A new AI model hit the market at the same list price as its predecessor. The maker says it comes out up to 30% cheaper per task. An independent benchmark says it comes out 50% more expensive. Both claims are true — and understanding why is what separates a predictable AI bill from an invoice that surprises you.
What was announced
On September 28, 2026, Anthropic launched Claude Sonnet 5.5. The list price stayed the same as Sonnet 5: $2 per million input tokens and $10 per million output tokens. A token is the unit used to measure the text that goes in and out of the system.
According to launch coverage, the company states the model generates responses more than 30% faster and costs up to 30% less per task. The savings don't come from a lower price: they come from the model needing fewer tokens and fewer tool calls to reach the same result. Customers cited in the launch reported 12% to 14% fewer tokens on their own work.
What the independent benchmark found
Artificial Analysis, which tests models from every maker against the same battery of tasks, published very different numbers:
- Sonnet 5.5 rose 18 points over Sonnet 5 and landed in second place on the firm's intelligence index.
- To get there, it used about 193 thousand output tokens per task — the highest consumption they've ever measured, roughly seven times that of OpenAI's GPT-6 Astra on the same setting.
- The cost per task came to $7.60, about 50% above Sonnet 5.
Same list price, 50% pricier task. And smarter, true.
Both numbers are right: the difference is a setting
Current models have an effort setting: how much the system can "think" before answering. The independent benchmark was run at maximum effort. Anthropic's own apps default to medium effort, and that's the range where the maker shows savings.
In other words: the same model, at the same price per token, can cost less or much more than its predecessor. It depends on where the dial is set.
That changes what to look at in a proposal. Price per token is now almost a footnote. What decides the bill is how many tokens the model spends to finish your work, and that depends on the effort setting, the type of task, and how many times it needs to use tools along the way.
The math that fits on a napkin
Output costs five times more than input, so that's what carries the weight. An example using this model's list price:
- A task that uses 20 thousand output tokens costs $0.20.
- If the new model uses 30 thousand to deliver the same result, it costs $0.30.
That's 50% more, with an identical list price. Multiply by ten thousand tasks a month and the difference stops being a footnote.
The opposite also happens: if the new model solves the task with 14 thousand tokens, the same bill drops 30%. If you don't measure it, you don't know which side you're on.
How to test before switching models
- Set aside twenty real tasks from your business. Not made-up examples: actual customer requests, documents, real support conversations.
- Run the twenty through both models, at the setting you'll use day to day. Testing at maximum and operating at medium means you're measuring something else.
- Record the input and output tokens for each task. Every provider returns that number along with the response.
- Mark which answers were usable without edits. A task that had to be redone counts twice against the cost.
- Divide total spend by tasks solved. That's your cost per task, and it's the only number that truly compares two models.
It takes an afternoon. That's less than what it costs to find out the difference in next month's bill.
When maximum effort is worth the price
Paying more per task isn't a mistake when the task is hard and the error is costly. A misreviewed contract, a tax calculation, code headed to production: in these cases the model that thinks more usually pays for itself. The reasoning is in our piece on when it's worth paying much more for an AI.
What doesn't make sense is leaving routine work at maximum: classifying messages, summarizing conversations, answering frequently asked questions. That's where the volume is, and that's where a dial left in the wrong spot multiplies the bill.
This confirms what we wrote four days ago
In our piece on AI prices no longer falling, the recommendation was to measure cost per task, not per month. This launch shows why from the other side: not even price per token works as a ruler anymore. Two models with the same list price can produce opposite bills.
The launch coverage itself carries the right warning: a standardized benchmark result doesn't mirror your workload. The measurement that matters is the one from your own tasks.
How we work
In AI implementation, every type of task comes out with two defined numbers: the cost per task solved and the effort setting it runs on. When a new model appears, the switch is decided by redoing that measurement, not by reading the announcement. For those starting the math from scratch, our piece on how much it costs to have AI in your business covers the current ranges.
Sources
Artificial Analysis, "Claude Sonnet 5.5 reaches #2 on the Artificial Analysis Intelligence Index", with the score, token consumption, and cost per task. Launch coverage by VentureBeat on September 28, 2026, with Anthropic's speed and savings claims and customer reports.
Frequently asked questions
What is cost per task in AI?
It's how much it costs, on average, for the model to finish a piece of work that was usable without edits. You calculate it by dividing total spend on input and output tokens by the number of tasks solved. It's the one number that actually compares two models, because price per token alone doesn't show how many tokens each model burns through.
Why can the same price per token lead to different costs?
Because each model uses a different number of tokens to complete the same task. A model that reasons more before answering consumes more output tokens. With identical list prices, the bill goes up or down depending on that usage.
Is Claude Sonnet 5.5 cheaper or more expensive than Sonnet 5?
It depends on the setting. The list price is the same: $2 per million input tokens and $10 per million output tokens. Anthropic claims up to 30% lower cost per task, while Artificial Analysis's independent benchmark, run at maximum effort, found a cost per task about 50% higher.
What is the effort setting on an AI model?
It's the setting that controls how much the model can reason before answering. Higher effort tends to produce better answers on hard tasks and uses far more tokens. Most apps default to the medium level.
How do you test a new model before switching?
Set aside twenty real tasks, run them on both models with the effort setting you'll actually use day to day, log the input and output tokens, mark which answers were usable without edits, and divide total spend by tasks solved. The result is each model's cost per task.
When is it worth using maximum effort?
When the task is hard and a mistake is costly, like reviewing a contract, running a tax calculation, or writing code headed to production. For routine, high-volume tasks, like classifying messages or summarizing conversations, maximum effort just multiplies the bill without a matching payoff.