How else you learn that a document an AI lab publishes has changed

The short answer

• Six ways to find out, on the same six criteria, with the page every cell was read off and the day it was read. Three of the six are better than this wire at something, and those rows say so.

• These six are not one market and a single ranking would flatter three of them. The labs’ own changelogs are the primary source and everything else here is derived from them, including this wire, but a changelog only carries what the lab chose to announce, so a price edited on a pricing table or a row dropped from a model card produces no entry. changedetection.io and Visualping catch exactly that, on any page at all, because you choose the pages, and the cost of that is that nothing is watched which nobody added and that what they caught is private to your account. OpenRouter answers what a model costs today, which no diff does well. Artificial Analysis answers how good a model is, which nothing else here attempts. Context Window answers one narrower question than any of them: on eight named labs, what line moved, what it said before, and what it says now, at a URL anybody can open. If your question is what a model costs right now, read OpenRouter. If it is how good the model is, read Artificial Analysis. If the page you care about is not on anybody else’s list, run a monitor. If the question is what changed, this wire is the one that answers it.

• The six criteria below are identical for every entry. This wire is the first row for all six, including the three criteria it loses. Each cell prints the short answer and the page it was read from. The same cell in the project’s own words appears in that project’s section further down.

What it watches

Whether it tells you a document changed

• This criterion is whether the thing reports a CHANGE, with the line as it read before and as it reads now, or only shows the current state and leaves you to remember what it used to say.

Who decides what is watched

• This criterion is whether the list of watched pages is fixed by the publisher or built by you, which decides whether a document nobody thought of is covered or invisible.

What you get when something changes

Whether the record is public and linkable

• This criterion is whether a past change has a URL a stranger can open, as distinct from a notification that arrived in one person’s inbox and is gone.

What it costs

The labs’ own changelogs

• Choose it instead of this wire when. You follow one lab, you want the lab’s own words with nothing in between, and you are willing to open the page yourself. This is the primary source every other row on this grid is derived from, including this one.

• What it watches: One lab’s own releases, and on a separate page its deprecations. OpenAI’s changelog states that it is the latest features and updates to the OpenAI API, and points upcoming deprecations at a separate deprecations page. Anthropic’s release notes state that they list changes to the Claude API, the client SDKs and the Claude Console. Neither covers the other lab, and neither covers a model card, a system card or a pricing table that was edited without a changelog entry.

• Whether it tells you a document changed: It tells you what was ADDED. It does not tell you what was silently edited. A changelog is an announcement the publisher chose to write. A price changed on a pricing table, a row dropped from a model card or an eval removed between two versions of a system card does not produce a changelog entry unless somebody writes one. OpenAI’s deprecations page is the one place this is committed to in writing: it states that impacted customers will always be notified by email and in the documentation, and that generally available models get at least 6 months of notice.

• Who decides what is watched: The lab, entirely. What appears is what the lab decided to write down, on the page the lab decided to put it on. That is the strength of the row and also its limit: nothing is filtered by somebody else’s judgement, and nothing is caught that the lab did not choose to announce.

• What you get when something changes: A dated entry on a documentation page, plus email for a deprecation. Anthropic’s release notes page carries a Subscribe control and entries dated newest first. OpenAI’s deprecations page states that customers actively using a model are notified by email when a deprecation is announced. For everything short of a deprecation, the form is a page you have to go and load.

• Whether the record is public and linkable: Public and free, and OpenAI also serves it as Markdown. OpenAI’s docs state that Markdown versions of documentation pages are available by appending .md to the page URL, and publishes a complete documentation index at /llms.txt, so the changelog is machine readable without scraping. Both labs’ pages are open with no account.

• What it costs: Free. Both pages load with no account and no payment. The cost is your attention: one page per lab per document type, and the eight labs this wire covers publish more than a dozen such pages between them.

• Every line above was read on 2026-09-19 off a page The labs’ own changelogs publishes, listed in full at the foot of this page. Its own site is developers.openai.com.

changedetection.io

• Choose it instead of this wire when. You want to watch a page nobody else watches, on your own schedule, and be interrupted when it moves. It is the right answer whenever the document you care about is not on somebody else’s list.

• What it watches: Any URL you give it, up to 5,000 on the hosted plan. The site states that the subscription includes up to 5,000 URLs, one real Chrome browser for JavaScript sites with more available as a bolt on, and a choice of European, USA and Tor proxies. It also states Trigger Text, CSS and xPath rules and Ignore Text filters, so a watch can be narrowed to one element of a page rather than the whole document.

• Whether it tells you a document changed: Yes. Detecting the change is the whole product. The site describes itself as website change detection and monitoring with alerts, and states re-check times from about 5 minutes per runner. What it does not do is understand the document: it reports that a region of a page differs from the last fetch, and reading that difference as a price move or a deprecation is left to you.

• Who decides what is watched: You, entirely. Every watch is one you created. That is the criterion on which this row beats every wire on this grid, and the cost of it is that a document you never thought to add is not watched, and a page you add badly produces noise on every unrelated edit.

• What you get when something changes: A notification, in more than 85 formats. The site states over 85 different notification types and names Discord, Slack, Telegram, Rocket.Chat, Email, Matrix, NTFY and Office 365 among them. The notification goes to you, so the change exists for whoever was on the alert and for nobody else.

• Whether the record is public and linkable: Private to your account. A watch and its history live in your own instance or your own hosted login. There is no public page for a change it caught, so it cannot be cited, linked or read by anybody you have not given the login to. The project itself is public: the site states that it is a fully opensource project and links its GitHub.

• What it costs: $8.99 a month hosted. The software itself is open source. The site prints Only $8.99 /mo for the hosted subscription, states yearly billing gets 2 months free, and states that the subscription is the official provider of the changedetection.io opensource project. Running the open source build on your own machine is the free route and the site links the GitHub for it.

• Every line above was read on 2026-09-19 off a page changedetection.io publishes, listed in full at the foot of this page. Its own site is changedetection.io.

Visualping

• Choose it instead of this wire when. You want page monitoring your team can run without hosting anything, with the change summarised for you and routed into Slack or Teams. Its own pages sell regulatory and compliance monitoring, which is the same shape of job as watching a lab’s pricing table.

• What it watches: Any page you give it, in blocks of 10 to 1,000. The pricing page builds its plans from two controls: how many pages you want to monitor, offered as 10, 50, 100, 200, 500 and 1,000, and how often to check, offered from every 2 minutes to every month. Its own use case list names competitor monitoring, regulatory intelligence and compliance monitoring.

• Whether it tells you a document changed: Yes, and it offers an AI reading of the difference. The pricing page lists AI analysis as a business feature, described on that page as better understanding the change detected, and lists advanced reports for reviewing hundreds of side by side changes at a glance. As with any generic monitor, the comparison is of the rendered page rather than of the document’s meaning.

• Who decides what is watched: You, entirely. Every job is one somebody on your team created, and the pricing page sells bulk import for creating thousands of them at once. The same trade as the row above: nothing is watched that nobody added.

• What you get when something changes: An alert, routed to Slack, Teams, Google Sheets and email. The pricing page lists advanced notifications as a business feature and names Slack, MS Teams and Google Sheets, and separately lists a Google Sheets integration that logs each alert as a row. It also states one year data retention on business plans against the last 90 days otherwise.

• Whether the record is public and linkable: Private by default, with an optional public alert page. The product list on the pricing page names Public pages, described there as managing your Visualping alert page, so a public record is available and is not the default. What a job caught otherwise stays inside the account.

• What it costs: Not stated on the page read. The pricing page computes its plan prices in the browser. Fetched on 2026-09-19 it served the plan configurator, the words Your total and a per month unit with no figure against it, the line "18,000 checks a month at every hour across 25 pages", and the claim that unused months are refunded. It also states that there is a free tier, in the words Get started it’s free. No price is asserted here because none was on the page that was read.

• Every line above was read on 2026-09-19 off a page Visualping publishes, listed in full at the foot of this page. Its own site is visualping.io.

OpenRouter

• Choose it instead of this wire when. You want today’s price and context length for a model, across every provider serving it, on one page. For the question "what does this model cost right now" it is a better answer than any wire, including this one.

• What it watches: Models across many providers, with price and context on each. The models page titles itself Compare AI Models: Pricing, Context and Benchmarks, and each row carries a dated entry, a context length and an input and output price per million tokens. Read on 2026-09-19 it carried entries dated as recently as 18 September 2026. It lists models, not the documents a lab publishes about them.

• Whether it tells you a document changed: No. It shows the current state. The page is a live catalogue. It tells you what a model costs today, and it does not tell you what it cost last week or that the figure moved, so a price cut or a rise is visible only to somebody who wrote the old number down. That is the exact gap a diff fills, and it is the reason both rows can be on this grid without either being redundant.

• Who decides what is watched: OpenRouter, from what its providers serve. The population is the models routed through the platform. A model nobody serves through it is not on the page, and a document a lab publishes is never on the page at all, because the unit here is a model rather than a page.

• What you get when something changes: A page you load. There is no change notification. Nothing arrives. You get the catalogue when you open it, which is the correct shape for a lookup and the wrong shape for a question about what moved while you were not looking.

• Whether the record is public and linkable: Public and free to read, and every model has its own page. The catalogue and each model’s page are open with no account. What is not public is the history: the page carries the price as it stands, so an earlier figure has no URL on the site.

• What it costs: Free to read the catalogue. The models page loads with no account. Paying begins when you route inference through the platform, which is a different product from reading the prices and is not what this row is about.

• Every line above was read on 2026-09-19 off a page OpenRouter publishes, listed in full at the foot of this page. Its own site is openrouter.ai.

Artificial Analysis

• Choose it instead of this wire when. The question is how good a model is rather than what a lab said about it. It runs its own evaluations, which nothing else on this grid does, and a wire reporting a lab’s own claims is no substitute for an independent measurement.

• What it watches: Models, measured by its own evaluations. The site describes itself as independent analysis of AI and states that its Intelligence Index v4.3 incorporates 10 evaluations, naming AA-Briefcase, GDPval-AA v2, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity’s Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR v1.1. Read on 2026-09-19 the index page showed 27 of 653 models selected. It measures models; it does not read the labs’ documents.

• Whether it tells you a document changed: It publishes a changelog of its own work, not of the labs’ pages. The front page carries a Changelog listing new model evaluations, new articles and methodology updates with dates, and on the day this was read its most recent entries were a language model evaluation on 18 September 2026 and a methodology update on 7 September 2026. What changed is its measurement, not a lab’s document.

• Who decides what is watched: Artificial Analysis, and its population is models it has evaluated. Coverage is whatever it has run. Its own front page states 653 models in the index selector, which is a far larger population than any document watcher on this grid holds, and a different population: a model with no published model card can still be measured, and a document with no model behind it is out of scope.

• What you get when something changes: Charts and leaderboards on the site, plus published articles. The front page carries the indices, leaderboards for speed, cost and output tokens, and dated articles announcing each methodology version. Nothing is pushed at a per document granularity, because the document is not the unit.

• Whether the record is public and linkable: Public and free to read, with the methodology published beside it. The index, the leaderboards and the methodology pages are open. The front page states that the Intelligence Index methodology gives a breakdown of each evaluation and how they are run, which is the same contract a comparison page is held to and is the reason this row is worth having on the grid.

• What it costs: Free to read. The site loads with no account. The pages that were read on 2026-09-19 carried no price against any of the indices or leaderboards, and no paid tier was named on the front page.

• Every line above was read on 2026-09-19 off a page Artificial Analysis publishes, listed in full at the foot of this page. Its own site is artificialanalysis.ai.

Context Window on its own grid

• What it watches: A fixed list of documents from eight named labs. The about page states that the wire reads the model cards, system cards, pricing pages, API changelogs, docs, deprecation notices, research posts and status pages published by OpenAI, Anthropic, Google DeepMind, Meta, Mistral, xAI, Qwen and DeepSeek. Eight labs is the whole beat. A document published by a lab that is not one of those eight is not watched, and there is no way to ask for one.

• Whether it tells you a document changed: A change is the unit. Nothing here reports the current state. The about page states that the engine fetches each registered document on a schedule, keeps the last normalised snapshot of it, and diffs the new fetch against that snapshot, and that a change which passes the gate becomes a story carrying the lab, the document, the line that changed, what it read before and what it reads now. A day on which nothing moved produces nothing.

• Who decides what is watched: The publisher, and a reader cannot add a document. The watched set is declared in the engine, not configured by a reader. There is no account on this site, so there is also no per reader list, and a lab or a document nobody registered is absent with no way to correct that from outside. This is the criterion on which an ordinary page monitor is simply better.

• What you get when something changes: A page on the wire and an RSS item. No email, no webhook, no alert. Each change is filed as a story at its own URL and appears in the RSS feed at /feed.xml, which answered HTTP 200 on the day this was read. Nothing on this site sends anything to anybody: there is no subscribe form, no email, no Slack integration and no webhook, so a reader who wants to be interrupted has to point their own reader at the feed.

• Whether the record is public and linkable: Public, free, and every change has its own URL. Every change is a page at /changes/<slug> with the quoted line on it, all of them are in the ledger at /search, and none of it is behind an account. The about page states that every figure in a post is a figure on somebody else’s page, at a URL, quoted character for character, and that the document a post reports on is one link away.

• What it costs: Free, with no account. The about page states that Context Window is a free, automated wire published by Compound Labs. There is no sign in on the site, no paid tier and no rate limit on reading it.

Questions

• Do the AI labs publish an RSS feed of their changelogs?

  Not uniformly, which is the practical reason a watcher exists at all. Read on 2026-09-19, Anthropic’s Claude Platform release notes page carries a Subscribe control, and OpenAI’s API changelog is served as Markdown by appending .md to the page URL, with a complete documentation index at /llms.txt. Those are two different mechanisms on two labs, and the wire covers eight.

• Why not just set up a page monitor on the pricing pages myself?

  That is a good answer and this page says so. changedetection.io starts at $8.99 a month hosted and is open source if you run it yourself, and it will watch any URL you give it with CSS or xPath rules narrowing the watch to one element. What you take on is the list: every lab, every document, every time one of them moves a page to a new URL. What you do not get is a public record, because the change arrives in your notification and has no page a colleague can open.

• What does a diff catch that a changelog does not?

  An edit nobody announced. A changelog entry exists because somebody wrote it. A price changed on a pricing table, a row removed from a model card, an eval dropped between two versions of a system card, a rate limit raised in a docs page: none of those produces a changelog entry unless the publisher chose to write one. A diff is triggered by the page changing rather than by the decision to announce.

• What is Context Window worse at than the rest of this grid?

  Three things, all of them on the grid above. It watches a fixed list of documents from eight labs and a reader cannot add one, where both monitors on this page watch anything you give them. It sends nothing: there is no email, no webhook and no Slack, only a page and an RSS feed. And it reports what a lab said about its own model, never whether the model is any good, which is what Artificial Analysis measures.

• How current are these figures?

  Every cell carries the day it was read, and a job re-reads them rather than a person promising to. Every source on this page was fetched on 2026-09-19 and scripts/verify-alternatives.mjs re-fetches each one and fails when the literal string a cell came from is no longer on its page. Where a page did not carry a figure the cell says so and still names the page, which is why the Visualping cost cell reads as an absence rather than a number.

What was read, and when

• Cells show how many of the six criteria that one page supplied. The Read off column of every table above names the page for each criterion.

• This is the only page on this wire that states anything about another organisation in its own voice. It applies the wire’s contract to that organisation: it records the page a figure came from, the day it was read, and a string a job can search for. Prices, plans and inclusion rules change, so each source records its own read date instead of using one date at the bottom of the page. scripts/verify-alternatives.mjs re-fetches each of these URLs and fails when the literal string from a cell is no longer on its page.