Seven dated changes landed this week. One of them is on a deprecation page.

Between Friday morning and yesterday afternoon, five vendors published seven changes to things builders depend on, each carrying a date. Not one of them was a launch. A weekly usage limit falls. A review default flips. A retention window goes from 28 days to indefinite. A payment term moves from arrears to upfront. A preview endpoint retires. A marketplace closes permanently. A supply contract ends.

Anthropic, Google and OpenAI all maintain a public deprecation page for exactly this purpose. Between them, those three pages contain one of the seven.

That gap is the useful thing here, and it survives the specific week that produced it.

The seven, with the place each one was published

Change	Effective	Published where
Claude Code weekly limits: the 50% boost is replaced by a permanent 25% baseline	September 14	@ClaudeDevs on X, August 29
Copilot code review default effort: Lite becomes Balanced	September 28	GitHub changelog, August 28
Copilot chat retention: 28 days becomes the life of the account	No earlier than September 28	Same changelog entry
Copilot Business and Enterprise seats charged upfront (card and PayPal customers)	October 1	Same changelog entry
gemini-omni-flash-preview deprecated	September 30	Gemini API changelog and deprecations page
Amazon Mechanical Turk closes permanently	September 30	A banner on mturk.com
OpenAI models withdrawn from Cursor (proposed)	November 12	openai.com post, August 28

The Claude Code line is the one worth reading twice, because the headline and the arithmetic point in opposite directions. Anthropic's developer account posted on August 29: "Starting September 14, we're permanently raising standard weekly limits in Claude Code by 25% for Pro, Max, Team, and seat-based Enterprise plans. Until then, the current 50% increase will be in place." A permanent raise, and true. But the 50% boost has been running since May 13 and was extended through the summer, so the baseline everyone has actually been working against for three and a half months is the boosted one. Anthropic said so itself in the follow-up post: "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code."

That number is the vendor's, not an inference. It is also not on anthropic.com/news, not in the Claude Code cost documentation, which describes the weekly window in detail without mentioning either the boost or September 14, and not on Anthropic's model deprecations page. It exists on X and in the coverage that picked it up.

What the three deprecation pages actually cover

Read the scope language and the gap stops being an oversight.

Anthropic's page opens with "This page lists all API deprecations" and narrows further: "The dates on this page apply to Anthropic-operated platforms: the Claude API, Claude Platform on AWS, and Microsoft Foundry." Partner platforms set their own schedules. The commitment attached is "at least 60 days' notice before model retirement for publicly released models." Retirement of models. The table's dates are phrased "Not sooner than."

Google's Gemini page is the most complete of the three: a table of model IDs, release dates, shutdown dates and named replacements. It also carries the most honest caveat any of them publishes. The shutdown dates "indicate the earliest possible dates on which a model might be retired. We will communicate the exact shutdown date to users with advance notice."

OpenAI's page reaches furthest past models into products. The Assistants API shut down August 26. The Evals platform, Agent Builder and reusable prompt objects shut down November 30. The Videos API shuts down September 24. Its most recent entry is dated July 20. What it does not carry, and was never designed to carry, is which third-party products may resell those models, which is the entire content of the Cursor decision published on Friday.

So the pattern is not vendor negligence. All three pages are registries of one object type: the model identifier, plus in OpenAI's case the first-party API surface around it. Exactly one of this week's seven changes is that kind of object, and it is the one that behaves well. gemini-omni-flash-preview has a date, a page, and a general availability replacement in gemini-omni-1.1-flash that shipped on August 27, a month before the retirement.

The other six are a subscription rate limit, two defaults inside a product, a billing term, a marketplace, and a supply agreement. None of those is versioned. Anything a vendor does not version, it changes by announcement.

Who has to do something, and by when

• You run Claude Code on Pro, Max, Team or seat-based Enterprise and your week already runs close to the cap. September 14 removes about a sixth of the capacity you have had all summer. Before then, run /usage and press w for the seven-day view, on a normal week rather than a quiet one. That gives you a pre-cut number to compare against, which you cannot reconstruct afterwards. If the answer is that you were regularly finishing the week near the ceiling, the choices are usage credits, moving routine work to a cheaper model, or accepting a shorter week.
• Your team uses Copilot code review with the effort level left at Default. On September 28 that value starts resolving to Balanced instead of Lite. Set it explicitly before the date, in either direction. The point is not that Balanced is worse; it is that a default that changes underneath you converts a decision you never made into a decision you now own.
• Your data policy states a 28-day retention window for Copilot chat. It becomes the life of the account. That sentence in a security review, a customer DPA or an internal policy document becomes wrong on a date, with no code change to trigger a review.
• You pay for Copilot Business or Enterprise by card or PayPal. From October 1, assigned seats are charged upfront at the start of the next billing cycle. Automation that assigns seats to contractors and reclaims them later now moves cash earlier than it used to.
• You run HIT-based annotation, human evaluation or red-team scoring on Mechanical Turk. Thirty-one days. Export HIT history and qualification lists yourself; a closing platform's FAQ is not a retention guarantee.
• GPT-5.6 appears anywhere in a Cursor workflow, admin model list or automation. November 12 is the proposed date, and the three routes out each carry different cost and data terms.

The column most registers leave out

A dependency list is keyed by what you use. An expiry register has to be keyed by what could stop, and that is a wider set which includes things that are not software: a promotional rate, a payment term, a retention default, a reseller agreement, a marketplace.

Two weeks ago the argument here was narrower, that a cost row needs a next-known-change-date column because published rates now carry expiry dates of their own. This week says the same column belongs on rows that have nothing to do with price. A rate limit is not a cost row. Neither is a retention policy or a reseller agreement, and all three moved on a schedule inside five days.

Four columns are enough. The thing. The date. The URL where the date was published. And what you do if the date arrives unchanged.

The fourth column is the one that gets skipped and the one that does the work. A date with no prepared response is a worry, not a plan, and the response is cheapest to write on the day you learn the date, while you still have the context that made you care. "September 14: re-measure weekly usage against the new ceiling, and decide on usage credits" takes twenty seconds to write in August and half a day to reconstruct in September.

Keep the source URL in the row for the same reason. Several of this week's dates are qualified in ways a one-line summary destroys: Google's shutdown dates are earliest-possible, Anthropic's retirement dates say "not sooner than," GitHub's unified relaunch says "no earlier than September 28th," and OpenAI's November 12 is described as proposed. A register that records those as hard dates will be wrong in the direction of panic, and one that records them without the qualifier will be unfalsifiable later.

Where the dates live, per vendor

• Google: ai.google.dev/gemini-api/docs/deprecations. A real table with replacements. Read the earliest-possible caveat before treating a row as a deadline.
• OpenAI: developers.openai.com/api/docs/deprecations. Models and first-party products. Nothing about third-party availability.
• Anthropic: platform.claude.com/docs/en/about-claude/model-deprecations. API models only, 60 days' minimum notice, tentative dates.
• GitHub: no deprecation page. The changelog at github.blog/changelog filtered on the copilot tag, which is where all three of this week's Copilot dates were published in a single entry.
• Plan terms, promotional capacity, retention defaults and supply contracts: nowhere durable. They arrive once, on a blog post or a social account, and are not restated.

That last line is the part that determines whether any of this is actionable. You cannot build the register by subscribing to something, because for most of what expires there is nothing to subscribe to. It gets built by writing the date down the first time you encounter it, from whatever channel you encountered it in, on the assumption that this is the only time anyone will say it out loud.

Six of the seven dates fall inside the next five weeks. If your register is currently empty, September 14 is the entry to write first, because it is the only one of the seven where the change reduces something you are already using every day, and the vendor has already published the size of the reduction.