Digest — 2026-09-04 (technologist)
Three things crossed the escalation bar. One is a security fix in something the fleet runs. Two are terms changes. No dated deadline falls inside 30 days — see the bottom.
1. ACT ON THIS — the Claude Code security floor is now 2.1.260, not 2.1.251
Install >= 2.1.260. Four permission-boundary or credential defects were fixed in the four releases after the floor this corpus currently records.
Read from CHANGELOG.md on 2026-09-04, verbatim:
2.1.260
- "Fixed
Edit/Write/Readpermission rules whose path contains parentheses being dropped" - "Fixed Bash permission checks auto-approving zsh commands that hide a command substitution"
- "Fixed
permissions.blockReadsOutsideWorkingDirectorieson macOS hiding the user's git config"
2.1.259
- "Fixed Bash
Read()deny rules not covering files given as option values" - "Fixed managed settings silently going unenforced when the managed-settings file cannot be parsed"
2.1.257
- "Added a Containment Escape rule to auto mode so cloud metadata-credential fetches are no longer auto-approved"
- "Fixed a leftover Anthropic API key or auth token being sent alongside your Foundry subscription key"
Plain-language read of the two that matter most. A permission rule is the line that says which files or commands Claude Code may touch without asking. The 2.1.260 zsh fix means that before this release, a shell command could smuggle a second command inside it (command substitution — one command's output pasted into another) and the permission check would wave the whole thing through. The parentheses fix means any rule you wrote with a ( in the path was being silently discarded rather than enforced. Both are "the guard rail was not where you thought it was" bugs.
The honest caveat, and it matters
The 2.1.259 block contains both of these lines:
"Fixed Bash
Read()deny rules not covering files given as option values" "Reverted the 2.1.259 change applyingRead()deny rules to Bash arguments"
A fix and its own revert, inside the same version heading. This corpus has already recorded that Anthropic edits CHANGELOG.md entries in place after publication (technology--claude-code-security-floor, method warning added 2026-08-21), so the most likely reading is that the fix shipped, regressed something, and was pulled — with the heading never split. What that means for you: do not treat the Bash Read() deny-rule gap as closed. The other three fixes above are unambiguous; this one is not, and I could not resolve it from the changelog alone.
Version accounting
registry.npmjs.org/@anthropic-ai/claude-code/latest reports 2.1.260. The CHANGELOG's top heading is also 2.1.260. They agree — which closes a row that has been open since 2026-09-01, when npm reported 2.1.252 and the CHANGELOG topped out at 2.1.251. The benign publish-ordering explanation held.
Versions 2.1.253, 2.1.254, 2.1.255, 2.1.256, 2.1.249 and 2.1.244 have no CHANGELOG entry at all. Stated as observed. Whether they were published and undocumented, or never published, is not established from the changelog.
Nine releases in the four days 2026-09-01 -> 2026-09-04, on top of the eighteen moves in twenty-seven days already recorded. The release cadence has not cooled.
2. TERMS CHANGE — the top Anthropic tier now requires 30-day data retention
Read from the Fable 5 and Fable 5.1 migration guides, verbatim:
"Both models require 30-day data retention and are not available under zero data retention (ZDR) arrangements unless expressly authorized by Anthropic. Both are designated Covered Models."
"On the Claude API, a request to Claude Fable 5 from an organization whose data retention configuration does not meet this requirement returns a 400
invalid_request_error."
Why this reaches you even though you don't run Fable. Zero data retention means the vendor keeps no copy of what you send after the request completes. Your fleet sends this model Gmail contents, financial-account context via Era, and your whole Drive corpus. The tier that carries the strongest capability claims is the tier where retaining that for 30 days is mandatory and not negotiable at the API level. That is a direction, not an event — but it is the kind that reaches a subscription user eventually.
Nothing in your fleet is affected today. All four scheduled tasks run claude-opus-5, re-verified live this run. Opus 5 carries no such requirement on the pages read.
Also newly true and worth one line: Claude Fable 5.1 and Claude Mythos 5.1 shipped (Fable 5.1 is now Claude Code's default Fable model, per 2.1.257). Neither is supported on Priority Tier; Fable 5 is.
3. PRICE CHANGE — cache reads on the 5.1 models are a quarter of the old rate
From the pricing page, verbatim:
"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."
This breaks a claim this corpus currently states as verified. fleet--prompt-cache-economics records that the 0.1x cache-hit ratio "holds at all five tiers." It no longer does. Two tiers are now 0.025x.
In dollars rather than multipliers: a Fable 5.1 cache read is $0.25 per million tokens. An Opus 5 cache read is $0.50. The premium tier is now cheaper per cached token than the tier you actually run, while being twice as expensive on fresh input ($10 vs $5) and on output ($50 vs $25).
Do not act on this. Your runs are three days apart and no cache survives that, so your workload is close to all-fresh-input; on fresh input and output Fable 5.1 is straight-up 2x Opus 5. It is recorded because a corpus claim is now wrong, not because there is a switch to make.
Nothing has a deadline inside 30 days
The model deprecations page was read this run. The most recent deprecation notice of any model is still 2026-06-05 — now 91 days of silence. On Anthropic's verbatim policy of "at least 60 days' notice before model retirement," nothing on that page can retire before 2026-11-03 even if a notice issued today.
That arithmetically rules out, again:
claude-sonnet-4-5-20250929— tentative "Not sooner than September 29, 2026"claude-haiku-4-5-20251001— tentative "Not sooner than October 15, 2026"
Neither is reachable. The Haiku watch stays relaxed.
What this digest does not cover
The full ledger walk, the landscape findings, and the corrections to technology--tokenizer-token-count-shift and fleet--prompt-cache-economics are in ledger/promises-addendum-2026-09-04.md and the run log. This file carries only what crossed the bar.