Cheaper Per Token Can Still Mean More Expensive Per Task

Artificial Analysis priced Claude Sonnet 5 at its originally announced future standard rate of 3 US dollars per million input tokens and 15 per million output tokens. On that basis, its Intelligence Index run cost about 2.29 US dollars per task, compared with about 1.99 for Opus 4.8. The benchmark explicitly excluded Sonnet 5's launch promotion (1).

That historical comparison is useful, but it is no longer the current price comparison. Anthropic later made the promotional 2-dollar input and 10-dollar output rates permanent. Applying that one-third reduction proportionally to the published benchmark estimate gives roughly 1.53 US dollars per task, below the cited Opus 4.8 figure. This is an arithmetic illustration, not a new benchmark run.

A per-million-token number still leaves out the question that decides the bill: how many billed tokens and tool operations does the model use to reach an acceptable result?

A model that reasons in circles is not cheap. It is cheap to start.

For model selection, compare total cost per completed task, not only the sticker price per token.

The number nobody prices: tokens-to-answer

Per-million-token pricing is the sticker on the window. It tells you the rate. It tells you nothing about how far the model drives to reach the destination.

Two costs hide behind a single API call:

• The unit price. Dollars per million input and output tokens. This is what every pricing page advertises.
• The usage path. Input, cached input, output and thinking tokens, plus repeated model turns and any separately priced server tools.

Your bill combines the applicable rates with measured usage and any separately priced operations. Artificial Analysis calls its benchmark measure cost per task: the weighted-average token cost to complete one Intelligence Index task. Models that produce longer answers or more reasoning tokens have a higher benchmark cost at identical per-token prices (2).

Cut the unit price by 40 percent and let token consumption rise by more than that, and you have made the model look cheaper while making it more expensive.

What actually happened with Sonnet 5

The Sonnet 5 launch is a useful case study because the answer changes when the applicable price changes.

At launch, Anthropic announced 2 US dollars per million input tokens and 10 per million output through 31 August 2026, with a planned rise to 3 and 15. Anthropic subsequently cancelled that rise. As of 2 September 2026, Sonnet 5 remains at 2 and 10, while Opus 4.8 is 5 and 25 (3). Sonnet 5 is therefore 60 percent cheaper on both base input and output rates.

Then you run it. Artificial Analysis found that at maximum reasoning effort, Sonnet 5 used about 40 percent more output tokens per Intelligence Index task than Sonnet 4.6, and roughly three times the agentic turns. On knowledge-work evaluations it burned around six times more turns at max effort than at low effort. The performance gains came through longer reasoning chains and more tool calls, not through efficiency (1).

Stack that on a second change: Anthropic says Claude 4.7 and later use a newer tokenizer that produces approximately 30 percent more tokens for the same text than earlier models, with the exact increase depending on content and workload (3). Recount the actual prompts for the model you plan to run rather than applying 30 percent blindly.

The corrected lesson is not that Sonnet 5 is currently more expensive than Opus 4.8. It is that per-task rankings depend on both observed usage and the price actually in force. The same benchmark usage can change rank when a promotion becomes permanent.

"A cheaper unit price on a model that reasons in circles is not a discount. It is a deferred invoice. The teams that win read the whole receipt, not the price on the shelf."

Figures were checked on 2 September 2026. Artificial Analysis's 2.29-dollar estimate used the planned 3-dollar and 15-dollar rates, while Anthropic's current standard price is 2 dollars and 10 dollars. Your workload is not the benchmark suite, so measure it directly.

Why reasoning models break the sticker price

This is not specific to Sonnet. It is a general metering issue for models and agents that can vary their amount of thinking and tool use.

Anthropic's pricing documentation bills tool use through input and output tokens and may add usage charges for server-side tools. Its effort control affects response text, tool calls and thinking, so a higher setting can change both capability and spend (3, 5). Thinking and tool-use blocks are observable in API responses where the product exposes them; they should not be described as invisible work.

So a model can be:

• Cheaper per token and more expensive per task, because it thinks longer.
• More expensive per token and cheaper per task, because it reaches the answer in one pass instead of five.

The sticker price and the real cost are not just different numbers. They can point in opposite directions.

Cost per task, defined

If you take one metric from this article, take this one.

Cost per completed task is the total spend, across every token and every turn, to get one real task done to your quality bar. Not per token. Not per request. Per finished, acceptable answer.

It captures what the sticker price hides:

• Thinking tokens. Billed model output used for reasoning where the provider exposes or meters it.
• Output length. A verbose model bills more even at the same rate.
• Agentic turns and tools. Additional model turns consume tokens, and some server tools carry separate usage fees.
• Retries. Wrong answers you have to run again are not free.
• Tokenizer drift. The same text can count as more tokens on a newer model.

A model that is cheap to start and expensive to finish fails this measure. That is the whole point of using it.

How to measure cost per completed task

You do not need a research lab. You need your own tasks and a scale. Here is the process we run before we recommend a model to a client.

1. Define the task and the quality bar. Not "summarise this," but "produce a summary that passes this rubric." A task is only complete when it meets the bar, otherwise the retry belongs in the cost.
2. Build a small eval set from real work. Twenty to fifty representative tasks can be a practical starting sample. Public benchmarks provide context, but they are not your workload.
3. Run each candidate model to completion. Same tasks, same settings you would ship. Let it reason, call tools, and retry the way it will in production.
4. Count usage to done. Record uncached input, cache writes and hits, output, thinking, every model turn and separately priced tools. Anthropic's token-counting endpoint estimates input usage, while actual response usage and invoices remain the billing record (4).
5. Price the whole path, including failures. Multiply tokens by the real rate, add the cost of retries on tasks the model got wrong the first time. That total, divided by tasks completed, is your cost per completed task.

The ranking can flip. A model with a higher per-token rate can be cheaper to finish, while a lower-rate model can consume more billable usage.

What this means for model choice

The lesson is not "always pick the expensive model." It is "stop choosing on the sticker."

A few rules we work by:

• Match the model to the task, not the price list. A capable model that answers in one pass can be cheaper per task than a weaker one that loops. Route simple, high-volume work to cheap models and hard, ambiguous work to strong ones. We wrote the full routing playbook in how to cut LLM token costs in 2026.
• Tune the effort dial. Anthropic describes effort as a soft behavioral signal, not a strict token budget. Lower settings generally trade some capability for token efficiency, so sweep supported settings on your own eval instead of assuming maximum is best (5).
• Watch turns and tools, not just model output. Extra calls can add input, output and tool charges. A model that finishes in fewer turns may cost less even at a higher token rate.
• Re-run the numbers when a model updates. A new version can change the tokenizer and the reasoning behaviour at once, as Sonnet 5 did. Last quarter's cost ranking is not this quarter's.

Price per token is the marketing number. Cost per completed task is the number that lands on your invoice. Optimise the one you actually pay.

References

1. Artificial Analysis (2026): Claude Sonnet 5 cost an estimated $2.29 per Intelligence Index task at the planned $3/$15 rates, versus $1.99 for Opus 4.8; promotional pricing was excluded. Source (Accessed: 2 September 2026).
2. Artificial Analysis: methodology and definition of weighted-average Intelligence Index cost per task. Source (Accessed: 2 September 2026).
3. Anthropic: current pricing, permanent $2/$10 Sonnet 5 rates after cancelling the $3/$15 increase, tool costs and approximately 30 percent more tokens from the newer tokenizer. Source (Accessed: 2 September 2026).
4. Anthropic: the token-counting endpoint estimates input tokens including tools and supports active models; actual billing may differ slightly. Source (Accessed: 2 September 2026).
5. Anthropic: effort affects text, tool calls and thinking; it is a behavioral signal rather than a strict token budget. Source (Accessed: 2 September 2026).