Anthropic files S-1 with $16.3B H1 revenue and ticker 'ANT', Claude Duet launches as proactive background agent, unified Chat+Cowork goes GA for Enterprise

Anthropic publicly filed its S-1 prospectus on October 26, targeting a Nasdaq listing under ticker 'ANT' with lead underwriters Goldman Sachs and Morgan Stanley. Key financials from the filing: first-half 2026 revenue of $16.3B (vs $4.6B for all of 2025); Q2 revenue of $11.6B up from $4.7B in Q1; Q2 operating income of $974M — its first operating profit — after a $1.7B operating loss in Q1; Q2 compute costs of $5.6B (up from $3.4B in Q1). Net losses were $80.7B in H1 2026 compared to $42B for all of 2025, driven by heavy training and infrastructure investment. Revenue per employee exceeded $6.6M over the past year (approximately 5,300 employees as of September 30). Major commitment disclosed: up to $120.5B in compute spending through 2029 under agreements with SpaceX AI, roughly $3.1B per month once capacity is delivered in December 2026, under an effort Anthropic calls 'Anthropic Direct' where it leases and operates its own data centers. Last private valuation: $965B (Series H, May 2026). Expected IPO range: $1.5T–$2T, per reports. Lockup structure is non-standard — some insiders can trade on day one, with staggered unlocks across 20+ windows over approximately one year, replacing a standard 180-day lockup. Key risks flagged: the insurance industry may not cover liability from a catastrophic AI incident; 'increasingly restrictive AI regulation'; and the prospectus states AI models might pose a 'catastrophic or existential risk to humanity'.

Claude Duet launched October 21 as a research preview for US Pro and Max subscribers by waitlist — no extra charge, but background tasks draw from your existing plan's usage limits. Available as separate iOS and Android apps and at claude.ai; runs on Claude Opus 5.5 with no model picker. Memory, skills, and connected apps from your existing Claude account carry over; over 6,000 connectors and plugins are supported including Gmail, Slack, QuickBooks, Stripe, Gusto, HubSpot, and Shopify. How a day works: a morning brief surfaces expiring contracts, unpaid invoices, candidates awaiting replies, and similar items across connected apps with suggested next steps. During the day it checks accounts every few hours and can run large tasks on a cloud VM while your laptop is closed. Approval gates are strict: Duet reads email, drafts messages, pulls data, and does research automatically, but always pauses before sending messages, making payments, signing documents, or deleting anything. Trading securities and moving money between accounts are blocked even with approval. A classifier screens outgoing messages for sensitive information. Competitor context: Meta Muse (September 8) targets personal tasks, OpenAI Dots (September 29) is tied to higher ChatGPT tiers and unavailable in EEA/UK. Team and Enterprise availability: no date announced.

The unified Claude experience — combining chat and Cowork in a single interface — became generally available for all Claude Enterprise plans on October 23, following its September rollout to Pro and Max. Users no longer choose between chat and Cowork before starting: conversations begin as chat, and when a task needs code execution or file work, Claude continues in the cloud and returns the finished result. Cloud work continues after the user closes their laptop and can be resumed from any device. Existing chats, projects, connectors, skills, and plugins carry over automatically. Admins get four usage controls: Efficient mode (caps agentic work to Sonnet and Haiku models), model entitlements per role, spend limits at org/group/member level, and pooled group budgets. Agentic work runs in isolated VMs with network egress off by default; connector credentials are kept outside the sandbox. Note: organizations using zero data retention for Claude Code, HIPAA-ready configuration, or CMEK remain on their current setup. Enterprise pricing stays usage-based.

Claude Security expanded from on-demand scanning to continuous scanning in a public beta for Enterprise customers. After a deep baseline scan, the system rescans code changes on a schedule based on their risk and complexity. Cross-repository findings are now supported — a project can span multiple repos to catch issues like a backend trusting input a client was supposed to validate. Fix tracking closes findings automatically when a patch lands (in Claude Code or elsewhere) and reopens them if the issue recurs. Each project gets a knowledge base describing deployment context and applicable policies, plus an editable AI-generated threat model. Each finding includes severity, an attacker path explanation, and sometimes a diagram. Integrations: API, webhooks (Slack, Jira), MCP connector for Claude.ai and Claude Code, and role-based access controls. Pricing: direct token costs with no platform fee, configurable spend limit; no Cyber Verification Program access required. Launch offer: up to $25,000 in credits per month for two consecutive months through December 31 — applies only to projects created under the new experience. A Claude Security plugin for Claude Code is also in beta for local scanning using models in your Claude Code account.

Seven releases shipped this week. v2.1.295 (Oct 20): onFailure: 'block' for command and HTTP hooks — a hook that fails, times out, or exits unexpectedly can now block the action it guards; terminal status signaling (OSC 7501). v2.1.304 (Oct 20): /auto-mode-setup guided setup for auto mode; /claude-api managed-agents-secure to compare agent settings against a security guide; Code Review gains Quick/Standard/Deep levels. v2.1.305 (Oct 20): sandbox.allowNestedUserNamespaces for Linux user namespaces (note: can bypass file protections — use with care). v2.1.306 (Oct 21): fixed claude -p inside cloud sessions reporting 'Not logged in'; fixed API errors misreported as 'prompt is too long' causing unnecessary compaction. v2.1.307 (Oct 22): /claude-api api-version to update code to API version 2026-10-21; claude_code.safety_event OpenTelemetry event; @claude review stop for Code Review. v2.1.309 (Oct 23): tightened sandbox and permission safeguards — Bash permission checks, MCP config protection; claude ultrareview requires workspace trust; sed -i now runs as a normal Bash command so /rewind no longer restores its edits. v2.1.310 (Oct 25): Monitor tool and plugin monitors available on Bedrock, Vertex, and Foundry; requests retry without beta values that some gateways reject; Enterprise custom roles can now control access to cloud sessions and Remote Control.