - Use the same rhythm for research: a weekly run that finds new papers in your field and files them as leads, feeding a [literature review](/cookbook/literature-review). - Keep a competitor watch the same way, with their public release notes as sources, and feed it into your [battlecards](/cookbook/battlecards). ## Where it falls short - know.sh does not watch the web; your assistant does, on its schedule. A failed or skipped run means a missed week, so glance at your assistant’s list of scheduled runs now and then. - Filing straight into know.sh during an unattended run is new and not supported on every plan or connector yet. The dependable pattern is to reply *file this* when the result arrives. - Assistants misread prices, limits and dates in announcements. The official page is the source; the finding is only your note about it. - Scheduled tasks count against your assistant plan’s limits; ChatGPT’s free plan, for example, runs tasks at most once a day. ## A note on sources and claims Treat every vendor claim as a claim: benchmark scores, speed and "state of the art" are the vendor's words until someone you trust has tested them. Say so in the finding. Read the licence before anyone uses an open-weight model; licences differ on commercial use, and a summary is no substitute for the text. Keep internal evaluation results, customer data and anything under NDA out of the monthly document you share. Indexed under: AI releases, Scheduled tasks, Release notes, Changelogs, Deprecations, Model licences, Watch lists.