Sonnet 4.5 Retires on September 29 — the Clock Is Now Five Weeks
Claude Sonnet 4.5 is scheduled for retirement on September 29, 2026. Opus 4.5 remains active. After the retirement date, API calls against the deprecated identifier stop working — the same hard failure mode that took out Sonnet 4 and Opus 4 on June 15 and Opus 4.1 on August 5.
Anthropic gives at least 60 days’ notice before retiring a publicly released model, which is generous by industry standards and still not enough time if nobody on your team is watching the deprecations page. The failure pattern is consistent: the main application gets migrated on time, and then something forgotten breaks — a nightly batch job, an internal eval harness, a Zapier action, a vendor integration that hardcoded a model string two years ago.
Concrete action for this week: grep for claude-sonnet-4-5 across every repo, config store, notebook and CI secret you control, not just the ones you think are live. Five weeks is comfortable if you start now and unpleasant if you start on September 22.
Yesterday’s Incident Closed at 08:30 UTC — 3 Hours 26 Minutes
Closing the loop on Monday. The multi-model incident opened at 04:50 UTC, Anthropic identified the cause at 05:27, error rates began stabilising around 07:47, and the incident was marked resolved at 08:30 UTC — a total of 3 hours 26 minutes, affecting Mythos 5, Fable 5, Opus 5 and Opus 4.8.
That is longer than the early reporting suggested. The gap between “root cause identified” at 05:27 and “stabilising” at 07:47 is two hours and twenty minutes of known-cause degradation, which is the interval worth designing around. Knowing what broke and having capacity to absorb the fix are different problems.
August 25 has been clean so far — third-party monitoring shows claude.ai at 100% uptime over the trailing 24 hours with two minor issues logged. After a month averaging an incident every three days, a quiet Tuesday is worth noting rather than assuming.
Inference Hooks and Skill Scanning Open in Beta for Enterprise
Two Enterprise-plan features that address the same anxiety from opposite directions. Skill and plugin security scanning (beta) checks third-party uploads for malicious content before they land in a workspace. Inference hooks (beta) give compliance teams real-time DLP enforcement across chat, Claude Code, Cowork and more, inspecting prompts and tool calls before they reach the model.
The second one is the structurally important change. Until now, enterprise controls on Claude have mostly been perimeter controls — who can log in, which connectors are enabled, what gets retained. Inference hooks move the enforcement point inside the request path, which is the only place you can actually catch a customer record on its way into a prompt.
It also quietly acknowledges where the risk now lives. With 94 plugins in the marketplace and skills that activate automatically based on context, the attack surface is no longer “what the user typed” — it is whatever a bundled skill decided to pull in on the user’s behalf. Scanning the bundle and inspecting the request are both necessary.
Claude Code 2.1.243 Fixes a Startup Crash on glibc 2.44
Short release, narrow blast radius, high urgency if it hits you. 2.1.243 (August 25) fixes a crash on startup on Linux distributions shipping glibc 2.44 — named examples are Arch Linux, CachyOS and Fedora Rawhide. If Claude Code stopped launching for you after a recent system update, this is why, and the fix is available now.
The likely trigger is the CLI’s move to native binaries. Shipping a compiled binary instead of a Node bundle buys startup speed and removes the runtime dependency, at the cost of inheriting the platform’s glibc ABI surface. Rolling-release distros hit the new glibc first, which is exactly the population that reported this.
Practical note: if you manage Claude Code across a fleet with pinned versions, this is worth an out-of-band bump rather than waiting for your normal upgrade window. A tool that will not start is not a degradation you can route around.
2.1.242 Makes /cost Use Your Contracted Rates, Not List Price
The day before the Linux fix, 2.1.242 shipped four settings worth adopting. The headline one is modelPricing, a managed setting that makes /cost and telemetry report your organisation’s contracted per-model rates instead of public list price. If you negotiated an enterprise agreement, every cost number Claude Code has shown you until now has been wrong — usually high.
promptCacheTtl and subagentPromptCacheTtl split cache lifetime between the main conversation and subagents, so you can hold a 1-hour cache on the long-running thread while subagents stay at 5 minutes. That is the right shape: the main conversation accumulates context worth keeping warm, subagents are short and disposable, and paying 1-hour cache-write pricing on both was leaving money on the table.
Rounding out the release: modelPicker lets an org curate /model down to an ordered, labelled list — useful for steering teams off models that are about to retire — and /usage gained a Loops breakdown showing run count, total tokens, tokens per run and last run for each loop.
950+ MCP Servers in the Directory as the Stateless Spec Lands
The connectors directory now lists more than 950 MCP servers, described as in use by millions of people daily. That is the ecosystem number; the more interesting one is the protocol change behind it. The MCP 2026-07-28 revision — stateless core, standardised extensions, hardened auth — is being brought into Claude.
Stateless matters more than it sounds. A stateless core means a server no longer has to hold a session open to be useful, which makes MCP servers deployable as ordinary serverless functions behind a load balancer instead of long-lived processes with sticky routing. That removes most of the operational reason a team would decline to expose an internal system over MCP.
If you maintain a server, the migration is the work: standardised extensions mean capabilities you previously bolted on non-portably now have a defined home, and hardened auth means some of what you built by hand is now specified. Worth reading the revision before your next server, not after.
Cowork Now Runs in the Chrome Extension Side Panel
Claude Cowork now runs directly in the side panel of Anthropic’s Chrome extension, with skills, plugins and connectors working in the browser for the first time and no extra setup required. The marketplace it draws on has grown to 94 plugins.
The significance is placement rather than capability. Cowork’s pitch has always been aimed at people who do not open a terminal — and for that audience, a desktop app is still a context switch away from the browser tab where the actual work is happening. Putting the same skills and connectors in a side panel next to the CRM, the invoice, or the shared doc is where that pitch either lands or does not.
The Public S-1 Is Expected Before the Month Closes
Six days left in August, and the expectation is still that Anthropic publicly files its prospectus by end of month, following the June 1 confidential S-1 and a $965 billion Series H mark. The roadshow is pencilled for September, with a target October Nasdaq listing.
The date is not arbitrary. SEC guidance requires the public filing at least 15 days before a traditional roadshow — so a September roadshow forces the flip in roughly the next two weeks whether the company feels ready or not. Secondary markets are currently pricing an implied valuation of $1.05 to $1.15 trillion, well above the last private round.
Ticker, exchange, price range and final valuation all remain unannounced. Every revenue figure in circulation is still reporting rather than audited disclosure; the prospectus is the first document that changes that.
The Seat-Count Arms Race: Deloitte 470K, KPMG 276K, Accenture 30K Trained
Worth seeing the professional-services deals as one number rather than three announcements. Deloitte is rolling Claude out to 470,000 people across its global network — Anthropic’s largest enterprise deployment to date. KPMG is giving access to all 276,000+ employees. Accenture is training roughly 30,000 professionals and became a premier partner for coding with Claude Code.
That is north of three quarters of a million seats at three firms, and the seats are the smaller half of the story. Each of these firms bills clients for AI transformation work, which means every deployed consultant is a distribution channel into the client base. Anthropic is not selling licences here so much as renting the implementation layer of the enterprise software market.
The Deprecation Treadmill Is the Tax Nobody Prices Into the Enterprise Pitch
Line up 2026’s retirement dates and a rhythm appears. Sonnet 4 and Opus 4 went on June 15. Opus 4.1 went on August 5. Sonnet 4.5 goes on September 29. That is a model family retiring roughly every seven weeks, each with 60 days’ notice, each requiring somebody to go find every place a string got hardcoded.
The 60-day policy is genuinely good practice — better than several competitors, and clearly documented. But policy quality is not the issue. The issue is cadence. A 60-day window is a comfortable amount of time to migrate one thing. It is a much less comfortable amount of time when the windows overlap, and by late September a large shop will have absorbed three separate migrations in a single quarter while also fielding new model launches it is being encouraged to adopt.
Put that next to today’s other stories and the tension is clear. Anthropic just shipped modelPicker so organisations can curate which models their teams see — a tool that exists specifically because model churn has become an internal governance problem. It is signing deals that put Claude in front of 470,000 Deloitte staff and 276,000 at KPMG. Those are populations that build things and then stop thinking about them. Every one of those deployments accrues quiet dependencies on specific model identifiers, and the treadmill does not slow down to accommodate them.
The honest version of the enterprise pitch would name this cost. Frontier capability and long-term stability are in genuine tension, and Anthropic has — reasonably, defensibly — chosen capability. But the buyer signing a five-year transformation programme should be told they are also signing up for a permanent migration function. The firms that handle this well will be the ones that treat model identifiers as configuration from day one, keep a live inventory of every place a model is pinned, and staff the upgrade path as ongoing work rather than an emergency. The ones that do not will discover their inventory on the morning a batch job starts returning errors.