Cohere North 2: Enterprise Agent Platform Gets Cost Controls
October 5, 2026. Cohere launched North 2 today, the most significant update to its enterprise AI agent platform since North reached general availability in August 2025. The new version targets two problems that have become persistent headaches for every enterprise AI team in 2026: runaway token costs and agents that reset their memory when a session ends.
North 2 introduces persistent memory across sessions, a redesigned multi-step orchestration harness, shared skill libraries, and per-user token spending caps. It still runs anywhere the data lives, from Cohere’s cloud to a fully air-gapped installation with no outside connection. And it accepts any AI model, not just Cohere’s own.
“We actually derive a sizable portion of our revenue from the harness,” Joelle Pineau, Cohere’s chief AI officer, told the Globe and Mail at launch. That framing matters. Cohere is not selling access to its models alone; it is selling the infrastructure layer that sits around any model a company chooses to deploy.
What Changed in North 2
North’s first version, launched in January 2025 with RBC as its flagship customer, behaved largely as a retrieval-augmented chatbot with some basic workflow automation. North 2 is a different product category.
The key additions, confirmed across multiple sources including VentureBeat and SiliconAngle, are:
Persistent memory. Agents now carry context across sessions. Instead of starting cold every time a workflow opens, an agent can recall prior state, prior decisions, and prior outputs. Paul Teyssier, VP of Product AI at Cohere, confirmed to VentureBeat that memory persists across sub-sessions and multiple agents, though Cohere has not published a specific retention window.
Redesigned orchestration harness. The harness manages complex, multi-step tasks autonomously and routes designated decisions to a human for sign-off. Users describe a workflow in natural language, and North 2 generates a runnable agent. From that same natural-language prompt, it can produce a slide deck, dashboard, or lightweight app that prototypes directly inside the chat window.
Skills and shared agents. Reusable capabilities are packaged as skills that agents call rather than rebuild. Teams create agents once and share them across the organization, turning repeated institutional knowledge into a shared internal library rather than siloed individual configurations.
Libraries. Shared document stores and asset libraries let every agent draw on the same knowledge base. Library access is managed by role, with new Library Access Manager and compliance-focused positions added in this release. Library API access is currently in beta.
Token spending controls. This is the addition that will get the most immediate attention from enterprise AI leaders facing budget pressure in 2026. North Admin now tracks usage down to the individual user and agent, sets consumption tiers by request and token rate, and sends alerts before limits are reached. Usage data can be exported to external analytics systems, broken out by model, user, and department.
North 2 Deployment Options
| Deployment mode | Description | Minimum compute |
|---|---|---|
| Cohere Cloud | Managed SaaS on Cohere infrastructure | Not applicable |
| Customer VPC | Runs inside customer’s virtual private cloud | Varies |
| On-premises | Self-hosted in customer data center | 2 GPUs minimum |
| Air-gapped | No external network connection | 2 GPUs minimum |
The air-gapped and two-GPU deployment options remain Cohere’s most distinctive competitive position. Most enterprise AI offerings assume customers are comfortable routing data through a vendor’s cloud. Regulated organizations in financial services, healthcare, defense, and government are not. North was built specifically for those environments, and North 2 extends that promise to the new agentic capabilities.
The Cost Problem North 2 Is Solving
AI agent costs compound quickly in ways that chatbot costs did not. Unlike a prompt-response exchange with a defined token count, a long-running agent that iterates, backtracks, and re-queries tools can consume tokens unpredictably. That unpredictability has become what Teyssier described to The Logic as “probably the number one rising concern for AI leaders.”
The problem is compounded by the push to deploy agents at organizational scale in 2026. Teams that shipped their first pilots in 2025 hit real production budgets this year, and many discovered that their token spending assumptions from the pilot phase did not survive contact with full organizational usage.
North 2 addresses this at the administrative layer rather than at the model layer. Rather than waiting for a model provider to ship cheaper tokens, enterprises can now set a budget ceiling before the spending happens. Automatic downgrade to cheaper models for simpler tasks is on Cohere’s roadmap but is not in this release.
Enera’s analysis of the enterprise AI pricing revolt covers how tokenomics conflicts have reshaped procurement decisions across enterprise AI teams throughout 2026. The spending cap features in North 2 are a direct product response to those pressures.
Model-Agnostic Positioning
North 2 accepts models from Anthropic, OpenAI, Google, DeepSeek, and Alibaba’s Qwen family alongside Cohere’s own Command and Embed families. Administrators assign which model each team or workflow uses. Teams can bring their own model contracts without rebuilding agent configurations.
This positioning puts Cohere in a different lane from OpenAI’s Workspace Agents and Anthropic’s Claude Code, both of which are tightly coupled to their own model families. The tradeoff is that North 2 gives up some model-specific depth for governance breadth. It cannot offer the deep Codex integration OpenAI delivers inside ChatGPT Enterprise, but it can manage a mixed fleet of models behind a consistent security and cost control layer.
For enterprise teams already running multi-model strategies, that consistency has measurable value. The risk is consolidation: as AI infrastructure vendors acquire one another and tighten integrations, a model-agnostic harness can face pressure to prioritize its own models over time.
North 2 launches in an unusual strategic moment on that front. Cohere and German sovereign AI company Aleph Alpha signed a merger agreement in September 2026. The combined entity keeps the Cohere name and positions itself as a transatlantic sovereign AI company. Aleph Alpha’s Kolibri model, which released its full weights under Apache 2.0 on October 3, is a natural addition to North’s model library for European enterprise customers with data residency requirements.
Nvidia Partnership and Throughput
Kari Briski, VP of Generative AI at Nvidia, confirmed at launch that North 2’s models produce more tokens per second per node on Blackwell and Hopper GPUs. Higher throughput reduces cost per token for enterprises running North on their own infrastructure. A faster model that serves the same number of requests at the same accuracy level simply costs less to operate per completed task.
Cohere does not publish North’s pricing publicly. The company prices it based on deployment scale, complexity, infrastructure requirements, usage patterns, support, and customization. That structure targets large enterprises with defined procurement cycles, not individual developers or small teams.
Governance and Certifications
North 2 ships with SOC 2 Type 2, ISO 27001, and ISO 42001 certifications. ISO 42001 covers AI management systems specifically. That matters for regulated industries where auditors are beginning to ask for documentation of AI governance frameworks alongside the usual information security audits.
Security controls sit in North Admin and apply at the individual agent level. Guardrails scan prompts and responses for personally identifiable information and detect prompt injection attempts. Autonomy policies limit agents to authorized actions and escalate critical decisions to a human. Detailed logs of all agent changes and activity are available for compliance reviews.
Enera’s post on the Snowflake Cortex AI Gateway covers how data platforms are building complementary governance layers for enterprise AI agents. The governance challenges are the same; the approaches differ.
Connectors at launch cover Slack, GitHub, SharePoint, and OneDrive. Cohere has announced planned connectors to PitchBook, S&P Global, and FactSet, signaling a push into financial services workflows where structured data access is central to agent utility.
What Enterprise Teams Should Take From This
North 2 is not trying to be the smartest platform. It is trying to be the most governable one. Memory, spending controls, shared libraries, and model portability are all tools for scaling AI agent deployments inside organizations that have already decided to deploy agents and now need to manage that deployment without losing control of cost, data, or behavior.
That is a different problem than the one AI labs are focused on, and it is the problem that enterprise AI leaders in finance, healthcare, government, and regulated manufacturing are dealing with every quarter.
The announcement is also a signal about where enterprise AI infrastructure competition is heading. In 2025, the race was to ship agents. In 2026, the race is to make those agents governable, affordable, and persistent enough to be genuinely useful across an organization. Cohere’s North 2 is a direct bet on that shift.
For teams evaluating how autonomous systems fit into their go-to-market infrastructure, Enera’s evergreen post on autonomous GTM systems covers the broader architectural decisions involved.
Sources: VentureBeat | SiliconAngle | The Globe and Mail | The Logic | CryptoBriefing