Claude Cowork and chat come together for Enterprise
B · Useful, limited Short Anthropic product announcement with concrete admin, security and cost-control details for enterprise buyers; cost figures are vendor claims and the customer quotes are promotional.
Summary based on the full text·Original: about 2 minutes to read
In short
An Anthropic announcement (23 October 2026) that chat and Cowork, its agentic mode for work with code and files, are now one product for Claude Enterprise, after the same change for Pro and Max in September. Conversations start as chat and move into a managed cloud environment when a task needs to run code or handle files, continuing after the user closes the laptop. The post sets out the security model and four admin controls for agentic usage cost.
Key numbers
- 40%lower running cost claimed for Opus 5.5 versus Opus 5 on typical workloads
- up to 30%lower running cost claimed for Sonnet 5.5 versus Sonnet 5 for most work
- fouradmin usage controls: Efficient mode, per-role model entitlements, spend limits, pooled group budgets
As reported by the source.
English·Web page·opens claude.com
What it reports
- One Claude across web, desktop and mobile; chats, projects, connectors, skills and plugins carry over.
- Agentic work runs in cloud sessions, each in its own isolated virtual machine, with network egress off by default for Enterprise and connector credentials kept outside the sandbox.
- Existing enterprise controls (SSO, Compliance API) apply; data is not used for training by default.
- Claimed cost: Opus 5.5 costs 40% less to run than Opus 5 on typical workloads; Sonnet 5.5 up to 30% less than Sonnet 5.
- Admin controls: Efficient mode (restricts to Sonnet and Haiku models), model entitlements by role, spend limits at organisation, group and member level, and pooled monthly group budgets.
- Organisations using zero data retention for Claude Code, a HIPAA-ready configuration or customer-managed keys stay on their current setup.
How they got it
Vendor announcement with four customer quotes. No benchmark or workload definition behind the cost claims.
Limits
- "Typical workloads" is not defined, so the 40% and 30% figures cannot be checked.
- Customer quotes are selected by the vendor; one claims weeks of work done in "under ten minutes of prompting" with no detail.
- Agentic tasks use more tokens than chat; the post gives no guidance on expected spend.
Why it matters for companies
When chat and agent modes merge, every licensed seat becomes a potential agent running code in the cloud, which changes the cost and risk profile without a new procurement decision. Admins should set role-based model entitlements and group budgets before rollout, not after the first invoice. Egress off by default and credentials outside the sandbox are the right baseline; security teams should confirm which connectors are enabled, since those define what the agent can reach.
Summary based on the full text. Read the original for the full argument and data.Read the original on claude.com
- Tags
- Enterprise AI, AI agents