Anthropic Stops Keeping Your Logs and Starts Watching Them in Your Bucket
Anthropic is replacing its enterprise data retention policy with Enterprise Frontier Safeguards, and the trick it pulls off is one the industry has been arguing was impossible. Customers get zero data retention. Anthropic keeps nothing. But automated misuse detection still runs — against logs stored in the customer’s own Amazon S3, Azure Blob Storage or Google Cloud Storage, under the customer’s own encryption keys and access policies, with no Anthropic human review required.
That is a genuine architectural answer to a fight that had been stuck in policy. The standing trade was blunt: accept retention so the vendor can catch abuse, or take ZDR and lose cross-session safety monitoring entirely. EFS routes around it by moving the data to the customer and sending the detection to the data. Anthropic says it built the system with more than 100 customers across financial services, healthcare, manufacturing, telecom, law, retail and the public sector — the exact roster that had been refusing to sign.
Two details tell you how badly it was needed. It is free — no premium tier, no security-as-upsell. And it ships everywhere at once: Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Google’s Agent Platform, Microsoft Foundry. Rollout is phased, beginning later this fall, with eligible customers getting ZDR on Fable 5 and Fable 5.1 during the transition. Companies do not rebuild their safety stack and give it away for free unless the old design was actively losing deals.
Claude for Teachers Goes Free for Every US School District
Claude for Teachers is now available to US K-12 schools and districts as a free Enterprise offering. Educators and school staff get free access; qualifying organizations get the full Enterprise suite — single sign-on, role-based access control, administrators who can add and remove staff, set policies and see adoption across schools. Sign up by June 30, 2027 and you get a full year free.
The product underneath is not a repackaged chatbot. It carries teaching skills grounded in learning science, connections to academic standards in all fifty states and the curricula beneath them, and hooks into the K-12 tools teachers already run. Anthropic describes its commitments as FERPA-aligned and says Claude for Teachers data is not used to train its models — which is the first question any district counsel asks and the one that usually ends the conversation.
Read the shape of the offer rather than the generosity. Free teacher accounts are a marketing expense. Free Enterprise with SSO and RBAC is an installation — it puts Anthropic inside the district’s identity system, its policy layer and its procurement paperwork. Districts do not rip out an identity integration that works. The year is free; the switching cost starts accruing on day one.
The Life Sciences Verification Program Opens the Biology Boundary, Selectively
Anthropic launched the Life Sciences Verification Program in partnership with the US government, so verified life sciences professionals can use Claude Mythos 5.1 with safeguards tuned for professional R&D. First participants are enrolled. The program is an invite-only beta that relaxes biology safeguards for research work while leaving every other safeguard in place.
This is the mechanism the Fable/Mythos split was built for. Same underlying system, two names, two safety postures: Fable 5.1 generally available, Mythos 5.1 gated behind trusted access. Rather than loosening biology restrictions for everyone and hoping, or holding them shut and watching legitimate researchers hit walls, Anthropic is opening the boundary per-verified-user — with a government partner doing the verifying. Access today is limited to a set of US organizations, with coordination underway to extend to more domestic and international partners.
What is missing is the part that matters for anyone hoping to get in. Anthropic has not named the participating agency, published eligibility criteria, or given a date for broader applications. That opacity is defensible for a biosecurity program and still leaves a real problem: a capability boundary policed by an unnamed government process is one that researchers cannot plan around, appeal, or audit. Worth watching whether the criteria ever get published.
Managed Agents Get a Spend Cap, a Passport, and Your Repo’s Skills
Three controls landed on Claude Managed Agents, and together they close the three questions that stop an agent reaching production. First, budgets: set a hard cap on a session’s spend, and the session pauses with a budget_reached stop reason when it hits the limit. Raise the budget, it resumes. A runaway loop now costs a defined number instead of a discovered one.
Second, geo-pinned inference. Set model.inference_geo to us to keep inference in-region, billed at 1.1x, or global to run wherever there is capacity at the standard rate. A ten percent premium on data residency is a remarkably honest piece of pricing — it says the cost is real, it is small, and it is yours to decide. For regulated buyers this was not a preference, it was a blocker.
Third, skills load straight from a mounted GitHub repository. Anything in .claude/skills is auto-discovered at session start, no manual copying into the agent environment. If your team already maintains skills for Claude Code, Managed Agents pick them up unchanged. That quietly makes the skills directory a shared asset across the local agent and the hosted one, which is a much bigger deal than a convenience feature reads like.
The Sonnet 5 Price Increase Is Cancelled
Quietly one of the more consequential lines in the release notes: Claude Sonnet 5’s introductory pricing is now the standard price. The increase to $3 / $15 per million tokens, scheduled for September 1, 2026, will not happen. Sonnet 5 stays at $2 / $10, permanently.
A promised price rise that gets withdrawn is worth more attention than a price cut, because it costs the vendor credibility to reverse and they do it anyway. Everyone who built a cost model around a 50 percent increase landing this month just got the money back, and the volume-tier workhorse of the lineup stays cheap at the exact moment the company is trying to show revenue growth is durable rather than promotional.
Two smaller API notes ride along. An anthropic-workspace-id response header now returns the workspace that a request’s API key or access token resolved to — small, but it removes a real class of “which workspace did that spend land in” debugging. And fast mode is gone for Claude Opus 4.7: requests sending speed: "fast" now return an error rather than degrading gracefully. If that string is in your codebase, it is a breaking change today.
2.1.266 Fixes the Gateway Regression, and Skills Learn to Audit Themselves
Claude Code 2.1.266 shipped as a one-change release, which is its own kind of signal. It fixes a regression that ignored the CLAUDE_CODE_USE_GATEWAY environment variable, restoring LLM gateway and proxy behavior. That variable is how enterprises route Claude Code through their own inspection and billing layer, so breaking it does not degrade an experience — it takes the tool out of compliance and out of the building. Hence the fast single-purpose patch.
The more interesting addition landed a few builds earlier. /skill-doctor lists which loaded skills a session never used and what each one costs you in context. Skills have quietly become a place where teams accumulate cruft — every one loaded is prompt real estate spent whether or not it is touched. A command that prices your unused skills is the first tool that makes pruning them a measurable act rather than a guess.
Alongside it: an “Organization policy” line in /status and claude doctor that says why a managed policy could not be loaded, refresh failures surfaced in /status, an explanation printed when declining the managed-settings dialog causes an exit, and helper timeouts reported as timeouts instead of generic failures. All of it is the same fix applied repeatedly: stop failing silently at the enterprise boundary.
Claudeforce Moves Toward Open Beta With 37 Prebuilt Sales Skills
Salesforce and Anthropic’s Claudeforce partnership is heading into open beta this month. The first shipping piece is Salesforce in Claude, a plugin carrying 37 prebuilt sales skills that let sellers and agents reason over live revenue context, automate pipeline updates and take governed action from inside Claude. It is with select pilot customers now.
The word doing the work in that description is governed. Salesforce is not exposing a data dump — it is putting its business logic, permissions model and action governance in front of the model, so an agent updating a pipeline is bound by the same rules a rep is. That is the difference between an integration that survives a security review and one that reads records into a chat window.
Note where the work is happening: inside Claude, not inside Salesforce. The CRM becomes the governed system of record; the assistant becomes the surface. For anyone building on this stack, 37 prebuilt skills is also a shape to study — it is what a major vendor decided the atomic units of its own workflow are, published for free.
Another Morning of Downtime Reports, and the Status Page Says Nothing
Claude drew a surge of user downtime reports on the morning of September 9, beginning around 7:30am Eastern on DownDetector. The official status page showed no incident at the time — which means either the problem was narrow enough not to trip a threshold, resolved before it did, or is simply not being reported the way users experience it.
Context matters more than any single morning. Claude logged twenty-one incidents between August 12 and September 2 — elevated errors and degraded performance across services. And on September 3, ChatGPT, Claude and Grok went down at roughly the same time, a coincidence that says more about shared infrastructure beneath the frontier labs than about any one of them.
The uncomfortable part is not the downtime, it is the gap between user reports and vendor acknowledgement. Every governance feature shipping this week — budget caps, domain allowlists, ZDR with detection — is aimed at making Claude safe to put in a critical path. Reliability signal is the other half of that, and a status page that stays green while users are down erodes it faster than an outage does. If you are running agents unattended, instrument your own error rates and do not wait for the dashboard.
The Constraint Is the Product Now
Line up everything that shipped this week and notice what almost none of it is. Not one is a capability. Zero data retention with detection that runs in your bucket. A hard spend cap that halts a session. Inference pinned to a jurisdiction for ten percent. Skills loaded from a repo you control. A command that tells you what your unused skills cost. A biology boundary opened only to people a government verified. Yesterday, a domain allowlist for web tools. Every one of them is a limit — a way to tell the model what it may not do, may not touch, may not spend, may not keep.
That is not a slow week. That is a company that has decided the binding constraint on revenue is no longer intelligence. Fable 5.1 more than doubled a scientific research benchmark eight days ago and it changed roughly nothing about who can deploy it. The deals that stall do not stall on whether the model is smart enough. They stall on a security review asking where the logs live, a finance team asking what happens if the loop does not terminate, a compliance officer asking which country the inference ran in. Capability opens the meeting. Constraints close it.
EFS is the clearest instance because it solves the problem by giving something up. Anthropic surrendered the data. It kept the detection and moved it to where the data now lives, under keys it does not hold. That is a harder engineering problem than retaining logs, and the only reason to solve it is that more than a hundred enterprise customers made retention the reason they would not sign. Then the company gave the result away free. Security is not being sold here — it is being removed as an objection, which is what you do when the objection is worth more than the upsell.
The pattern repeats at every level of the stack, and the pricing tells you it is deliberate. Data residency: 1.1x. EFS: free. Sonnet 5’s promised 50 percent increase: cancelled. Anthropic is putting a small, honest, legible number on control and refusing to make it a luxury tier. Compare that to the alternative playbook — residency in the enterprise SKU, ZDR behind a call with sales, safety features as a badge on a pricing page. Charging for control caps your ceiling at the customers who can afford it. Giving it away moves your ceiling to everyone who was blocked by it.
Which is why this reads as a coherent strategy four weeks before a prospectus rather than a grab-bag of housekeeping. The K-12 giveaway is the same move in a different market: free Enterprise with SSO and RBAC is not charity, it is installation. LSVP is the same move again — not loosening biology safeguards, but building a verified door through them. The story Anthropic needs to tell is not that Claude is the most capable model; that claim expires with the next release from anyone. It is that Claude is the one you are allowed to actually deploy — in a bank, a hospital, a school district, a lab. Capability is a leapfrog game. Permission compounds.