Tokenmaxxing: The Bill Nobody Showed You

Companies laid off workers to save money. Then the AI bill arrived — and it was higher than the salaries they had just cut. This is the story of tokenmaxxing: what it is, what it revealed, and what it means for companies that have not caught up yet.

Key Takeaways

The most powerful AI tools are expensive — expensive enough that, for most jobs, keeping a human is still cheaper. Tokenmaxxing — measuring how many AI operations your team triggers instead of how good the results are — burned through budgets and exposed an uncomfortable truth: most companies do not know what good AI use actually looks like. For Germany, which is still catching up on AI adoption, there is a rare chance to skip the expensive mistakes — but only if companies know what to avoid.

The Leaderboard Nobody Questioned

Meta (the company behind Facebook and Instagram) built an internal competition called Claudeonomics. All 85,000 employees could see a live ranking of who was consuming the most AI. Top users earned titles like “Token Legend” and “Session Immortal.” The bottom of the list was visible too.

Think about what that leaderboard actually rewarded. Not the best output. Not the most useful work. The highest token consumption. An employee who ran the same prompt ten times appeared more productive than one who asked once, got it right, and moved on. The incentive was not to work well with AI — it was to use as much AI as possible, as visibly as possible, to avoid appearing at the bottom of a company-wide ranking seen by 85,000 colleagues.

In one month, Meta consumed between 60 and 73 trillion tokens. The estimated cost ran into the hundreds of millions per month, heading toward billions annually. Nobody in the programme was asking whether those tokens produced anything worth that price. The metric was volume. Volume was climbing. The programme was, by its own measure, a success.

The leaderboard was shut down 48 hours after the numbers leaked to the press. Not because the costs were wrong. Because once the numbers were public, the question became unavoidable: what exactly did all of that buy?

Around the same time, Uber rolled out an AI coding assistant to around 5,000 engineers. No leaderboard — just a tool that billed by usage. Average monthly costs ran $150–$250 per engineer, with heavy users reaching $500 to $2,000 per month. By April 2026, Uber had spent its entire AI budget for the year — with eight months still to go. No competition required. The wrong incentive structure produces the same result either way.

What Tokenmaxxing Actually Is

A token is the basic unit AI systems use to process text — roughly three-quarters of a word. Every message you send and every answer you receive consumes tokens. Automated AI tasks running in the background (so-called agentic workflows, where the AI works through a series of steps on its own) can consume enormous numbers of tokens without anyone noticing — or approving the spend.

Tokenmaxxing is what happens when companies start treating token volume as a sign of productivity. The more AI operations your team triggers, the more “efficient” they appear. This sounds like a small measurement error. It is not. When volume becomes the main goal (the KPI — Key Performance Indicator), people stop asking whether the output was actually good and start optimising to use more AI, whether it helps or not.

The Bill Nobody Showed You

Not all AI models cost the same. The price gap between mainstream models is up to 300× — from $0.10 per million tokens (cheap, fast models like Gemini Flash-Lite or GPT-4.1 nano) to $30 per million output tokens (powerful flagship models like GPT-5.6). The most capable models — so-called frontier models (the best-performing AI systems available) — like Claude Opus 4.8 sit at $5 input / $25 output per million tokens.

Reasoning models (AI systems that think through a problem step by step before answering) add a hidden layer of cost: those internal thinking steps are billed even though you never see them. A short response can quietly generate ten times more tokens behind the scenes.

Per-token prices have dropped around 98% since 2022 — but enterprise AI bills have tripled over the same period. Lower unit prices were swallowed by far higher consumption, driven by automated workflows that run continuously in the background.

Cost Map: Tokens, Humans & Design Systems

The following overview shows how AI tool costs compare to human salaries in design roles, what happens when a design system is missing, and how Figma’s role has shifted as a result.

When AI Costs More Than the Workers It Replaced

Forbes, July 2, 2026: “AI Costs More Than The People It Replaced.” That was not a provocation. It was the market correcting a calculation error.

The original model looked obvious: replace a €85,000 salary with a €1,800 AI subscription — a 97% cost reduction on paper. The model had five rows missing. Oversight time for checking AI output (25–35% of capacity still required). Error correction — AI mistakes look plausible and take 2–3× longer to catch than obvious ones. Agentic token escalation — automated AI workflows use 10–100× more tokens than projected, as Microsoft discovered when it told engineers to stop using an AI coding tool entirely. Institutional knowledge that leaves with the person and cannot be reconstructed. And a rehire premium when companies reverse course — recruiter fees plus roughly six months to get someone back up to speed.

MIT’s Project Iceberg (November 2025, Massachusetts Institute of Technology) put a precise boundary on where the math actually works: AI is cost-competitive in roles covering 11.7% of the US workforce. For the other 88%, the evidence from 2025–2026 shows companies are now paying to find that out.

The Boomerang: Why the Layoff Logic Is Reversing

Klarna (a Swedish payments company) announced in 2024 that its AI agent could do the work of 700 customer service employees. It made international headlines. Then, quietly, Klarna started hiring customer service staff back — no press release, no correction to the original headline.

Ford rehired 350 veteran engineers to fix quality problems that automated systems could not catch. Their conclusion: giving AI design requirements does not produce a quality product without experienced human oversight.

These are not exceptions. They are the pattern. IBM replaced large parts of its HR department with AI that handled 94% of routine requests without issues. The remaining 6% — ethical decisions, complex escalations, judgment calls — broke the process. IBM subsequently announced plans to triple its US entry-level hiring.

55% of executives now regret replacing human workers with AI. 73% of organisations that made AI-driven cuts failed to improve their financial results. 29% have already rehired for the exact positions they eliminated.

For most organisations, the assumption that AI makes people obsolete has already been tested — and failed.

Token Loyalty: The New Corporate Obedience

When volume becomes the goal, a new kind of employee behaviour emerges: token loyalty. Not loyalty to the work. Loyalty to the metric. People optimise for what is measured, regardless of whether it produces anything useful. The person at the top of the leaderboard did not necessarily do the best work — they most completely handed their thinking to the machine.

This is the deeper problem behind the leaderboard: a company with 85,000 employees could not define what good AI use looks like. So it measured the only thing it could count — and built a culture around it.

The Skill Nobody Is Protecting

When people stop doing parts of their job and hand them to AI, those skills do not stay sharp. The ability to evaluate whether an output is actually correct — not just plausible-sounding — comes from practice. Tokenmaxxing optimises for volume. It does not reward that judgment.

The long-term risk: employees who outsourced the most are now least equipped to know when AI gets it wrong. They delegated before they developed the ability to evaluate. Now they cannot reliably tell the difference between a correct answer and a confident-sounding mistake.

What This Means for Germany

Germany is typically a second mover on new technology. According to Bitkom (Germany’s digital industry association), only 26% of German companies currently give employees access to AI tools. That sounds like a disadvantage — and in some ways it is.

But there is another reading: while US companies spent 2025–2026 learning expensive lessons, most German companies were still writing their AI strategy documents. They did not go through the costly phase, because they had not started yet. That is a rare chance to do it right — if they avoid simply importing the US playbook, delays and all.

Germany also has structural safeguards most US companies lack. Works councils (Betriebsrat) have a legal right under §90 BetrVG (the German Works Constitution Act, which sets out employee participation rights) to be informed and consulted before AI systems are introduced in the workplace. A works council that understands what it is looking at can prevent tokenmaxxing-style programmes before they are ever launched. That is not bureaucracy. That is governance that works.

The EU AI Act — in full effect from August 2026, requiring AI systems to be classified by risk, documented, and made transparent — adds further obligations. Companies rolling out AI without governance will have to retrofit compliance under time pressure. Building it in from the start costs less.

Figma, Claude Design & the Source of Truth

Anthropic (the company behind the Claude AI) launched Claude Design on April 17, 2026. It generates live, clickable web prototypes (interactive test versions of a website or app) from a plain text description — no design canvas, no dragging components around. Figma’s stock dropped 7% the day it launched. The conclusion many drew — that Figma is being replaced — was wrong.

Figma is not dying. Its role is narrowing, and that narrower role is becoming more important, not less. What is being automated away is generation: producing a user interface (UI — the buttons, screens, and menus users interact with) from a description. Claude Design, Cursor, v0, and Lovable all do this now. What cannot be automated is the canonical record — the single source of truth every AI tool reads to understand what “correct” looks like for your specific product.

Figma’s response was not to compete with Claude Design but to partner with Anthropic. Figma Config 2026 (June 23–25, San Francisco) made the direction explicit: a new MCP server (Model Context Protocol — lets AI tools read live context from GitHub, Slack, or Notion), Code Connect (links each Figma component directly to its real codebase counterpart), and Figma Motion (a built-in animation timeline). The message was clear: Figma is becoming the rulebook that all AI generation tools read from.

This has a direct cost implication. A designer with a well-documented, MCP-connected design system gives AI tools the context to get things right the first time. Fewer corrections mean fewer tokens and lower bills — the tokenmaxxing problem solved at the source.

What Comes Next

What we are living through right now has a name: the trough of disillusionment. Every major technology goes through it. The Gartner Hype Cycle (a well-known model that maps how technologies move from hype to reality) describes the pattern precisely: a peak of inflated expectations, followed by a crash when reality does not match the promise, followed eventually by a plateau where the technology is used in ways that actually work.

AI hit its peak somewhere around 2024. Tokenmaxxing leaderboards, mass layoffs, 97% cost-reduction slides — that was the top of the curve. The evidence of 2025–2026 — the failed financial outcomes, the rehiring waves, the bills that tripled while per-token prices fell — is the descent. This is not AI failing. This is the hype correcting itself.

The companies that come out ahead are the ones that use the disillusionment productively. Token spend is now treated as a real cost line (part of the P&L — Profit and Loss, i.e. the company’s financial bottom line), not just a usage report.

Salesforce CEO Marc Benioff introduced “agentic work units” as the new benchmark — and the concept is worth understanding, because it reflects exactly what the tokenmaxxing era got wrong. An agentic work unit is a completed, verifiable output: a support ticket resolved, a contract reviewed, a design iteration accepted, a bug fixed and tested. Not prompts sent. Not tokens consumed. Work done.

This reframes AI from a usage metric into a service with a price per outcome — and it puts the human back at the centre. Every work unit needs to be defined, accepted, and verified by someone who understands what good looks like. Tokenmaxxing removed that person. The agentic work unit model assumes they are there. That is what the plateau of the hype cycle looks like in practice.

Sources & Further Reading

• Forbes. AI Costs More Than The People It Replaced. July 2, 2026.
• Axios. AI can cost more than human workers now. April 2026.
• Fortune. Uber burned through its entire 2026 AI budget in four months. May 2026.
• Fortune. Microsoft reports are exposing AI’s real cost problem. May 2026.
• Forbes. After Tokenmaxxing, Token Spend Has Become the New Metric to Watch. July 10, 2026.
• Vucense. Meta’s Claudeonomics: 85,000 Employees Competing on AI Token Usage. 2026.
• MIT / Fortune. MIT Project Iceberg: AI can already replace nearly 12% of the U.S. workforce. November 2025.
• Great Learning. Tokenmaxxing Explained: The Hidden Cost of Workplace AI. 2026.
• Bitkom. KI-Nutzung in deutschen Unternehmen 2026. 2026.
• VentureBeat. Anthropic launches Claude Design, challenging Figma. April 2026.
• Figma Blog. Introducing the Figma MCP Server. 2026.
• BenchLM. LLM API Pricing Comparison July 2026. July 2026.
• CNBC. Employers who laid off workers for AI are reversing their decisions. July 1, 2026.
• Fast Company. The great AI layoff is turning into the great AI rehire. 2026.
• Forbes. Companies That Fired Workers for AI Now Want Them Back. May 21, 2026.
• IBTimes. AI Layoffs Backfire: 32% of Bosses Rehire Roles They Thought Robots Could Do. 2026.
• Futurism. Large Study Finds That Replacing Workers With AI Is Backfiring Badly. 2026.