Salesforce and Anthropic Launch Claudeforce — and Salesforce Lends Out Its Suffix
Salesforce and Anthropic announced Claudeforce, a substantial expansion of a partnership that until now was mostly a model-availability arrangement. Claude becomes a default reasoning model across Agentforce, Slack and Salesforce developer tools, with Salesforce’s enterprise harness — data, workflows, business logic, permissions and governance — exposed to agentic experiences wherever the work happens.
The concrete deliverable at launch is Salesforce in Claude, a plugin carrying 37 prebuilt sales skills. Sellers can reason over live revenue context, automate pipeline updates and take governed actions without leaving Claude. Pilot access is open now, with an open beta targeted for September. More integrations across Claude, Salesforce and Slack are promised.
The naming is not a throwaway detail. This is the first time Salesforce has attached its “force” suffix to another company’s product — a suffix it has spent two decades reserving for things it builds. Marc Benioff framed it on the Q2 earnings call as an answer to “SaaSpocalypse” anxiety, which tells you what the deal is really about: Salesforce placing a bet that the safest position in an agent-first world is to be the system of record the agents have to go through.
Claude for Teachers Opens Free to Schools and Districts
Claude for Teachers is now available to schools and districts as a free Enterprise offering, extending a program that launched for individual US K-12 educators in July. Organizations that sign up by June 30, 2027 get a full year of free access.
The product is deliberately educator-only — no student accounts — with teaching skills grounded in learning science and lesson content mapped to academic standards in all 50 states. Privacy terms include a K-12 Data Processing Addendum written for FERPA compliance, and Anthropic states that data submitted through Claude for Teachers is not used for model training.
The district-level move is the meaningful change. Individual teacher signups are a marketing channel; district deployments are procurement, IT review and multi-year habit. OpenAI and Google are running the same play, and classrooms are turning into one of the more contested distribution fights in the industry. Some educators have pushed back on the whole category, which is worth reading alongside the announcement rather than after it.
Computer Use, Browser Use, Skills API and Files API Are All GA
The agent stack came out of preview together. Computer use, the Skills API and the Files API are generally available on the Claude Platform, joined by a new browser use tool built specifically for agents working in web applications.
The upgrade that matters most in practice: computer use now takes several actions per turn instead of one action per model call. Tasks that used to burn a round trip on every click finish in a fraction of the calls. The browser use tool goes further — alongside the screenshot it reads the structure of the page and acts on a named field or button rather than a screen coordinate, which is the difference between an automation that survives a CSS change and one that does not.
The Skills API gives agents versioned folders of instructions, scripts and templates loaded only when a task calls for them, and the Files API gives them reusable documents and finished outputs. Together they are the supported path for building on Claude Managed Agents: automate software that has no API, carry your team’s expertise into the agent, and get real files back.
Claude Code 2.1.250 — Restricted Mode, Cross-Session Messaging, Enterprise Credits
The release line is now at 2.1.250. Three additions are worth acting on.
Restricted mode (--restricted or CLAUDE_CODE_RESTRICTED=1) strips the built-in tools that run commands or code, drops WebFetch unless you name it in --tools, confines file tools to the working directory, refuses bypassPermissions, and ignores user, project and local settings files. This is the flag for running untrusted repos or shared CI, and it is what you point at during a security review.
Cross-session messaging pairs two tools: ListAgents to discover which of your other sessions are reachable, and SendMessage to deliver work to one by name. Note the limits before you design around it — it needs v2.1.224+, runs on macOS and Linux (including Linux under WSL 2, but not native Windows), and is unavailable on Bedrock, Claude Platform on AWS, Google Cloud’s Agent Platform and Microsoft Foundry. Recent patches added a fallback to a private per-user /tmp directory and a 30-second timeout on inbox connections that never send a complete line.
Also landed: usage credits for Enterprise, better diagnostics for server-managed settings, improved agent workflows, Remote Control fixes, and prompt caching improvements, plus a long list of session, login, MCP and terminal bug fixes.
Admin API Goes GA, and Managed Agents Get Domain Allowlists
Quieter releases, but these are the ones that unblock procurement. The Admin API user-management endpoints are now generally available for Claude Enterprise organizations, covering members, invites, groups and custom roles. The broader Admin API now reaches organization info, workspaces, API keys, rate limits, service accounts and customer-managed encryption keys, and is exposed through the ant CLI and multiple SDKs rather than raw HTTP.
For Managed Agents, you can now restrict which sites an agent’s web_search and web_fetch tools can reach by setting allowed_domains or blocked_domains on the tool entry in the agent_toolset_20260401 configs array. Pair that with the session budgets that shipped this week and an autonomous agent finally has both a spending limit and a network boundary.
Also in beta for Enterprise: skill and plugin security scanning, which checks third-party uploads for malicious content before they land in a workspace. As skills become shareable artifacts, that scan stops being optional.
Thirty-One Days Until Sonnet 4.5 Retires
Claude Sonnet 4.5 retires September 29, 2026. After that date, calls against claude-sonnet-4-5 fail outright — the same hard cutoff that took Sonnet 4 and Opus 4 on June 15 and Opus 4.1 on August 5. Opus 4.5 is unaffected.
If you are still on anthropic-sdk-python below v1.0, do both migrations in one pass. The pinned model identifier and the removed temperature, top_p and top_k parameters live in the same call sites. Sweep repos, config stores, notebooks, CI secrets and vendor integrations — not just the services you think are live.
Anthropic Pitches Investors a $30 Trillion Market
Ahead of its listing, Anthropic is reportedly telling prospective investors its total addressable market exceeds $30 trillion. That would top the $28.5 trillion figure SpaceX put in front of investors before its June IPO, and it sits in the same neighborhood as US GDP — roughly a quarter of global output.
The reasoning behind the number is the scope of work AI models could eventually perform, which is to say the number is a claim about labor, not software. The revenue underneath it is moving fast in its own right: Anthropic more than doubled revenue to $11.6 billion in Q2, up from $4.73 billion in Q1. Reported IPO parameters put the raise as high as $100 billion at a $2 trillion valuation, with prospectus documents expected shortly and a debut possible as early as September or early October.
Analysts have been blunt. Similar TAM figures have been called aggressive marketing rather than serious math, and Fortune’s read was that the number’s implausibility may be the point — an anchor set high enough that a $2 trillion valuation starts to sound modest by comparison. Worth remembering that none of this is confirmed by Anthropic, and the only document that settles it is the S-1.
Two Hyperscalers Looked at the Same Megawatts and Said No
We covered the $45 billion Nscale deal on Wednesday — roughly 460 megawatts over six years at the Monarch Compute Campus in Mason County, West Virginia, running NVIDIA Vera Rubin hardware, with first power expected late 2027. New reporting from Semafor adds the part that changes how it reads.
Microsoft and Google were each in talks for this exact lease and both walked away. Microsoft exited over the summer during a portfolio review that scrutinized on-site power generation. Google’s conversation ended amid an internal look at computing needs and capex management. Two of the most capacity-hungry companies on earth evaluated the same 460 megawatts and passed.
That is not a verdict on demand. It is a verdict on this site, this power arrangement, this timeline. Anthropic taking it means either a different risk tolerance, a different read on identical facts, or simply less optionality — a company without its own cloud has fewer places to put a bet this size. The tell to watch is late 2027. Data center timelines slip routinely, and a six-year commitment anchored to a campus that does not exist yet carries schedule risk a signed contract does not remove. Nscale, reportedly prepping its own IPO as soon as next month, is using this contract as the anchor.
Claudeforce Is Salesforce Choosing Which Half of Its Business to Defend
Strip away the branding and Claudeforce is a company deciding, publicly, which part of itself is durable.
The “SaaSpocalypse” argument that has hung over Salesforce for two years goes like this: if an agent can read your data, apply your business rules and take actions on your behalf, the interface layer — the screens, the tabs, the seat licenses — is the part that evaporates. Salesforce sells seats. That is the fear, and Benioff answered it on the earnings call by shipping the thing the fear predicted.
Look at what each side actually gave up. Salesforce gave up the assumption that its UI is where the work happens. Salesforce in Claude means a seller can update a pipeline without opening Salesforce. Anthropic gave up a slice of neutrality: Claude is now a default reasoning model inside a specific vendor’s stack, and the suffix on the name makes that hard to walk back quietly.
What Salesforce kept is the part it believes cannot be cloned: data, business logic, permissions and governance. That is the actual bet. If agents commoditize interfaces but not systems of record, Salesforce ends up more entrenched, because every agent action still has to pass through its permission model and land in its objects. If agents eventually build their own systems of record — and there is nothing structural preventing that — Salesforce has spent this deal training customers to work somewhere else.
The 37 prebuilt skills are the detail I would watch. Skills are portable, versioned, and now scannable and shareable as first-class platform artifacts. A skill that encodes “how a Salesforce pipeline update works” is useful precisely because it abstracts Salesforce. That abstraction is a gift while the partnership holds and a liability the moment it does not. Salesforce is betting the governance layer is the moat. The next two quarters, and specifically whether the September open beta stays exclusive, will say whether that bet was made from strength or from the other thing.