The most revealing number in the AI industry right now is not a benchmark score. It is a ratio. At Anthropic, spending on compute has reached 2.3 times the company’s entire payroll. A senior engineer costs roughly $224,000 per year in fully loaded compensation; the compute budget that corresponds to each engineer works out to about $515,000 annually. The machines now cost more than the people who use them.
When an AI lab’s compute bill overtakes its salary bill, efficiency stops being a technical preference and becomes a survival constraint. That single shift explains why Anthropic deleted 80 percent of Claude Code’s system prompts, why OpenAI compresses its models’ inner monologues into telegraphic shorthand, and why enterprises that once encouraged unlimited AI usage are suddenly imposing hard caps. The era of burning tokens for spectacle is ending, and the economics of the entire industry are being rewritten in real time.
Claude Code Was Designed to Burn
To understand why Claude Code became so expensive to run, you have to understand what it was actually built to do. Boris, the product’s lead, framed the founding question as: if models become smart enough, what does code look like? The goal was never to help developers economize on tokens. The goal was to showcase how intelligent the model was.
That philosophy shaped everything. Five minutes and $200 of compute is not a malfunction in Claude Code; it is the design working as intended. Every sub-agent spawn, every polished UI animation, every verbose reasoning trace exists to make you watch the screen and think: this model is brilliant. Claude Code is, in a genuine sense, a marketing instrument — Anthropic absorbed enormous token costs to purchase developers’ emotional conviction in the model’s capability, and that conviction converted into subscriptions and enterprise contracts.
The system prompt told the same story. With every tool and feature bundled in, it once swelled to 65,000 tokens; even with most features disabled it still carried 12,000. Before the model wrote a single line of code, it was already hauling an instruction manual on its back. For contrast, Pi, a deliberately lightweight coding agent, boots with fewer than a thousand tokens of context. The gap between those two numbers is the gap between a product that treats context as a stage for demonstration and one that treats it as a budget.
Prompt debt also ages worse than code debt. When code rots, it announces itself — features break, tests fail, bugs surface. When prompts rot, the model just quietly gets worse. Users perceive that “Claude Code seems less sharp lately,” when the real cause is an outdated system prompt that never caught up with a new model. The liability sits on the books invisibly until someone audits it. Anthropic’s 80 percent deletion was not radical; it was an overdue write-down.
The Bill Arrives: Token Apocalypse in the Enterprise
Earlier this year, developers were still bragging about how many tokens they consumed, treating usage as a productivity leaderboard. Within months the mood inverted, and the industry coined a new phrase: Token Apocalypse. The turn was fast for a simple reason — the invoices arrived.
Anthropic’s 2.3x ratio is the most dramatic example, but the pattern runs across the corporate world. Amazon, Adobe, Atlassian, and Citigroup have all imposed strict controls on AI usage: restricting which models employees can access, setting individual quotas, and in some cases revoking permissions entirely. Uber caps every engineer at $1,500 of tokens per month. Walmart recently halted the use of several internal AI tools altogether.
Employees are now receiving deeply contradictory signals. On one channel: AI makes you a hundred times more productive, and you are required to use it. On another: stop burning the company’s money. Neither message is wrong. The contradiction exists because AI tools were shipped without the guardrails that every other enterprise software category eventually grew — cost governance, usage quotas, model tiering, and context management were all deferred in the rush to get capabilities into hands. Enterprises are now doing the unglamorous catch-up work.
Two Philosophies: Burn for Experience, Compress for Margin
The deepest split in the industry runs between two design philosophies. Anthropic burns tokens to buy experience; OpenAI compresses tokens to protect margin. Neither is accidentally on its side of the line.
OpenAI has spent years squeezing tokens out of everything: compressing reasoning traces, optimizing model efficiency, treating “same work, fewer tokens” as a core engineering objective. Codex 5.5 is the cleanest demonstration. GPT-5.5 at medium reasoning effort consumed about 20,000 tokens to post a striking SWE-bench score, while Opus 4.8 spent 50,000 tokens and scored lower. Efficiency was not a side effect; it was the competitive strategy.
Use core new nodes. Need infer. Need add VAE encode for images. Try. Try period.
That fragment comes from leaked OpenAI internal reasoning traces, and it reads nothing like normal English. It reads like compressed engineering shorthand — and that is the point. A model reasoning internally does not need politeness, completeness, or flow. It needs actions, objects, judgments, and next steps. As long as the final answer comes out coherent, the intermediate thinking can be as terse as physics allows.
Compressing reasoning pays compounding dividends in agentic workflows. Agents execute in chains: each step’s thinking becomes the next step’s input. Every token of thought the model saves is not a one-time saving but a reduction that propagates down the entire execution chain. A model that “thinks” 60 percent shorter is, in effect, running a cheaper brain on every subsequent turn of a long task.
Here is the uncomfortable corollary for Anthropic: if it fixed the verbosity problem, its revenue would visibly drop. If developers can accomplish the same work with fewer generated tokens, the difference is billing that never happens. A company whose income scales with token consumption faces a genuine conflict of interest when asked to make its models more concise. OpenAI’s incentive structure points the opposite way, which is precisely why its efficiency lead is likely to persist.
Caveman: When Users Start Rationing Tokens Themselves
The clearest signal that the market has changed came from an unlikely place: a plugin called Caveman that went viral this year. The name is literal — it makes the model talk like a caveman. No pleasantries, no filler grammar, no transition phrases, only the load-bearing content.
Caveman save you token, save you money.
It looks like a joke and works like a lesson. Creator Julius Brussee built it in early April after noticing, during heavy Claude Code use, how much of his token spend was wasted on unnecessary language: greetings, vague hedging, transitional phrases that do nothing in an agent loop. His benchmarks showed Caveman cutting output tokens by 65 to 75 percent compared with default behavior. More telling again: Shayne Sweeney, an engineering director at OpenAI, contributed code to the project. When a rival lab’s engineering leadership helps users of a competitor’s product spend less on it, the efficiency race has officially crossed company lines.
Users reaching for their own token-saving tools means the industry has passed through a phase boundary. There was a period when any cost was acceptable if the experience was magical. That period is over. Tooling that economizes on tokens is no longer a curiosity; it is a procurement category.
What to Do When Intelligence Is Priced by the Token
If you build AI products: stop treating “burn tokens for delight” as the default design philosophy. The next competitive axis is not which model is smarter but how few tokens the same intelligence costs. Run a token audit on your own product: instrument where generation happens, measure what fraction of output is filler, and treat verbosity as a defect rather than a feature. Anthropic deleting 80 percent of its prompts is the opening move of a filtering every AI product will eventually face.
If you are a developer: quit stacking rules into system prompts. Stronger models want lighter instruction. Audit your CLAUDE.md, your agent configuration files, your tool descriptions — separate what is genuinely necessary from what was added “just in case,” because the second category is silently eating your token budget on every single request. The habit that made you effective in 2024 is a cost center in 2026.
If you allocate budget at a company: read the structural signal correctly. Anthropic cutting its own prompts, OpenAI compressing its reasoning traces, and users installing Caveman are three independent data points saying the same thing: the industry has moved from a capability race to an efficiency race. The winner of the next phase will not be the company with the most powerful model but the one delivering the lowest cost per unit of intelligence. Your procurement, your usage policies, and your internal tooling should already reflect that shift — because your competitors’ invoices already do.
