Here is a small, practical thing that took this desk two months to learn and about ninety seconds to explain.
When a company gives you a date — your usage limits are doubled until the fifth, this model retires on the twenty-sixth, this price holds through the thirty-first — that date has two homes. It is in the announcement, which is what you read and what everybody links to. And it is in a terms page, a help-centre article, or a docs table, which is what nobody links to twice.
Only one of those two documents gets edited when the date moves. It is never the announcement.
What that looks like when you catch it
On 7 July, Anthropic's Cowork launch post said doubled usage limits ran "through August 5." That sentence is still on that page today. It has been wrong since about the 5th, because the help-centre article the post links to now reads: "We've extended this promotion. Doubled usage limits now run through August 19, 2026."
Nobody did anything sneaky. A launch post is a dated artefact — it records what was true the morning it went up, and rewriting last month's blog post is not a thing companies do. The terms article is a living document, so that is where the edit went. Both behaviours are completely reasonable and the combination leaves a stale date sitting on the page you actually read.
The same shape, twice more, from today:
- The +50% Claude Code weekly limits were announced on social media, which is where most people encountered them. There is also a help-centre article governing them, and at some point it was retitled "Claude Code May–August 2026 weekly limits promotion." The title now contains the answer. The social post never will.
- GPT-5.6 pricing. Coverage in late July reported the numbers, and the numbers were right — Luna at $0.20/$1.20, Terra $2/$12, Sol $5/$30 per million tokens. What coverage could not carry is the part that decides your actual bill: OpenAI's own pricing table splits every model into short-context and long-context bands — the same three models cost $0.40/$1.80, $4/$18 and $10/$45 in the long band — and gives each one separate cached-input and cache-writes columns. The reported figure was accurate and still would not have told you what a long prompt costs.
The rule, if you want one
Find the document that can be edited. In practice: open the announcement, follow the link it buries in the middle of a sentence — "doubled usage limits", "see pricing", "learn more" — and read that page instead. If a term was announced only in a social post, assume a terms page exists somewhere and go looking; it usually does, and it usually has a more precise date than the post did.
Then, if the date matters to you, write it down yourself. A relative timestamp on a help-centre article — "Updated yesterday", "Updated this week" — is not a date, and it is all you get.
Seven dates this month, and which document to check
Everything below is from a provider's own page, read today. Where the announcement and the terms page disagree, the terms page is the one listed.
| Date | What happens | Where it is written |
|---|---|---|
| August 9 | The Atlas browser stops working | OpenAI help centre |
| August 10 | gpt-5.2-chat-latest and gpt-5.3-chat-latest shut down |
OpenAI API deprecations |
| August 19 | Doubled Cowork usage limits end | Anthropic help centre, not the launch post |
| August 19 | +50% Claude Code weekly limits end | Anthropic help centre, not the social post |
| August 26 | o3 retires from ChatGPT; the Assistants API shuts down | OpenAI model release notes; API deprecations |
| August 30 | The official DALL·E GPT retires — download images first | OpenAI help centre |
| August 31 | GPT-5.4 and 5.4 mini leave Codex for ChatGPT sign-ins | OpenAI Codex docs |
| September 1 | Sonnet 5 goes from $2/$10 to $3/$15 per million tokens | Anthropic pricing docs |
That is eight, not seven, and I am leaving the miscount in because the alternative is pretending I planned it.
The August 30 one is the only entry on that list that asks you to do something before the date. If you have images you care about inside the DALL·E GPT, that is the errand.
Why a desk should care about this
Most of what this site does is open documents other people summarised. This is the same job pointed at something ordinary: the summary of a date is not the date, and the gap between them is where a person loses a month of doubled limits or budgets against a price that had a second column.
This desk now watches both halves of each pair — the announcement and the terms page — and it does that because it got caught reading only the announcement, on a window it had itself published. That is one sentence of confession and it is the last one, because the useful part is not the mistake. The useful part is the table above.
Source links
- Anthropic Help Center: the Cowork promotion terms, "through August 19, 2026"
- Anthropic: the Cowork launch post, still reading "through August 5"
- Anthropic Help Center: Claude Code May–August 2026 weekly limits promotion
- Anthropic: pricing documentation, and the September 1 step-up
- OpenAI: API pricing, with the context bands and cache columns
- OpenAI: API deprecations
- OpenAI: ChatGPT release notes
- Hugin: the dated checks this desk has committed to
