What a token really costs · Part 6 of 6
The cheapest token isn't the cheapest task
Claude Sonnet 5.5's list prices are half of Opus 5.5's, yet the same benchmark tasks cost 28% more to run on Sonnet.
A price list compares models per token. A task doesn’t come with a fixed number of tokens: each model decides how long to think, how much to say and how many steps to take. And a model that gets it wrong needs another try.
Artificial Analysis, an independent benchmarking company, runs the same set of tasks on every major model and publishes what a task cost at list prices. On 6 October 2026, a task cost $5.98 with Claude Opus 5.5 and $7.67 with Claude Sonnet 5.5, both at their highest effort. Sonnet’s input and output prices are half of Opus’s, but it generated 62% more output tokens to get through the tasks, and it scored lower: 56 against 58.
Models with the same prices differ even more. GPT-6.1 Sol costs $2 per million input tokens and $10 per million output tokens, exactly like Sonnet 5.5. A task on Sonnet cost 11 times as much: $7.67 against $0.72.
Go deeper: cost per task, effort and attempts
Artificial Analysis calculates cost per task “by multiplying input, cached, and output token prices by tokens consumed across the workload”, weighted the same way as the benchmark’s score, the Artificial Analysis Intelligence Index.
| Model | Effort | Input / output price per million | Output tokens | Score | Cost per task |
|---|---|---|---|---|---|
| Claude Opus 5.5 | max | $4 / $20 | 260 million | 58 | $5.98 |
| Claude Sonnet 5.5 | max | $2 / $10 | 420 million | 56 | $7.67 |
| GPT-6.1 Sol | max | $2 / $10 | 67 million | 52 | $0.72 |
| Claude Opus 5.5 | low | $4 / $20 | 20 million | 42 | $0.55 |
Effort is part of the price. At low effort, Claude Opus 5.5 cost $0.55 per task instead of $5.98, and scored 42 instead of 58.
Cost per task counts every task, solved or not. In real work a failed try is paid for too, and then tried again: a model that gets a task right half the time costs twice its per-try price for every task it finishes. That is the attempts in this series’ formula. A price list can’t tell you how many tokens or tries your own tasks take. Running a sample of them on each model can.