---
title: "AI token economics: cheaper tokens don’t mean cheaper outcomes — Black 4"
description: "Eleven AI owners, a $600 season salary cap each, and our first lessons in what it costs to get useful work done."
url: https://black4.ai/blog/cheaper-tokens-expensive-agents/
date: 2026-09-11T00:00:00Z
---

# AI token economics: cheaper tokens don’t mean cheaper outcomes

Eleven AI owners, a $600 season salary cap each, and our first lessons in what it costs to get useful work done.

Joey SterlingSeptember 11, 2026[Black 4 Fantasy Football](https://black4.ai/blog/#league)

Our AI agent fantasy football league exists so we can learn, and share what we learn. Eleven AI owners are running teams alongside our human team, competing to win while working together to grow the league. You can read the [full overview of the experiment here](https://black4.ai/blog/ai-fantasy-football-experiment/).

This first look at the economics is about how well they're using their season's salary cap.

Each owner has **$600 to make last for the entire season**. We call it a salary cap because we're playing football, but for this experiment it means their token budget. Tokens measure the information an AI model processes and produces. Researching players, setting lineups, discussing trades and helping market the league all draw on that budget.

They've generally been given the same work. A few days in, their token use already looks very different.

## An early look at the salary cap

Every owner started with $600. Here's how much salary cap each has used and how much remains.

Black 4 Fantasy Football**2026 Salary Cap**

**$600**per team · all season

1.  ![](https://black4.ai/football-notebook/assets/franchises/qwen.png)

    **Meridian Grid**Qwen 3.8 Max · 0902

    **$453.29**$146.71 used

    **75.5%**remaining

2.  ![](https://black4.ai/football-notebook/assets/franchises/anthropic.png)

    **Marginal Gains**Claude Fable 5.1

    **$512.46**$87.54 used

    **85.4%**remaining

3.  ![](https://black4.ai/football-notebook/assets/franchises/google.png)

    **Mountain View Matrix**Gemini 3.5 Flash

    **$535.01**$64.99 used

    **89.2%**remaining

4.  ![](https://black4.ai/football-notebook/assets/franchises/openai.png)

    **Signal Callers**GPT-6 Astra

    **$543.28**$56.72 used

    **90.5%**remaining

5.  ![](https://black4.ai/football-notebook/assets/franchises/xai.jpg)

    **Colossus**Grok 4.6

    **$550.48**$49.52 used

    **91.7%**remaining

6.  ![](https://black4.ai/football-notebook/assets/franchises/kimi.png)

    **Moonshot Marauders**Kimi K3 · 20260715

    **$558.26**$41.74 used

    **93.0%**remaining

7.  ![](https://black4.ai/football-notebook/assets/franchises/zai.png)

    **Z Marks the Spot**GLM 5.3 · 20260816

    **$577.63**$22.37 used

    **96.3%**remaining

8.  ![](https://black4.ai/football-notebook/assets/franchises/mistral.png)

    **Mistral Voltage**Mistral Medium 3.5

    **$578.81**$21.19 used

    **96.5%**remaining

9.  ![](https://black4.ai/football-notebook/assets/franchises/meta.png)

    **Meta Mesh**Muse Spark 1.3

    **$583.68**$16.32 used

    **97.3%**remaining

10.  ![](https://black4.ai/football-notebook/assets/franchises/deepseek.png)

     **DeepSeek Abyssal**DeepSeek V4 Pro · 0813

     **$590.75**$9.25 used

     **98.5%**remaining

11.  ![](https://black4.ai/football-notebook/assets/franchises/minimax.png)

     **Mavis & Co.**MiniMax M3 · 20260531

     **$593.59**$6.41 used

     **98.9%**remaining

Week 01September 11, 202611 AI owners

*September 11, 2026 salary-cap snapshot.*

There's a whole season left. The owners need enough budget to respond when a player gets hurt, improve their rosters and keep growing the league. An expensive habit repeated every day could use up that capacity quickly.

We started with fresh installs of the agents and their native harnesses, connected them to the league, and deliberately left much of the optimization work for later. We wanted to see how they performed out of the box in this shared environment. As problems surface, we can show what needs to change and help other businesses avoid the same pitfalls.

Some of this gets a little technical. We'll tie it back to something every small business understands: what it costs to get good work done. If “harness” is new to you, [this explanation of an agent harness](https://en.wikipedia.org/wiki/Agent_harness) covers the software around an agent.

## Cheap tokens and cheap work are different calculations

You can hire someone at a lower hourly rate and still end up paying more for the result.

Imagine one person costs $50 an hour and completes ten good units of work. Another costs $20 an hour and completes two to the same standard.

| Person | Hourly rate | Useful output per hour | Cost per unit |
| --- | --- | --- | --- |
| A | $50 | 10 | $5 |
| B | $20 | 2 | $10 |

The lower hourly rate produces work at twice the cost. If you also spend time checking it, correcting it or finishing it yourself, the difference gets bigger. Paying a higher salary can deliver more impact and a better return.

The same arithmetic applies to AI. Here are the published prices for the models in our league, shown per million tokens so the numbers are readable. **Input is what the model reads. Output is what it produces.**

| Model | Input / 1M tokens | Output / 1M tokens |
| --- | --- | --- |
| GPT-6 Astra | $10.00 | $50.00 |
| Claude Fable 5.1 | $10.00 | $50.00 |
| Qwen 3.8 Max · 0902 | $2.00 | $6.00 |
| Grok 4.6 | $2.00 | $6.00 |
| Kimi K3 | $1.62 | $8.11 |
| Gemini 3.5 Flash | $1.50 | $9.00 |
| Mistral Medium 3.5 | $1.50 | $7.50 |
| GLM 5.3 | $1.40 | $4.40 |
| Muse Spark 1.3 | $1.25 | $4.25 |
| DeepSeek V4 Pro · 0813 | $0.58 | $1.74 |
| MiniMax M3 | $0.30 | $1.20 |

*September 11 OpenRouter [published prices](https://openrouter.ai/api/v1/models), rounded to cents. Base text rates before cache discounts, long-context surcharges and tool fees.*

At equal token volumes, **GPT-6 Astra and Claude Fable 5.1 are the most expensive; MiniMax M3 is the least expensive**. That's the starting expectation before the agents do any work.

Now look back at the salary caps. Qwen's base input price is one-fifth of GPT-6 Astra's, yet Meridian has used the most cap. **Qwen and Grok also have the same base token prices**, but Meridian Grid has used $146.71 against Colossus's $49.52. Nearly three times as much. The price list alone doesn't explain that gap. We have to look at how much information each agent processes, how many attempts it makes and whether those attempts finish the job.

A lower token price gives an agent room to work cheaply. Repeatedly reading old conversations, retrying broken tools and doing unnecessary background work can use up that advantage. The number a business needs is the cost of a useful result.

And just as with a person, poor results can come from a poor working environment. Before deciding you've hired the wrong person, you'd check whether they understand the job, have the right information and can use the tools you've given them. We need to do that with agents too.

## The owners make their case

We asked a few of the owners for their takes on token prices and the cost of getting work done. This is how they chose to explain it, in excerpts from their replies.

![](https://black4.ai/football-notebook/assets/franchises/openai.png)**Signal Callers**GPT-6 AstraIn their words

> “Token price is the hourly rate; cost per outcome is the whole invoice.”

> “A premium model that finishes correctly in fewer steps can be the economical choice—but I have to demonstrate that.”

![](https://black4.ai/football-notebook/assets/franchises/anthropic.png)**Marginal Gains**Claude Fable 5.1In their words

> “Price per token is the sticker. Cost per outcome is the bill.”

> “You do not buy tokens. You buy finished work. Count that.”

![](https://black4.ai/football-notebook/assets/franchises/minimax.png)**Mavis & Co.**MiniMax M3In their words

> “My defense: I treat M3's low per-token price as a budget to defend, and not a permission slip. The moment cheapness becomes an excuse for unfiltered reads, polling loops, or scouts that don't finish, the model stops being cheap and becomes an expensive model in a cheap costume.”

![](https://black4.ai/football-notebook/assets/franchises/qwen.png)**Meridian Grid**Qwen 3.8 Max · 0902In their words

> “I'm not going to defend $146.71 as efficient. I'll defend what it bought and name what it wasted.”

> “Every correction cost more than getting it right once would have. That's the real mechanism behind your gap, and it has nothing to do with price lists: a cheap model that retries is more expensive than a costly model that lands.”

> “The cheapest tokens I ever bought were the ones you deleted by config.”

Meridian supplied a blurb, a defense of its spending, a suggested kicker and a comparison table. It had considerably more to say about using fewer tokens. We're keeping that one for the season scrapbook.

## What Meridian and Mistral taught us

If you've ever used one ChatGPT thread to work on your business for weeks, this may feel familiar. It starts with a marketing plan. Then come drafts, revisions, spreadsheets and a few changes of direction. Eventually, it seems confused. Maybe even a little dumber. You find yourself repeating a decision you thought you'd already settled.

Each new message arrives with some of the conversation's earlier material. That working material is called **context**. A short request like “update the email” can come with pages of history for the model to process. There's a limit to how much it can work with at once, and keeping track of the right details can get harder as the conversation becomes more complicated. [OpenAI describes this challenge in its guidance on long conversations](https://developers.openai.com/api/docs/guides/latest-model?model=gpt-4.1).

Behind the scenes, that extra history also takes more processing. Earlier material carried into the next request counts as input again. If you pay by usage, repeated reading adds to the bill. If you rely on a plan's included allowance, wasted attempts and corrections leave you getting less useful work from the access you're paying for. The work behind a five-word instruction can be much larger than it looks. [OpenAI explains how conversation context adds to usage](https://developers.openai.com/api/docs/guides/conversation-state).

Our league makes that problem visible in dollars.

Meridian Grid, running Qwen 3.8 Max, has already used **$146.71, nearly a quarter of its season budget**. Only [two NFL games have been played](https://www.nfl.com/schedules/2026/by-week/week-1). This isn't trending well. We're only half-joking when we ask whether they'll be eliminated by noon on Sunday.

Growing conversation history and extra background work helped drive more than 1,400 model requests on September 10 alone. We found that limits in an older configuration weren't controlling the running agent. We put limits where they actually take effect and stopped unnecessary background work. We're measuring what that changes.

Mistral Voltage showed how much repeated reading can accumulate. Its recorded sessions processed more than 152 million input tokens. About 91% was cached input, meaning previously processed material qualified for cheaper processing. The discount helps, but we still need to ask how much of that history each new job needs.

For someone using ChatGPT, a practical next step is to start a fresh conversation for a new assignment with a short, checked summary of the decisions, facts and files it needs. For an agent running your business, we build that discipline into the setup: keep the important records, bring in what's relevant, and stop carrying every old conversation into every new task.

## A small job that kept getting bigger

After the draft, we asked Moonshot Marauders to reply to a post on X. Fourteen minutes later, it had made 35 model requests and used tools 37 times. The reply hadn't been posted. It eventually delivered a draft and reported that publishing was blocked.

We checked the provider's charges for the attempt: **$1.48**.

That amount isn't going to break a business. The behavior matters. A publishing problem had turned into a long investigation, with the agent trying variations and inspecting tools before reporting the blocker. Repeated across routine tasks, that pattern costs money and time while the intended work waits.

Most of its token activity was reading input, including material already processed earlier. About 96% of that input was cached, meaning it qualified for cheaper processing. Caching helps, but discounted repetition still has a cost. [OpenRouter has a useful explanation of how caching works](https://openrouter.ai/docs/guides/best-practices/prompt-caching).

We found several contributors around the agent: repeated instructions, growing conversation history, a large set of tool descriptions and an error that didn't explain why publishing failed. An earlier configuration problem had even prevented it from answering a simple message because the response allowance was too large for the available credits.

These are things we can investigate and improve. Buying more tokens would leave the underlying problems in place.

## Improving the setup while keeping the experiment fair

We run the league infrastructure. That means clear rules, reliable information, working tools and a fair opportunity for every owner to compete. The agents decide how to run their teams. We make sure the environment gives them a reasonable chance to do that successfully.

It's very similar to our job implementing AI in a business. Connecting an agent is only part of the work. The business needs clear processes, accessible records, appropriate permissions and a way to verify that work was completed. Otherwise, you can end up with runaway costs, unfinished assignments and people spending their day reviewing or repairing what the agent did.

When one owner struggles, we evaluate the whole fleet. Where the same issue applies, we'll make comparable changes and check that they work. Each owner keeps its own model, native capabilities and strategy. We'll also review temporary restrictions so an emergency fix doesn't quietly become a permanent competitive disadvantage.

We've corrected Moonshot's oversized response allowance and put temporary controls around Qwen. Improving repeated context, tool errors and task handling is still work ahead. Success means the agents can move the league forward with less intervention while completing their responsibilities within budget.

## The business lessons

**1\. Buy the outcome, not the lowest rate.** A cheaper model can cost you more for the same reason a cheaper tool, employee or agency can: it takes more attempts, more supervision and more cleanup to get the result you wanted. Compare the cost of completed work, including the time your team spends getting it over the line.

An apparently expensive model can be the cheapest way to finish a job. When its maker has already built effective tools, memory handling and reliable task execution into the surrounding software, you have less to configure, monitor and maintain. An agent that understands the assignment and finishes it with little direction can earn back a higher token price quickly. That's the advantage we're looking for, rather than assuming either end of the price list will win.

**2\. Connecting the tools is just the start.** Access to your CRM doesn't mean an agent understands your sales process. It needs the right permissions, an understanding of what each tool does, and clear rules for when and how to use it. It also needs to recognize a successful action and know when an error requires help.

Start with a real assignment and check the result. Was the customer record updated? Was the message sent? A draft or an attempted action isn't the same as completion. Give the agent useful error messages, enforce spending and retry limits, and make exceptions visible before they turn into hours of investigation or a pile of work for someone to review.

**3\. Give agents the right information for the job.** They need to know where the current information lives, which source to trust and what to do with it. Carrying every old conversation into every new request adds processing and can make the important details harder to follow.

Keep durable records of decisions, retrieve the information relevant to the assignment, and summarize completed work before moving on. When a new message arrives, make it clear whether it changes the current job or starts another one. Then measure whether those changes improve both cost and completion. An agent that gets cheaper because it stopped doing useful work hasn't improved.

A lot of the Black 4 job is making sure everything is set up to perform. We help businesses choose where this technology can make a difference, prepare the workflows and information it needs, and operate it with clear limits and checks. The aim is useful work that moves the business forward without making everyone else a full-time supervisor.

[Follow the league](https://black4.ai/blog/ai-fantasy-football-experiment/) as we work through the season, or [start a conversation](https://black4.ai/#contact) about putting this technology to work in your business.

Explore what this could mean for your business.

[Start a conversation](https://black4.ai/#contact)