CISA and Canadian partners release guidance on AI token consumption

CISA and the Canadian Centre for Cyber Security issued joint guidance called Managing AI Token Consumption for Operational Resilience on October 22, 2026, to help organizations manage token use from third-party superintelligence services and keep critical operations running.

The document helps organizations understand and manage how they use tokens to support operations. CISA, the Canadian Centre for Cyber Security, and partners issued the document jointly on October 22, 2026. The release targets organizations that depend on external AI models for core business functions.

When token consumption is not managed, it can push costs up unpredictably, cut availability, degrade performance, exhaust tokens, and enable denial-of-tokens or denial-of-wallet attacks. These outcomes can undermine operational resilience for organizations that depend on external AI systems. The guidance ties token management to continuity for critical operations.

Critical infrastructure owners and operators can use the document to spot mission-critical functions that rely on superintelligence. They can also map service provider and operational dependencies and prepare for lower superintelligence availability during outages, incidents, or periods of high demand. The document targets critical infrastructure owners and operators as well as cyber defenders. It helps them anticipate how an outage or demand spike could affect token-dependent services.

To address these risks, the guidance outlines steps for allocating tokens ahead of time, building efficient workflows, and helping users write better prompts. It also recommends using caching to reuse context, optimizing data before submission, and tracking token consumption over time. These measures help organizations match token use to operational priorities and maintain service continuity. The steps are practical for teams managing external AI services.

Cyber defenders can use the document to oversee superintelligence-enabled cyber defense workflows while preserving human oversight, governance, and strong cybersecurity fundamentals. The recommendations apply to automated security tasks that rely on external models. Human oversight remains central to the approach, and governance ensures accountability.

Tokens are the measurable units that determine how much computational effort a task requires, such as generating responses or analyzing inputs. They split into input tokens, which are the text a user provides, and output tokens, which the AI generates in response. Producing output tokens requires more computational work than processing input tokens. The count includes both the text sent to the model and the text the model generates.

Since AI models are stateless, users must send the full conversation history with each interaction, which drives up token usage. Usage rises with longer prompts, detailed conversation history, and larger, more complex models. Stateless operation means the model does not retain memory between requests, so every exchange repeats prior context.

Teams can cut token usage by writing concise prompts and caching context for reuse. Organizations can also set response length limits and select the right AI model for the task. Storing reusable context, like system prompts or tool definitions, can reduce the cost of repeated computation. Dynamic toolbox optimization sends only the tools needed for each request, which cuts input token costs.

Microsoft Foundry’s toolset includes Foundry Playground, Model-Router, Agent-Optimizer, and Foundry IQ for optimizing AI systems and managing token usage. The tools cover testing, model selection, agent tuning, and related optimization tasks. Organizations can use them to align model choice with token budgets and avoid surprise bills.

TechCrunch reported that per-token prices have fallen even as AI adoption and autonomous agents have pushed token consumption higher. Companies that bought all-you-can-eat subscriptions in early 2025 are working to understand spending, cut costs, and protect return on investment.

New models released in November, including Anthropic’s Claude Opus 4.5, OpenAI’s GPT-5.1, and Google’s Gemini 3 Pro, improved agentic tools and multiplied consumption even as per-token prices fell. The agentic capabilities caused usage to rise even when individual tokens became cheaper.

Goldman Sachs projects that global token usage will grow 24 times by 2030.

Uber used up its full 2026 AI coding budget by April, and Microsoft took back its developers’ Claude Code licenses months after enabling them. A Priceline employee said a routine Cursor contract renewal cost 4-5 times as much. One company reportedly received a $500 million Claude bill because it did not set usage limits for employees. Vitaly Gordon of Faros AI said a chief technology officer described an engineer who spent $40,000 on tokens in a single month.

J.R. Storment of the FinOps Foundation said companies reported running 3x over their full 2026 token budget by April. Faros AI published a two-year study of 20,000 developers in April that found output was increasing even as bugs and rewrites also rose. Jellyfish found that engineers using the most tokens were roughly twice as productive as lighter users while spending 10 times the tokens. Nicholas Arcolano of Jellyfish said per-developer AI consumption climbed about 18.6 times over nine months because of agentic features.

Jellyfish, Waydev, and Faros AI sell AI agent monitoring that helps companies prove the return on investment of developer tools. Ramp entered AI spend management, and Datadog and New Relic added cloud cost management, token-level observability, and GPU monitoring. Pay-i monitors and improves the costs and performance of GenAI investments, while Paid lets developers track costs, measure usage, and bill users based on actual value instead of subscription fees. Factory, a startup that builds AI agents for enterprises, released a model router that chooses the right model for each task automatically.

The Linux Foundation announced plans for the Tokenomics Foundation, a standards body for AI token cost discipline like FinOps for cloud spend. It develops a canonical definition and framework for tokenomics and open standards covering AI token usage and billing. It also plans metrics like cost-per-intelligence or tokens-per-watt, along with measures of token factory effectiveness and consumption efficiency. The Tokenomics Foundation plans to launch formally in July and announce more members at the FinOps X conference, though its first deliverable is still months away.

Amazon Web Services is expected to roll out new financial management features for enterprise AI spending at the FinOps X conference. Tiffany Luck, a partner at NEA, expects to build token efficiency and observability into the harness or app layer. Vitaly Gordon expects frontier labs and model providers to use OpenRouter-style optimization to route queries to the lowest-cost models. Nishant Gupta, chief availability officer at Salesforce, said token economics calls for a different operational muscle than the industry developed for cloud.

Chris Reed, senior director of IT finance at Priceline, found gaps between a vendor’s reported usage and Priceline’s internal data. Reed likened AI adoption to the crack-cocaine epidemic, and Priceline began placing token limits on certain groups.

J.R. Storment described tracking token costs as a trillions-of-rows-a-month data problem, compared with hundreds-of-millions-of-rows-a-month for cloud costs. According to Storment, most of the 180 vendors in the FinOps Foundation are moving toward AI agent monitoring and spend management.

Alexander Embiricos, OpenAI’s head of enterprise, said customer conversations moved from capability to spending visibility, auditability, token controls, and model efficiency. He added, “Our conversations are never about that now.”