Cohere North 2 adds token spending caps and air-gap support
Cohere North 2 introduces hard token spending caps, cross-session memory, and air-gapped deployment options for enterprise AI agents.
The enterprise platform brings hard limits on token usage, persistent session memory, and support for servers cut off from the internet.
On October 5, 2026, Cohere announced the release of Cohere North 2, an overhaul of its enterprise software aimed at running and governing autonomous agents. According to company disclosures and reporting by VentureBeat and SiliconANGLE, the release includes more than 15 features designed to address two persistent operational barriers: unexpected cloud bills triggered by runaway inference loops and corporate policies preventing sensitive data from leaving private infrastructure.
What Cohere North 2 does to prevent budget overruns
Deploying agents on multistep workflows often creates financial surprises when automated routines get stuck in repetitive queries. In North 2, the North Admin console tracks token consumption down to specific teams, users, and individual agents.
Engineering leads can enforce organizational spending ceilings, set per-user token quotas, and configure rate caps on requests per minute. The system issues preventative alerts before teams breach their allotted budget. The orchestration layer also keeps human-in-the-loop checkpoints in place, requiring manual sign-offs before agents can trigger high-risk actions.
Persistent memory and model flexibility
When Cohere introduced the first version of North in August 2025, agents started from a clean slate every time a user opened a new chat. North 2 changes that behavior by preserving context, working state, and session history across interactions.
Workflows now separate into two core building blocks: Skills (modular capabilities that different departments can reuse) and Libraries (shared reference document collections). The platform also adopts a Bring Your Own Model (BYOM) design: while optimized for Cohere’s Command family, it connects to proprietary third-party models brought by enterprise customers to reduce lock-in risks. On the interface side, North 2 adds a drag-and-drop workflow builder alongside chat-based prototyping tools that turn plain-language prompts into slides, dashboards, and lightweight micro-apps.
Isolated networks, security, and integrations
For businesses facing strict regulatory rules in sectors like finance, healthcare, and defense, North 2 supports deployments inside virtual private clouds (VPCs), on-premises server racks, and air-gapped environments disconnected entirely from the public internet.
Cohere reports compliance with SOC 2 Type 2, ISO 27001, and ISO 42001, supported by built-in filters that detect personally identifiable information (PII) leaks and block prompt injection attempts. The release comes with pre-built connectors for workplace tools such as Slack, Microsoft SharePoint, Outlook, OneDrive, GitHub, Jira, and Linear, as well as support for the open Model Context Protocol (MCP). To handle hardware throughput, Cohere collaborated with NVIDIA to tune performance across clusters running Blackwell and Hopper GPUs.
Closed pricing and roadmap gaps
Cohere confirmed production deployments with enterprise customers including South Korea’s LG CNS and Canada’s Bell Cyber. Several critical commercial questions, however, remain unanswered. The company has not published fixed pricing or per-seat costs, leaving platform access tied to private enterprise sales negotiations.
Furthermore, direct connectors to financial market data providers like FactSet, PitchBook, S&P Global, and Crunchbase appear only as future roadmap targets without shipping dates, keeping them classified as RUMOR items for now. There are also no public, independent benchmarks evaluating the rebuilt orchestrator against open-source alternatives like LangGraph or AutoGen. North 2 delivers the operational controls that enterprise IT teams have been requesting, but the true return on investment will depend on how reliably these guardrails hold up under daily production workloads.