"How many tokens did we spend today, and what's the ROI?" It's a question C-suite leaders are starting to ask about their AI strategy. A year ago it would have sounded like engineering trivia. Today it belongs in the boardroom, right next to headcount and cloud spend.
**Short answer: **Tokenmaxxing is when a company burns through AI tokens, and budget, because nobody designed a consumption strategy. You fix it with six steps: audit consumption, measure consumption ROI instead of raw usage, go LLM agnostic and match the model to the task, teach teams to prompt well, cache repeated content, and invest in the harness, the orchestration layer that quietly controls your AI economics. ** What did Meta's AI usage memo reveal? ** Meta sent an internal memo to 6,000 employees warning that internal AI usage costs are approaching billions of dollars in 2026. In roughly 30 days, employees consumed 73.7 trillion tokens, tracked on an internal leaderboard nicknamed "Claudeonomics." CTO Andrew Bosworth put it bluntly: "All motion is not progress and token usage alone is not a measure of impact of any kind." Meta's response was a centralized AI Gateway dashboard and formal token budgets starting in 2027.
Read that again. One of the most sophisticated engineering organizations on the planet is only now building the guardrails. If Meta is struggling with tokenmaxxing, what chance does the average enterprise have? ** What is tokenmaxxing?**
Tokenmaxxing is what happens when nobody designs a consumption strategy. Every prompt goes to the biggest, most expensive model. Every context window gets stuffed to the brim. Every agent re-reads the same documents on every run. The result: AI bills that grow faster than AI value.
How do you fix tokenmaxxing? 6 steps to control AI token costs
Here are the six steps we use at Gaspar AI to keep token consumption under control, for ourselves and for our clients. **
- Audit before you optimize**
You can't fix what you can't see. Break down consumption by use case, by team, by agent. Most companies discover that 20% of their workflows drive 80% of their spend, and half of that is waste. Meta needed a leaderboard to see the problem. You need a dashboard before the bill arrives, not after. ** 2. Measure consumption ROI, not just consumption**
This is exactly Bosworth's point: token usage alone is not a measure of impact. Tie every token stream to a business outcome. Cost per resolved ticket. Cost per processed case file. Cost per qualified lead. Once you price outcomes instead of usage, the decisions make themselves: workflows with strong ROI get more budget, workflows that burn tokens without moving a metric get redesigned or killed. A leaderboard measures activity. An ROI model measures value.
3. Go LLM agnostic and match the model to the task
This is the single biggest lever. You don't need a frontier model to classify a ticket or extract a date from an invoice. Distinguish between open source and proprietary LLMs: open source models (Llama, Qwen, Mistral Small) can run on your own infrastructure with predictable costs and full data control, while proprietary frontier models earn their premium on complex, reasoning-heavy work. There is also a human factor most strategies ignore: not every employee in the enterprise has appetite for one single LLM. Developers, analysts, and support teams have different preferences and different workflows, and forcing everyone through the same model creates shadow AI usage you can't see or budget for. An LLM agnostic architecture solves both problems. You swap models per use case, give teams flexibility within governed limits, and negotiate with every vendor from a position of strength. ** 4. Teach your teams to prompt like AI gurus**
Bad prompts are expensive prompts. Vague instructions trigger long, rambling outputs and multiple retries. Best practices that pay for themselves immediately: be specific about the output format, set length limits, provide examples instead of explanations, and strip irrelevant context before sending. A one-hour prompting workshop can cut token spend by double digits.
5. Cache, don't repeat
If your agents keep processing the same knowledge base, policy document, or product catalog, you are paying for the same tokens again and again. Prompt caching, embeddings, and a knowledge graph layer mean you retrieve exactly what is relevant, not the whole library, every time.
6. Invest in the harness, the secret weapon
Models get the headlines, but the harness (the orchestration layer around the model) determines your economics. A well-designed harness decides which model handles which step, how much context each call actually needs, when a human should validate instead of another expensive LLM loop, and when to stop an agent that is spinning. Two companies can use the exact same model and see a 10x difference in cost, purely because of the harness. ** The takeaway**
AI cost control is not about using AI less. It is about using it deliberately. So the next time someone in the boardroom asks "how many tokens did we spend today, and what's the ROI?", you should have an answer. If your AI bill is growing faster than your AI results, you don't have a model problem. You have a strategy problem, and it is fixable. You don't have to wait until 2027 to fix it.
**Frequently asked questions ** What is tokenmaxxing? Tokenmaxxing is when an organization burns through AI tokens, and budget, because nobody designed a consumption strategy. Every prompt goes to the biggest, most expensive model, every context window is stuffed to the brim, and every agent re-reads the same documents on every run. The result is AI bills that grow faster than AI value.
How can enterprises reduce their AI token costs? Enterprises can reduce AI token costs in six steps: audit consumption by use case, team, and agent; measure consumption ROI rather than raw usage; go LLM agnostic and match the model to the task; train teams in prompting best practices; cache repeated content using prompt caching, embeddings, and a knowledge graph layer; and invest in the harness, the orchestration layer that decides which model handles each step and how much context each call needs.
What is consumption ROI in AI? Consumption ROI ties every stream of token spend to a measurable business outcome, such as cost per resolved ticket, cost per processed case file, or cost per qualified lead. Instead of measuring how many tokens are used, it measures the value those tokens produce, so strong workflows get more budget and wasteful ones get redesigned or removed.
Why does an LLM agnostic approach lower AI costs? An LLM agnostic approach lets you route each task to the most cost-effective model instead of sending everything to one expensive frontier model. Simple tasks go to smaller open source models such as Llama, Qwen, or Mistral Small, often on your own infrastructure, while proprietary frontier models are reserved for complex reasoning. It also accommodates different team preferences, reducing shadow AI usage, and strengthens your negotiating position with vendors.
What is a harness in AI systems and why does it matter for cost? A harness is the orchestration layer around the model. It decides which model handles which step, how much context each call needs, when a human should validate instead of running another expensive LLM loop, and when to stop an agent that is spinning. Two companies can use the exact same model and see a 10x difference in cost, purely because of the harness.