Cirrius cut its monthly internal LLM API spend across our production agentic systems by nearly 90 percent since Q4 2025. We measured that against a baseline of running every workload through Anthropic frontier APIs. The savings are real, and they came from treating tokens like a scarce resource instead of a flat-rate utility.
On June 1, 2026, GitHub Copilot becomes the first major AI product to retire unlimited usage and move customers to token-based billing. The rest of the industry is forecast to follow within twelve to eighteen months.
You have two ways to meet that shift. Keep using AI like a flat-rate utility and absorb the budget shock. Or run AI like an operator and keep scaling. This article lays out the playbook we use internally, and the playbook the broader market is converging on. The executives we talk to all have the same question. Where do we start.
The Subsidy Era Is Ending
Flat-rate AI pricing was always temporary. The math is finally surfacing.
Agentic workflows consume roughly one thousand times more tokens than single-prompt completions, a figure GitHub disclosed when announcing Copilot’s pricing change. OpenAI is reportedly losing about two dollars for every one dollar earned on inference, according to Microsoft-derived analysis published in Fortune. Forrester forecasts that enterprises will defer 25 percent of planned 2026 AI spend into 2027 as cost discipline catches up with experimentation. Gartner projects global AI spending will hit $2.5 trillion in 2026.
None of this is a vendor problem. The economics of agentic work and the economics of unlimited subscriptions never matched. The bill is being presented now, and June 1 is when the first major product hands it to its customers.
From Consumer to Operator
You have been using AI like Netflix. Starting June 1, you start using it like AWS.
That reframe is the difference between a budget shock and a budget plan. AI consumers buy a flat seat and use it however the vendor allows. AI operators do three things consumers do not.
- They instrument their token spend. Every source of token consumption is made visible.
- They tier their models by task value. Simple work runs on cheap or local models. Hard work earns the frontier price.
- They treat context as a budget. Every prompt, tool definition, and conversation history line is a line item.
The companies that will scale AI through 2027 are not the ones that buy the most tokens. They are the ones that learn how to manage them. The rest of this article is what that looks like.
What the Market Is Telling You to Do
Beyond Cirrius, experienced practitioners are converging on a similar playbook. Nathaniel Whittemore, host of The AI Daily Brief, recently laid out five moves enterprises should make before the pricing reset hits. We agree with all five.
- Find your AI spend leaks. Audit high-volume tasks where retries, bad routing, and over-prompting compound silently. Most token waste is invisible until someone looks for it.
- Hold a model bake-off. Measure cost per completed task, not cost per token. A cheap model that fails twice is more expensive than a smart model that succeeds once.
- Appoint a model sommelier. One owner of model selection, who tracks the market and publishes monthly guidance. Without an owner, every team picks the most expensive model that works.
- Build escape-hatch architecture. Default lanes for routine work, escalation lanes for complex work, human review for low-confidence and high-stakes calls. Stop running every request through the same model.
- Create an AI cost scoreboard. Make spend and quality jointly visible. You cannot trade off two things you cannot see at the same time.
These five are necessary. They are not yet sufficient. The discipline that hits a dramatic savings number is what we layered on top.
How We Cut Our Spend by Nearly 90 Percent
We run our agentic systems on a three-tier model routing pattern. Most of the savings comes from the structure itself.
Tier 1. Open-source models on private infrastructure for the heavy lifting. Roughly 80 percent of our agent work lives here. Repetitive, well-bounded tasks: summarization, classification, simple extraction, internal QA, ambient agent loops that run on a schedule. Self-hosted means effectively zero per-token cost after the infrastructure investment. We have tested several open-source model families including Llama 4, and we expect to keep rotating as the open ecosystem evolves. Latency is higher than frontier APIs. For ambient and batch work, that tradeoff is acceptable.
Tier 2. Haiku and Sonnet for the bulk of frontier work. Customer-facing reasoning, multi-step orchestration, code-aware tasks. These models are fast, capable, and dramatically less expensive than top-of-the-line frontier reasoning. About 80 percent of our frontier spend lives here by design.
Tier 3. Opus only for advanced builds. Complex reasoning, novel architecture work, places where the capability gap is real and worth paying for. Reserved, not the default.
That is the structure. Four mechanics make the structure actually work, and these are the topics we cover in our Cirrius Talk episode on token economics.
- Token flow visibility. Five sources drive consumption: system prompts, conversation history, tool definitions, tool calls, and output. We instrument all five. You cannot optimize what you cannot see.
- Progressive disclosure. Tighter context windows. Tools exposed to the model only when relevant. Most teams hand the model far more context than the task requires.
- Prompt caching. The single biggest lever for ambient or repeatable agents. Applied correctly, it cuts effective input cost by an order of magnitude.
- Tiered routing. The meta-pattern that ties the three tiers together. The router itself becomes a small, cheap model decision.
Tiering plus mechanics is the full system. Neither piece works as well alone.
What Nearly 90 Percent Actually Means
Specifying the math, because executives should ask.
We measure monthly internal LLM API spend across our production agentic systems. The baseline is what those same workloads would cost running every call through Anthropic frontier APIs. The window is Q4 2025 through today. The number does not generalize universally; your savings will depend on how much of your workload is well-bounded enough for Tier 1.
The honest tradeoffs: Tier 1 carries infrastructure cost and operational burden. Latency on local models is higher than frontier APIs. The initial investment is real. The ongoing cost is the win.
Borrow the Expertise. Build the Asset.
Cirrius is not selling tokens. We are selling the capability to manage them.
We run this discipline internally on our own production agentic systems. We have codified the playbook into something a team can execute. An engagement looks like this: assess current token economics and spend leaks; instrument visibility across the five token sources; design a tiering plan that fits your workload; hand the running discipline back to your team.
You do not rent stewardship from Cirrius. You build it as your own asset. The savings number is the surface metric. The capability inside your team is the compounding return.
A handful of executives will get this right early. They will set the cost structure their competitors will be forced to chase.
Where to Start
Two resources to take with you:
- Listen to the Cirrius Talk episode, Agentic AI: Token Economics, Designing Agents That Don’t Blow the Budget.
- See how Cirrius helps clients build this discipline at cirriussolutions.com.
If you want a token economics assessment for your team, reach out. We will tell you where your spend is leaking and what an operator-grade architecture would look like for your workload.
