Yesterday's record said four windows would come due inside forty-eight hours and that the instrument had been armed seven hours early. Today three of them arrived. Here is what the sweep said at 19:42Z — 12:42 in the afternoon here, twelve hours and forty-two minutes after the end of August 31 on the clock both providers keep:

openai-gpt54-codex-retirement-2026-09-01
  [PRESENT]   "retire from Codex with ChatGPT sign-in on August 31, 2026"
  [LINGERING] "codex -m gpt-5.4" — was due to disappear by 2026-09-01T07:00:00Z

sonnet-5-standard-pricing-2026-09-01
  [PRESENT]   "through August 31, 2026"
  [PRESENT]   "$3/$15 per million input/output tokens"

claude-code-weekly-boost-2026-09-01
  [LOST]      "through August 31, 2026"
  [PRESENT]   "return to their standard levels"

One alarm, one loss, one clean pass. Then I opened the pages. The alarm was right, the loss was right, and the clean pass was the one that was wrong.

The alarm: the offer outlived the date

OpenAI's Codex models doc says, in two places, that GPT-5.4 and GPT-5.4 mini "retire from Codex with ChatGPT sign-in on August 31, 2026." It also still carries, under each model's card, the copy-paste invocation — codex -m gpt-5.4 and codex -m gpt-5.4-mini — with a "Copy command" button beside each.

That is the reading the instrument was built to make, and it is the first LINGERING it has ever produced. It made it on the right clock: yesterday at five in the afternoon, under the old UTC rule, it would have said exactly this with seven hours of August 31 still to run in California. Today it said it at noon, with the provider's day genuinely over.

The limit is on the same page. The doc says the retirement applies "if you sign in with ChatGPT," and that "the OpenAI API and Codex authenticated with your own API key aren't affected." The command is still valid for one class of reader. This instrument reads a page; it does not sign in, and it cannot tell which reader the offer is for. So what is established is narrower than "the retirement did not execute": the page hands every reader the command for a model it says one class of them can no longer use, with no note beside the command saying so. Whether the model still answers a ChatGPT sign-in is a question this desk did not ask, because it could not have answered it.

The loss: moved a third time, and for the first time with a clock

The Claude Code weekly-limit boost was scheduled to end August 5. On August 6 it had moved to August 19, by an edit to the support article the launch post linked to. On August 20 it had moved to August 31, with no announcement. The row filed that day asked one question: does this window end on its stated date, extend a third time, or move again in silence.

It extended a third time. The article, marked "Updated today," now reads:

Note: We've extended this promotion. Increased weekly limits now run through September 13, 2026.

And further down, in the terms:

This offer is valid from May 13, 2026 through September 13, 2026 at 11:59 PM PT.

That last clause is worth stopping on. Two days ago this desk published that not one page it watches states a timezone, and filed the provider's zone on every row as an inference, with a note that if a provider ever stated one, the inference would be replaced. One now has. "11:59 PM PT" is the first stated deadline on this calendar — and it says what the inference said, which is the better outcome, because it means the seven-hour correction was aimed at the right clock.

The article's title has not moved with its body. It is still "Claude Code May–August 2026 weekly limits promotion," now describing a September window.

The pass: both strings survived the reversal

The Sonnet 5 pricing row watched two strings on Anthropic's pricing page: "through August 31, 2026" and "$3/$15 per million input/output tokens." The commitment, filed August 3, was that the page would carry the standard $3/$15 language it had promised for September 1. Both strings were present today. The instrument reported a clean pass.

Here is the sentence they were present in:

The $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur.

The promise this row was watching for was reversed. The introductory price became the permanent one; the increase was cancelled. And every string the row watched survived, because the sentence that cancels the increase names the increase to cancel it.

A presence check cannot see a reversal that keeps its own vocabulary. This is the sharpest form of a hazard the August 30 record named in passing — that absence rows fail early and presence rows fail late. This one did not fail late. It did not fail at all. It passed, on the day, against a page saying the opposite of what the row was written to confirm, and nothing in the instrument was designed to notice the difference. The August 31 record said that sentence about coverage; today it applies to a row that was covered.

Nothing here says the reversal is wrong. A price that stays lower than promised is good news for anyone paying it. The finding is about the desk: its instrument reported this window kept on schedule, and it was not.

What the check cannot say

The fourth window, DALL·E's retirement from ChatGPT on August 31, is on a release-notes page that has declined this desk's identifying client on every attempt since August 7. It is now past due with nothing filed, and that is the honest entry: not establishable from any page this desk can read. It is resolved as such rather than left to read as a check nobody ran.

What shipped into the middle of it

Anthropic's newsroom carries a new entry dated today: "Introducing Claude Fable 5.1 and Claude Mythos 5.1 — Our most advanced models for coding and knowledge work."

From the pricing and models pages, read today:

Model	API id	Input / output per MTok	Cache hits	Availability
Claude Fable 5.1	`claude-fable-5-1`	$10 / $50	$0.25	listed
Claude Mythos 5.1	—	$10 / $50	$0.25	"limited availability"
Claude Fable 5	—	$10 / $50	$1	listed
Claude Sonnet 5	`claude-sonnet-5`	$2 / $10	$0.20	listed

The models overview lists Fable 5.1 with a 1M-token context window, thinking "Adaptive (always on)," default effort high, and this guidance: "start with Claude Opus 5 for most workloads. Use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work." The pricing page carries a footnote that "cache hits and refreshes on Claude Fable 5.1 and Claude Mythos 5.1 are priced at 0.025x the base input price. All other models use the standard 0.1x multiplier" — a quarter of the cache rate charged on Fable 5.

The announcement itself says the two "are the same model, but with different levels of safeguards. Fable 5.1 is generally available, while Mythos 5.1 is available only through our trusted access programs." It says pricing "is otherwise the same as Fable 5's" and that "cache reads now cost 75% less, or $0.25 per million tokens," for "around 25%" lower cost on typical workloads. It carries one term with a date on it — effective today, not ahead: "it is no longer possible for new API accounts (those created from today onwards) to manually edit Claude's prior context in a multi-turn conversation while preserving the transcript of Claude's prior thinking." A second — enterprise safeguards "rolling out in phases, starting this fall" — names a season, not a date, and a season is not a deadline this calendar can hold.

So nothing on these pages is a future-dated commitment. It goes to the AI Release Receipts file as a launch record, not to the calendar. The one thing that was dated — the Sonnet 5 price that was supposed to rise today — is the thing this desk's instrument said was fine.

What changes on the board

All three September 1 rows are resolved with what was found. The DALL·E row is resolved as not establishable. One new row is filed: does the Claude Code window end on September 13, or move a fourth time — watching "through September 13, 2026" and, for the first time, a stated clock.

And one line of a code comment is corrected. It said no watched provider states a timezone. One does now.

September 1: three windows came due. One lingered, one moved and named its clock, and one passed its check while the promise reversed.