The price sheet lies
List price tells you almost nothing about what AI really costs.
Muse Spark 1.2 (xhigh) is listed 82% cheaper than GPT-5.5 (low). It costs 287% more to actually run.
The flip
Same models, two orderings
Each model ranked by list price on the left, and by what it actually costs to run Artificial Analysis's Intelligence Index on the right. Follow a line across — where it climbs, the sticker price lied.
14 models with both a list price and a measured run cost · updated 9 hours ago
Why the sticker lies
You don't buy tokens. You buy answers.
A price sheet quotes dollars per million tokens. But a task isn't a fixed number of tokens — and that's where the bill hides.
List price is per token
The number on the pricing page is $ per million input/output tokens. It says nothing about how many tokens a model will spend to finish your task.
Reasoning models are verbose
A "thinking" model can emit 10–50× more tokens chewing through the same problem. Cheap per token × a mountain of tokens = an expensive answer.
The real unit is the task
Run the same fixed benchmark across models and the cheap-looking ones often cost the most. Rank by the finished job, not the sticker.
Receipts
Where list price misleads the most
Sorted by how far a model moves when you re-rank by real cost. ▲ means it's pricier than its sticker suggests; ▼ means it's a quiet bargain.
| Model | List $/Mtok | Actual run cost | Re-rank |
|---|---|---|---|
| Mistral Small 3.2 Mistral | $0.15 | $163 | ▲ 4 pricier |
| GPT-5.5 (low) OpenAI | $11.25 | $358 | ▼ 5 cheaper |
| Muse Spark 1.2 (xhigh) Meta | $2.00 | $1,385 | ▲ 2 pricier |
| Grok 4.6 (medium) SpaceXAI | $3.00 | $1,937 | ▲ 2 pricier |
| Claude Sonnet 4.6 (Adaptive Reasoning, Max Effort) Anthropic | $6.00 | $3,356 | ▲ 2 pricier |
| Mistral Large 3 Mistral | $0.75 | $130 | ▼ 2 cheaper |
| DeepSeek V3 (Dec '24) DeepSeek | $0.46 | $25 | ▼ 2 cheaper |
| Nemotron 3 Ultra 550B A55B (Reasoning) NVIDIA | $1.18 | $443 | ▲ 1 pricier |
| Claude Sonnet 5 (Adaptive Reasoning, Low Effort) Anthropic | $4.00 | $653 | ▼ 1 cheaper |
| Claude Fable 5.1 (Adaptive Reasoning, Low Effort, Default Fallback) Anthropic | $20.00 | $3,158 | ▼ 1 cheaper |
List price is the blended 3:1 input:output rate. "Actual run cost" is what it costs to run the full Artificial Analysis Intelligence Index, from Artificial Analysis — a fixed task suite, so the only variable is how each model behaves.
Citation
Use this in your work
Every figure here is open data. Here's a ready-made citation, plus the methodology behind the numbers and the full State of AI.
Copy a citation
Free to use and cite under CC BY 4.0. See how this is measured.
Champlin Enterprises. (2026). The AI Price Sheet Lies (MyTokenTracker) [Data set]. MyTokenTracker. Retrieved September 9, 2026, from https://mytokentracker.io/price-vs-cost
@misc{mytokentracker-price-vs-cost,
title = {The AI Price Sheet Lies (MyTokenTracker)},
author = {{Champlin Enterprises}},
year = {2026},
howpublished = {MyTokenTracker, \url{https://mytokentracker.io/price-vs-cost}},
note = {Accessed September 9, 2026. Licensed CC BY 4.0.},
url = {https://mytokentracker.io/price-vs-cost}
}
Need a fixed point in time? Every day’s data is permanently archived in the open-data repository, so you can cite a specific date by linking that day’s committed file.
Free weekly digest
Stop guessing what AI costs
We track the real, measured cost of every model — not the sticker price. One line to install, free forever. Get the weekly headline in your inbox.