Why cost per successful task beats token price

“Token prices fell again” is becoming the least informative sentence in an AI budget review. Between March and August 2026 the effective price US companies paid per token dropped 41%. Over the same window, the heaviest 1% of AI spenders cut per-employee AI spend by nearly 10%. Price and bill moved the same way, but not by the same order of magnitude — the unit of account is wrong. Cost belongs per successful task.

What the payment data shows: price down 41%, bill not recovered

The Ramp AI Index tracks payment data across roughly 70,000 companies. Its August 2026 reading produced three numbers:

• Average effective token price of $0.68 per million tokens, down 41% from the March peak of $1.15;
• 56% of companies paid for AI products, up just 0.4 percentage points month over month;
• Per-employee AI spend at the heaviest 1% of spenders fell nearly 10%, to $7,205.

Ramp chief economist Ara Kharazian attributes the drop to two forces stacking: falling token prices, and customers trading down to cheaper models — ChatGPT 5.6-Terra or Anthropic Sonnet instead of the more expensive frontier releases. In his words, competition is “not just driving the price down, but driving spend down at the top 1% of companies that previously the market was expecting to drive much of the growth going forward” (via TechCrunch, 9 September 2026).

One number in the same dataset gets overlooked: only 6.4% of AI-spending businesses used a model-serving or inference platform in August. Open-weight and hosted inference are real growth directions, but the base is still too small to change the shape of the aggregate bill.

Two caveats travel with these numbers. First, Ramp’s sample skews tech-heavy: a US Census Bureau survey updated on 23 August found just 22% of businesses report using AI, so the two figures cannot substitute for each other. Second, this is a single August month — a vacation month. The same Ramp index showed little to no adoption growth between August and October 2025 before picking up at year end. One month does not make a trend.

Quality spreads 1.4x. Cost spreads 30x.

Artificial Analysis moved the unit of measurement in its Coding Agent Index from “model” to the combination of model, harness and execution settings. The reasoning is plain: the same model inside a different harness changes in performance, token consumption and cost. Its May 2026 launch note produced a set of spreads that are hard to ignore:

• Cost per task on API token pricing: $0.07 to $2.26, a spread above 30x;
• Composite score: 43 to 61, a spread of roughly 1.4x;
• Tokens per task: 1.7M to 4.8M, a spread above 3x;
• Time per task: roughly 6 to 40 minutes, a spread above 7x.

Note the version: the index has since moved to v1.5 with updated benchmarks and combinations. The figures above come from the v1.0 launch note of May 2026 and should always be cited with that version attached. Current readings belong to the live Coding Agent leaderboard page.

The most instructive entry is the most expensive combination measured: GLM-5.1 in Claude Code at $2.26 per task. It got there mainly through token volume — 4.8M tokens per task, which the launch note partly attributes to the model entering loops — not through the highest per-token price. Another open-weight model in the same harness, DeepSeek V4 Pro in Claude Code, costs $0.35 per task. That is a 6.5x cost difference for a 3-point score difference.

What this breaks is the default assumption that price ranking equals cost ranking. The first explanatory variable for cost is token consumption, and token consumption is set by harness behaviour: prompt structure, cache hits and loop control. The same index records cache hit rates ranging from 80% to 96% across combinations, while cached input typically prices below half of standard input. Change the routing for a model and its unit economics move.

Our judgment: move the budget unit to cost per successful task

Three actions, all executable:

1. Reporting unit. Monthly reporting should carry two numbers: cost per successful task and tokens per successful task. That requires defining a successful task first — who judges it, against what — or the metric will be optimised into a flattering number.
2. Procurement baseline. Any quote must pin the harness version and the reasoning effort setting. Token consumption for the same model can differ more than 3x across harnesses, and that difference lands directly on the bill. A side-by-side comparison without a pinned baseline measures nothing.
3. Do not cross-validate price indices. Payment-side (Ramp at $0.68 per million tokens), gateway-side (Silicon Data’s LLM Token Expenditure Index, $0.98 per million tokens on 10 September) and vendor list prices are three different measurements. A CFO report should use cost per successful task computed from your own bill; industry indices are for direction, not for budget amounts.

On Jevons: this window neither proves it nor disproves it

The widely quoted Jevons argument runs: lower unit price, higher usage, higher bill. The new demand-side data points the other way. In this window, price cuts were not offset by volume. As TechCrunch put it, the labs have yet to make up for the price cuts with growing volume.

The safe formulation: unit prices really are collapsing, but the heaviest customers’ bills have not risen to match — the process by which price cuts get bought back through volume has not happened yet in this window’s data. That is not enough to overturn a multi-year conclusion (the window is short, single-month, and a vacation month), but it is enough to require any budget model that cites Jevons to label “volume will compensate” as a hypothesis, not a fact.

One question for buyers

If your AI budget model carries “price per million tokens” as its cost variable, answer this first: this quarter, how many tokens did one successful task consume, and is that number up or down on last quarter? Without that number, price movements have no explanatory power over your bill.

References: Ramp AI Index, August 2026 (Ramp payment data, ~70,000 companies; $0.68 per million tokens, July top-1% median $7,400 per employee, 6.1% on model-serving platforms), ramp.com/data/ai-index-august-2026 (curl HTTP 200) | TechCrunch, 9 September 2026 (56% paid adoption +0.4pt, top-1% per-employee spend down nearly 10% to $7,205, 6.4% on inference platforms, US Census 22% comparison, Kharazian quote), techcrunch.com/2026/09/09/ai-spend-per-employee-slumped-at-top-firms-in-august-summer-doldrums-or-a-warning-sign (curl HTTP 200) | Artificial Analysis Coding Agent Index v1.0 launch note (cost per task $0.07 to $2.26, composite score 43 to 61, 1.7M to 4.8M tokens per task, cache hit rates 80% to 96%, GLM-5.1 in Claude Code $2.26 at 4.8M tokens, DeepSeek V4 Pro in Claude Code $0.35, time per task ~6 to 40 minutes), x.com/ArtificialAnlys/status/2053865095076438427 (May 2026) | Current leaderboard and methodology (v1.5), artificialanalysis.ai/agents/coding-agents (curl HTTP 200) | Silicon Data LLM Token Expenditure Index (gateway usage-weighted, $0.98 per million tokens on 10 September), silicondata.com/products/silicon-index/llm-token-expenditure-index (curl HTTP 200) | DX, AI coding assistant pricing and ROI guide ($200 to $600 per engineer per month including seat and tokens), getdx.com/blog/ai-coding-assistant-pricing.