What AI actually costs to run in 2026
Three charts made the rounds this month. Stacked together they tell a clear story: the price spread between the cheapest and priciest model is enormous, quality stops improving well before price does, and even the biggest AI spenders are quietly pulling back per employee. Here's what the numbers actually say.
The 270× price range
Run the same task through 21 current models and the cost per run ranges from $0.23 to $62.43 — a 271× spread for what is, on paper, the same job. Lined up cheapest to priciest, the jump isn't gradual. It's a cluster of models under $2, a second cluster in the single digits, and then a handful of frontier flagships that cost 10–30× more than the cheap cluster.
Almost every task people reach for a flagship model for — drafting, summarizing, classifying, first-pass coding — runs fine on something in that sub-$2 cluster. The $17–$62 tier exists for the sliver of work that genuinely needs top-end reasoning, not as the default.
Cost vs. quality: the Pareto frontier
A separate eval plotted cost per review against F1 score against human reviewers — how closely a model's judgment matched a person's, on a real review task. The model names were anonymized in the original, so we've kept it that way here. The shape of the plot is the point, not the brand names.
Frontier A costs 8.5× what the in-production model costs, for a lower F1 score. It isn't on the frontier at all — it's just expensive.Reading the chart, not a quote
The same eval flagged a separate case: one closed frontier model scored $2.22 per review on its default harness and $0.94 on a second, leaner harness — a 2.4× cost swing from the same model, purely from how it was wired up. The harness is sometimes a bigger cost lever than the model choice.
Enterprise AI spend per employee is falling — at the top
The Ramp AI Index tracks real business spend on AI — LLM subscriptions, coding agents, API tokens, and GPU cloud — by percentile of spend per employee. From January 2024 through August 2026, every tier climbed. But the top 1% of firms by AI spend recently turned over.
Top 1% of firms
Top 10% of firms
Median firm
Median and top-10% firms are still ramping spend hard. But the top 1% — the heaviest spenders, the ones who were first to put AI on the corporate card — pulled back 10% per employee even as their total AI usage kept growing. Read together with the first chart, the likely explanation isn't "less AI." It's routing: swapping flagship-tier calls for the sub-$2 tier wherever the task doesn't need the expensive model.
What this means if you're choosing models
In plain English
- Default to the sub-$2 cluster. DeepSeek V4, GPT 5.6 Terra-tier, Gemini Flash-tier models now cover most drafting, classification, and first-pass work at a fraction of flagship cost.
- Check whether you're even on the frontier. The scatter chart's clearest lesson: the most expensive point tested wasn't Pareto-optimal at all — a cheaper model beat it on quality. Expensive isn't a proxy for best; measure it.
- The harness can cost more than the model. A 2.4× swing from harness choice alone means the wrapper — retries, context length, tool calls — deserves the same scrutiny as the model pick.
- Reserve the top tier deliberately. Opus 5 and Fable 5.1-class models still lead on hard reasoning. Route to them on purpose, for the work that needs it — not as the default for everything.
- Even the biggest spenders are optimizing, not just scaling. The top 1% cutting per-employee spend while usage grows is cost discipline, not cost-cutting. Worth copying before a finance review forces the question.
Building AI workflows that don't waste spend
Model routing, cost-per-task tracking, and picking the right tier for the job — this is the day-to-day of the automations built here.
Keep reading
- Inside Anthropic — the $965 billion AI company — the founders, the Pentagon clash, and the Mythos model.
- AI Workflows — how the automations behind these routing decisions are built.
- AI case studies — interactive walk-throughs of live builds.