If you are choosing between two adjacent Sonnet versions for the same job, the useful questions are what each costs per call, how fast each is, and how much of the prompt each gets from cache. We have figures for both from our own use.
Over the last 90 days, two Anthropic models handled book.questions.solver across 453 calls: Sonnet 5.5 took 337 of them and Sonnet 5 took 116. Neither returned a single failure. Sonnet 5 was last used on 2026-09-24, and Sonnet 5.5 is the one still in use.
Below we give each model's cost per call, median time, token figures and dates of use, then say where the comparison stops being reliable.
The two tools and when each ran
Sonnet 5 is anthropic/claude-sonnet-5. Sonnet 5.5 is anthropic/claude-sonnet-5-5. Same vendor, same job, adjacent versions.
Sonnet 5 ran first: 116 calls, first used 2026-09-23 and last used 2026-09-24. Sonnet 5.5 ran 337 calls, first used 2026-10-08 and last used 2026-10-15.
Sonnet 5.5 is the one still in use. Sonnet 5 is the one we stopped using. The record does not say why we stopped.
Cost per call
Sonnet 5.5 cost $0.01798 per call, $6.0594 in total over its 337 calls. Sonnet 5 cost $0.02414 per call, $2.8008 in total over its 116 calls.
Sonnet 5 cost 1.34 times what Sonnet 5.5 did per call, a gap of $0.00616 per call.
These are what we paid on our own volume. They are not a rate card, and they say nothing about what the vendor lists on its pricing page.
Speed and failures
Sonnet 5.5 had a median of 4.2 seconds per call. Sonnet 5 had a median of 5.2 seconds, which is 1.24 times slower, a gap of 1.0 seconds.
Neither tool had a failure in this window: 0 of 337 calls for Sonnet 5.5 and 0 of 116 for Sonnet 5.
That is the count of calls that came back as failures on our side. It is not the vendor's uptime, and a window with 453 calls and no failures is a narrow thing to lean on.
Tokens and the prompt cache
The token figures only make sense read together. Sonnet 5.5 averaged 7,375 input tokens per call in total. Of those, 6,561 were billed at full rate and 777 came from the prompt cache, which is 10.5 percent. It produced 469 output tokens per call.
Sonnet 5 averaged 10,083 input tokens per call in total, and all 10,083 were billed at full rate. Its cached input was 0 tokens per call, so 0.0 percent. It produced 398 output tokens per call.
So the two jobs were not the same size on average, 7,375 input tokens against 10,083, and only Sonnet 5.5 got any of its prompt from cache. The facts do not say why the input sizes differed, so we do not claim a cause.
Sonnet 5 paid full rate on its whole, larger input and cost more per call. These facts show both things sitting together, and they do not let us say how much of the cost gap comes from the cache and how much from the larger input.
Billed input alone would mislead here. Sonnet 5.5 shows 6,561 billed tokens per call against 10,083 for Sonnet 5, but part of that gap is simply that the total input was smaller, and part is the 777 tokens that came from cache.
The limits of this comparison
This is not a benchmark. These are our prompts, our one job, and our volumes, measured over the window in the facts.
The calls were not split between the tools in any controlled way. Sonnet 5 ran on two days in September and Sonnet 5.5 on a later stretch, with 116 and 337 calls respectively. The total input per call differed between them, so the two did not necessarily receive identical work.
A different job, different prompts or a different volume may produce different numbers. We have nothing here about answer quality, only cost, time, tokens and failures.
We also have nothing about what changed between the two stretches of use besides the model name, so we cannot attribute the cost, speed or token differences to the models alone.
For this job we would pick Sonnet 5.5 today: it was cheaper per call at $0.01798 against $0.02414, faster at a 4.2 second median against 5.2, had no failures, and it is the one we kept using. We would pick differently if your own measurements on your own prompts showed otherwise, or if you need a quality comparison, since our figures carry none.
Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.
Top comments (0)