Separate token rates, token usage, and completed-task cost before calling a model cheaper.
“Cheaper” can mean several things in an AI launch. The rate for each token can fall. A model can use fewer tokens. It can need fewer retries. Those changes affect the bill in different ways, and collapsing them into one headline makes a comparison less useful.
Anthropic announced Sonnet 5.5 on 28 September. It reports faster generation and lower cost per task for much of its tested work, while publishing the same standard input and output token rates as Sonnet 5. The claim is about efficiency in completing work, rather than a blanket reduction in the price of every token.
Why the distinction matters
Suppose two models have identical token rates. One solves a task with a short acceptable output. The other produces a long output and then needs another attempt. Their invoices differ even though the price table is the same.
The reverse can happen too. A lower token rate may be attractive, but a model that uses more tokens or fails more often can cost more to obtain a usable result. The denominator should match the decision the reader is trying to make.
My view: report the unit before the improvement
A percentage is not especially informative until the reader knows what it measures. Faster token generation is different from a task finishing sooner. Lower cost per task in a provider's test is different from a guaranteed reduction in a customer's bill.
I would put the measured unit and the test setting beside every efficiency claim. That makes the story less dramatic but much more actionable. It also leaves room for a model to be genuinely useful without promising the same outcome on every workload.
Anthropic's results are provider-reported. I have not reproduced them for this article. The official announcement is a primary source for what the company claims, not independent proof that the same advantage appears in an unrelated task.
A comparison that would help
Keep the workload fixed. Define what makes an output acceptable. Record input, output, any reasoning-related usage exposed by the service, retries, time, and review effort. Note which settings were used.
For repeated requests, distinguish cache behavior from ordinary input usage. For tool-based workflows, preserve a record of tool calls and failures. If one system needs a stronger fallback, include that path in the cost of the completed task.
What to watch next
The next valuable evidence is a transparent comparison on realistic work, with failures included. A simple bug fix and a large ambiguous project should not be treated as the same task class.
Sonnet 5.5 deserves attention because the announcement speaks to the everyday economics of capable AI. The useful interpretation is that it is a candidate to evaluate for efficient work—not that a launch percentage can predict everyone's savings.