The FinOps Foundation recently added a sister organization: the Tokenomics Foundation.
It does not address the token in the blockchain context, but as the “atomic unit” of generative AI. Its aim: contribute to the emergence of standards and best practices to maximize value creation… and to control consumption.
A foundation of techniques for consuming fewer tokens
We’re not starting from zero. The FinOps Foundation already has a Token Economics working group. This group notably conducted a community survey that helped establish a hierarchy of cost-optimization levers.
When considered individually, the biggest savings potential lies in selecting a model that matches the task at hand. The intelligent routing frameworks can cut the cost per request by 60 to 80%, the FinOps Foundation adds.
| Lever | Potential savings | Implementation effort |
| Rightsizing of the model | 60 to 90 % | Medium |
| Batch processing (API batch) | 50 % | Low to medium |
| Prompt caching | 50 to 90 % on cached tokens | Low |
| Context window management | 20 to 60 % | Medium to high |
| Control of output length | 10 to 40 % | Low to medium |
| Volume/engagement discounts | 10 to 30 % | Low (purchases) |
Context window management includes, among other things, sliding-window techniques, conversation summarization, and compression of tool outputs. Output-length control can be achieved by using structured response schemas.
A three-step FinOps strategy
Among all the ways to leverage AI capabilities, direct API access (OpenAI, Anthropic, etc.) is identified by the FinOps community as the most challenging category to manage. Billing is even more opaque because it does not natively incorporate concepts like business units, cost centers, or applications. Frequent changes in model catalogs don’t help. And predicting usage peaks, as well as capping them, remains difficult.
The FinOps Foundation proposes that organizations implement a three-phase strategy:
- Months 1 to 3: inventory of expenditures, establishment of API key governance, light tagging, production of a basic dashboard, per-account budget alerts
- Months 3 to 9: attribution and showback by model, review of model rightsizing, integration of batch and prompt caching, optimization of the context window, activation of anomaly detection
- Beyond: chargeback, dynamic routing, negotiation of volume discounts, integration of cost estimation into CI/CD, consolidation of cost-metrics reporting with those of other technologies
Native tools potentially sufficient… in a single-vendor environment
Governing API keys means associating each with a team, an application, or a use case, with an owner and a clearly designated cost center. The attribution features (application inference profiles on AWS Bedrock, project-level keys on OpenAI, and workspace-level keys on Anthropic…) have advanced, the FinOps Foundation notes. They are generally sufficient for organizations that operate with a single vendor. Those with multiple vendors will rely on feature-level attribution, policy enforcement per request, or gateways such as Helicone, LiteLLM, and Portkey.
Organizations that have built a FinOps team typically entrust it with AI cost governance. It sometimes falls to the platform team or to an AI center of excellence. In all cases, it involves framing the scope within which engineering teams can operate autonomously: lists of approved models, maximum context length per use case, data-classification rules for external providers, architecture review requirements for agentic workflows, etc.
Prices are indeed continuing to fall, but more slowly than in the past, the FinOps Foundation notes. And the trend mainly affects the cheaper models. Not those used for reasoning and agentive tasks… which consume 5 to 40 times more tokens per task.
