Enterprise AI & Work
North 2's Cost Controls Need a Model-Aware Budget
Cohere's North 2 adds agent cost controls, but a 66.7x public API price spread shows why token caps need model-aware budgets.
Cohere’s October 5 North 2 launch puts agent budgets inside the enterprise workspace, but buyers should test what those budgets actually measure. Cohere’s own public model prices span 66.7x per output token—a useful warning that equal token allowances do not imply equal spending, even before a North contract enters the calculation.
A token allowance is not a dollar allowance
Cohere’s North 2 announcement combines a redesigned agent orchestration system with reusable skills, shared libraries, persistent memory, and applications created through chat. North Admin adds user quotas, rate limits, organization-wide caps, and consumption visibility. The platform also supports customer-selected models and private deployment. The commercial promise is coherent: make useful agents available across a company while giving the company control over their behavior and consumption.
The buying question sits underneath that promise. A finance team wants a monetary ceiling. A platform team may configure requests or tokens. An application owner wants completed work. Those are different units, and a procurement review must establish how the product translates among them.
Cohere lists Command A at $10 per million output tokens and Command R7B at $0.15 per million output tokens. Dividing $10 by $0.15 produces 66.7x, rounded to one decimal place. The input prices show the same spread: $2.50 divided by $0.0375 also equals about 66.7. Both calculations use current public API list prices retrieved for this edition, not negotiated North rates.
This comparison establishes a budgeting problem, not a model recommendation. It says nothing about equivalent answer quality, successful task completion, private deployment economics, or which models a particular North customer has enabled. A cheaper model that requires more attempts may not be cheaper per finished job. Nevertheless, a policy expressed only as a common token allowance cannot be interpreted as a common dollar allowance across differently priced models.
That distinction matters because North is sold through custom enterprise pricing. The public API ratio cannot become a forecast of a customer’s North bill. It is an audit prompt: ask which expenses the cap covers, which price schedule it uses, and whether changing the model changes the permitted consumption. An enterprise quote should identify platform licensing, inference, capacity commitments, and additional services clearly enough to reconcile usage with invoices.
The infrastructure option creates a second translation problem. Standard Vault’s pricing documentation distinguishes fixed instance commitments from a flexible arrangement with baseline capacity and autoscaling; additional flexible capacity is billed by instance-hour. Limiting token use on provisioned capacity may restrain load without reducing an existing commitment. Buyers should therefore separate preventing consumption spikes from reducing expenditure. A control can succeed at the first while leaving the second unchanged.
This is the same procurement discipline raised by today’s Mistral preview exit-clock analysis: the measurable unit in a product announcement has to connect to the decision in the contract. A quota dashboard is evidence of a management interface. A budget guarantee requires an agreed financial boundary and observed enforcement.
Prove the boundary before expanding autonomy
The most consequential qualification is in Cohere’s current Flow Control documentation. It labels that feature Alpha and advises development evaluation rather than production use. The documentation says interface settings have no effect until the enforcement services are enabled through the deployment configuration. It also describes an initially unlimited default tier. These are properties of the documented implementation; they do not prove that every hosted North 2 offering lacks production spending controls. Buyers need Cohere to identify which version and implementation their contract actually includes.
The same page describes independent request, input-token, and output-token limits, with rules that can distinguish models and providers. It also exposes a choice to allow or deny requests if the rate-limiting service fails. Those details make the acceptance test more concrete. Configuring a number is insufficient: the operator must observe the right request being stopped, under the intended identity and model, when the relevant limit is reached.
Our earlier analysis of Snowflake’s agent governance gateway made a related distinction between a control-plane proposal and dependable operational coverage. North now warrants the same treatment. Central management can make budgets legible, but coverage must be demonstrated for the traffic the enterprise actually runs.
Existing North customers with a specific workflow ready for broader deployment should move first into a bounded evaluation. Their useful experiment is to replay representative work under the proposed policy and compare accepted results, consumption records, invoice treatment, and requests rejected at the limit. Record model selection alongside each result. Otherwise, an apparent spending improvement may merely reflect an unnoticed change in answer quality or workload composition.
Prospective buyers should make that evaluation part of procurement. The implementation cost includes mapping identities and departments to policies, reconciling the billing units, and assigning responsibility when useful work is stopped. The finance owner should approve the financial definition; the application owner should approve the interruption behavior. Neither decision belongs solely to the administrator who enters the quota.
The strongest counterpoint is that a hard monetary ceiling can obstruct valuable work. Blocking an urgent workflow may cost more than the inference it saves. That does not invalidate controls; it means operators need an explicit exception process, an accountable approver, and separate treatment for workflows with different consequences. A single company-wide ceiling is a blunt instrument unless the operating procedure around it is equally clear.
Evidence would change the verdict quickly. A customer-specific production specification, reconciled invoices, and observed enforcement across the chosen models would support expansion. Unexplained usage, mismatched billing units, or a deployment whose controls remain experimental would support delay. North 2 deserves an evaluation from enterprises that need private, governed agents. It earns a larger rollout when its budget means the same thing to the software, the finance team, and the person responsible for the interrupted job.