---
source_url: https://marquard.org/posts/agent-cost-per-task/
title: "Tokens got cheap. Tasks didn't."
date: 2026-07-21
summary: "Per-token prices fell 50× in three years. The bill for a finished agent task went the other way, and the spread is the number nobody benchmarks."
categories:
  - "operator notes"
word_count: 1011
reading_time_minutes: 5
---


Mozilla published its first [State of Open Source AI report](https://stateofopensource.ai/) this month, and the headline numbers are real: GPT-4-class inference fell from $20 to $0.40 per million tokens in 36 months, a 50× collapse that outran the bandwidth and compute price curves of the dotcom era. The report's framing is that the price of intelligence collapsed and open weights won the commodity layer.

I run agents against my own work all day, and my meter tells a more complicated story. The price of a token collapsed. The price of a finished task went sideways, and in some places up.

![Two lines over 2023–2026: the price per million tokens collapses 50× while the cost of a finished task stays flat and then rises, the widening gap between them labeled "the spread nobody benchmarks."](chart.svg)

## The strongest open model ever released costs triple per task

Kimi K3 landed last week: 2.8 trillion parameters, an Artificial Analysis intelligence index of 57, third overall behind only the top two closed models, and [ahead of Claude Opus on several coding benchmarks](https://aidailybrief.ai/e/2026-07-17). Best open-weight model ever shipped, by six clear points. On the per-token menu it undercuts the frontier: $5.40 blended per million tokens against $9 for Opus and $10 for GPT-5.5.

Then look at what a task costs. Artificial Analysis [clocked K3 at $0.94 per benchmark task](https://aidailybrief.ai/e/2026-07-17#cost-per-task-tripled), triple its own predecessor, Kimi 2.6, from less than three months earlier. On one widely shared test it burned 13,241 reasoning tokens to produce 3,417 tokens of output. Capability jumped 13 index points; the bill for finishing something tripled. Meanwhile DeepSeek V4 Pro completes a task for four cents, which puts a 23× cost spread inside the open column of the same leaderboard. Part of that gap is DeepSeek doing less thinking, which is the whole point. What a finished task costs is set by how hard you make the model work, not by the sticker price per token.

That spread deserves more attention than the ranking. Reasoning models buy their benchmark position with thinking tokens, so every leaderboard jump ships with a bigger token appetite. Cheaper per token and pricier per outcome can both be true of the same release, and lately they usually are.

{{< callout >}}
The best open model ever shipped undercuts the frontier per token and costs triple per task. Both numbers are real. They just answer different questions.
{{< /callout >}}

## My own meter says the same thing

For about six months I've run an [agentic OS over my own work](https://marquard.org/posts/agentic-os-six-months/) — a chief-of-staff agent that reads my inbox, writes my morning brief, preps my meetings, drafts in my voice. I meter every session at API list prices, because a number I don't track is a number I can't reason about.

Over the past five weeks: 125 sessions, about $929 metered at list prices.

The interesting part is the spread inside that number. My morning brief, a real task that reads overnight email, my calendar, my task list, and the news, lands around $6 a run. The voice-editing pass that makes a blog post sound like me instead of like a model averages $32 a run, on the same models at the same prices. One bad run of it stretched to 452 conversational turns before it settled.

Nothing about the models explains a 5× gap between those two tasks. The shape of the work explains it: how many turns the task takes, how many times the agent re-reads context, how many verification passes I demand before I'll trust the output. My most expensive workflows are the ones where I've raised the quality bar, because a higher bar means more rounds of checking before I'll accept what comes back. So I don't try to optimize that pass down — the voice fidelity is non-negotiable, and I'd rather pay for the extra rounds than put something generic out under my name.

{{< callout >}}
The workflows that cost me the most are the ones where I raised the quality bar. Every extra pass of checking gets paid for in tokens. The bar is a line item now.
{{< /callout >}}

That matches what the wider market is finding. Mozilla's own report concludes that value is moving up the stack, out of the commodity model layer and into the agentic harness. I'd extend that: the bill moved up the stack too. The harness decides how many tokens a task consumes, which means the harness, not the model's price tag, decides what your product costs to run.

## The questions that predict the bill

If you're a PM putting an agent feature into production, the per-million-token price on the pricing page is close to useless for forecasting cost. These are the numbers I'd want before writing a unit-economics line into any plan:

1. **Cost per finished task, at the median and at p95.** Not per call, not per million tokens. A task that's cheap on average and occasionally runs 50× long will find your worst customer and live there.
2. **Turns per task, and what triggers a retry.** Turn count was the whole difference between my $6 task and my $32 one. Retries compound quietly.
3. **The verification multiplier.** How many passes of checking does the output need before you'd put your name on it? Every notch of quality you demand is paid for in review tokens.
4. **Cache hit rate.** Agents re-read their own context constantly. Mine run 82–93% cached depending on the workflow, and the difference between those two figures is real money at volume.
5. **The cost trend of your task, not the model.** Model prices fall on a curve. Task costs follow your product's appetite for reasoning and checking, and that appetite has been growing faster than prices have been falling.

None of this argues against open weights, and none of it argues against reasoning models. K3 is a remarkable release, and $0.94 for a completed task of frontier-grade work is still absurdly cheap by any pre-2023 standard. My argument is narrower: the number everyone quotes is the input price, and the number that decides whether your agent product has a margin is the output price of a finished, verified task. Those two numbers have been moving in different directions for a year.

Benchmarks price intelligence by the token. Your CFO pays for it by the finished task. Which number is in your board deck?