Anthropic to SpaceX: We’ll Take That Record
Bloomberg reported yesterday that Anthropic has told investors it expects its IPO to match or exceed SpaceX’s record offering. For scale, that is the number to beat: SpaceX raised $75 billion — $86.2 billion with the over-allotment — at a $1.77 trillion valuation in June. Fortune’s newsletter this morning adds a telling detail: CFO Krishna Rao skirted the valuation question in recent investor briefings rather than putting a figure on it.
The context behind the ambition: Anthropic was last valued privately at $965 billion in May after a $65 billion Series H, investors have separately floated a $2 trillion target for an October listing, and the revenue run rate reportedly crossed $65 billion last week. Two caveats before anyone builds a spreadsheet on this. First, the “match or beat SpaceX” framing is sourced to people familiar with the briefings, not to an Anthropic statement. Second, a CFO who won’t name a valuation to his own investors is usually signaling that the number is still moving — in either direction. OpenAI is believed to be targeting 2027.
Enterprises Can Keep Their Retention Data in Their Own Cloud
Anthropic said yesterday it is building a new safety system, shipping later this year, that keeps the 30-day retention requirement for prompts and outputs on “covered models” — currently Mythos 5 and Fable 5 — but lets enterprises hold that data in their own cloud infrastructure rather than Anthropic’s. The company says the design has been shaped by coordination with more than 100 customers, including Salesforce.
This is a walk-back, and worth naming as one. The retention mandate introduced over June and July triggered a well-documented enterprise backlash — the Wall Street Journal covered the Silicon Valley reaction — because for a regulated buyer, “your prompts sit on the vendor’s infrastructure for 30 days” is not a policy detail, it is a procurement blocker. Moving custody to the customer’s own cloud preserves the safety property Anthropic wanted while removing the part legal departments actually objected to. If a covered-model deal stalled at your company this summer, the objection you filed is the one being answered.
Computer Use Ships GA, and Playground Replaces the Workbench
Computer use is now generally available as computer_toolset_20260801. The GA toolset drops the beta header, adds batch actions so Claude can take several actions in a single turn, turns zoom on by default, and allows per-member configuration through a configs array. It runs on Fable 5, Mythos 5, Opus 5, Sonnet 5 and Opus 4.8. Older beta versions still work, but the request shape and tool handling changed, so there is a published migration guide from computer_20251124.
Alongside it, the Console’s Workbench has been rebuilt and renamed Playground. It now supports every Messages API parameter, ships runnable templates for features like code execution and web search, and shows the full SDK request next to the API response for each run — which is the part that makes it useful for debugging rather than just prompting. Note the housekeeping that went with it: the legacy Workbench and the experimental prompt tools endpoints (generate_prompt, improve_prompt, templatize_prompt) were retired on August 17. If you had scripts calling those, they are erroring now.
Browser Use: 27 Tools, All Running on Your Infrastructure
The more interesting sibling to computer use GA is browser_toolset_20260801, a brand-new client toolset for driving a browser that your application hosts. Where computer use screenshots and clicks a whole desktop, browser use works inside a viewport and reads the page’s accessibility tree, elements, forms and tabs — so Claude gets element references and structured form input instead of guessing at pixel coordinates.
One entry in your tools array expands to 27 member tools by default (navigate, read_page, left_click, screenshot and friends) plus four opt-in ones: javascript_exec, file_upload, read_console and read_network. The architectural point is the one to keep: nothing executes on Anthropic’s side. Every call runs against your own browser automation, in your own network, behind your own auth. For anyone who wanted agentic browsing but could not put a third party inside the session boundary, that is the whole ballgame — and the last two opt-in tools quietly make Claude a competent frontend debugger.
2.1.238 Unlocks Private Plugin Marketplaces
Claude Code v2.1.238 shipped yesterday, a 39-item drop, and the headline item is headersHelper for plugin marketplaces. It runs a command to mint HTTP headers — a short-lived token, typically — for catalog and same-origin archive fetches. In plain terms: authenticated, private, enterprise-gated plugin marketplaces are now possible. It only fires on install or update, the command is displayed first, and it sits behind a [y/N] prompt (or -y).
Also in the release: a keybindingFlavor setting — set it to "readline" for Bash-style Ctrl+W that deletes back to whitespace, with "classic" as the default. Two new self-hosted runner flags: --defer-shutdown-max-min <minutes>, which on SIGTERM keeps serving attached sessions and parks the rest before exiting, and --proxy-authorization-command / --proxy-authorization-file for egress proxies that need a fresh header per connection. Plus a fix worth knowing: unbounded memory growth in long interactive sessions is resolved, with subagent tool results now released once they leave the display window. One behavior change to note — Ctrl+L / Cmd+K in fullscreen only repaints now; the double-press /clear shortcut is gone, because one-row nvim terminals were triggering accidental clear loops.
OpenAI Aims Straight at the Retention Policy
Two days ago OpenAI previewed Private Safety Processing for select customers, and the targeting is not subtle. It is an agentic, cross-session misuse-monitoring system that retains zero customer data — extending Zero Data Retention from per-session to what OpenAI calls “long-horizon” monitoring across multiple conversations, emitting only a “narrowly defined signal” when something trips.
That is a product built to be pitched in the same room where Anthropic’s 30-day retention requirement is being objected to. Anthropic’s counter-position on the record: human review happens only through a controlled access path with a small set of approved reviewers, logged in a tamper-proof log the reviewers cannot suppress or modify. Read the two announcements together with yesterday’s retention news and the sequence is clear enough — OpenAI shipped the attack on Wednesday, Anthropic announced the fix on Thursday.
Ramp: Anthropic Still Leads Business Users, but the Gap Is Closing
New Ramp data published yesterday, covering 70,000+ American businesses, puts Anthropic at nearly 44% of paying business AI users in July against OpenAI’s nearly 40%. The trend line is the story: OpenAI is growing faster in Q3 to date. Ramp economist Ara Kharazian credits GPT-5.6 Sol’s developer traction, and offered a blunt assessment of Anthropic’s side: Fable 5 “disappointed both in adoption and real-world application given price + data retention requirements imposed by regulators.”
Worth holding two things at once. The overall pie is still expanding fast — the share of Ramp customers paying for any AI rose from 50% in March to nearly 56% in July — so this is a growth-rate contest, not a zero-sum one. But the specific reason cited for Fable 5 underperforming is price plus data retention, which is exactly the objection Anthropic moved to fix yesterday. When an analyst and a policy reversal name the same cause in the same week, that is not coincidence.
Three Incidents in 24 Hours
Anthropic opened a multi-model incident at 19:16 UTC yesterday — elevated errors on “some Claude models” — which put claude.ai, api.anthropic.com, Claude Code and Cowork into partial outage simultaneously while Console and Claude for Government stayed up. That split points at the model-serving path rather than auth or admin. It was the third incident in roughly 24 hours, after a Google connectors outage (Sheets, Docs, Slides, Chat; 18:32–19:01 UTC) and Wednesday’s Opus 5 / Haiku 4.5 event.
The 90-day uptime figures on the status page: claude.ai 99.38%, API 99.48%, Claude Code 99.4%, Cowork 99.49%. Those are not catastrophic numbers, but they are roughly a four-nines-short-of-four-nines profile for infrastructure that a lot of people have quietly moved into production paths, and the incidents are clustering rather than spreading out. The advice is the same as it was on Tuesday, and it stays here until the streak breaks: if Claude sits in a production path, have a failover model configured and tested, and make sure your retry logic backs off instead of amplifying.
The Retention Reversal Is the Most Important Thing That Happened This Week
It will not get the headlines. The IPO number will. But look at what actually moved: OpenAI shipped a zero-retention monitoring product on Wednesday explicitly aimed at Anthropic’s policy, Ramp published data on Thursday citing price plus data retention as why Fable 5 underperformed, and Anthropic announced a fix for the retention half on the same day. Three data points, one week, one cause.
The interesting part is what got conceded. Anthropic did not drop the 30-day requirement — the safety property survives intact. It changed who holds the bytes. That is a precise concession, and it suggests the company worked out, across 100-plus customer conversations, that the objection was never really about retention as a concept. It was about custody. Legal departments at regulated buyers can write a policy around data they control on infrastructure they own. They cannot write one around data sitting on a vendor’s servers under a rule the vendor can change.
Read that against the rest of the week — four surfaces leaving beta, computer use going GA, browser use arriving as a client-side toolset that runs entirely on your own infrastructure — and a pattern shows up that is more coherent than any single release note. Anthropic is systematically moving the trust boundary back toward the customer. Your browser, your sandbox, your cloud, your headers on your private plugin marketplace. For a company weeks from asking public markets to price it above SpaceX, that is the right instinct: the ceiling on enterprise AI revenue was never model quality, it was the number of review boards willing to let a vendor hold the keys. Now the harder question. All of that architectural care is being delivered on a status page that logged three incidents in a single day. Custody you can fix with a design document. Availability you cannot.