The choice is between Sonnet 5.5 and Sonnet 5 for the same job. Both are Claude models, and both did editorial.boxes.anchor.read for us: 365 calls over the last 90 days, 329 on Sonnet 5.5 and 36 on Sonnet 5.
We used Sonnet 5 first, from 2026-09-25 to 2026-09-29, and Sonnet 5.5 from 2026-09-30 to 2026-10-26. Sonnet 5.5 is the only one of the two still in use.
This post sets out what each cost us per call, how long each took, how often each failed, and what the token figures show. It also says where the comparison stops being useful.
What each one cost us per call
Sonnet 5 (anthropic/claude-sonnet-5) cost $0.01007 per call over its 36 calls, $0.3624 in total. That made it the cheaper of the two per call.
Sonnet 5.5 (anthropic/claude-sonnet-5-5) cost $0.01755 per call over its 329 calls, $5.7731 in total. That is 1.74 times the cost of Sonnet 5 per call, or $0.00748 more per call.
These are the amounts we paid on our own volume, not a vendor rate card. They say what this job cost us, and nothing about what either model costs on someone else's workload.
Speed and failures
Sonnet 5.5 had a median of 2.0 seconds per call. Sonnet 5 had a median of 2.6 seconds, which makes it 1.3 times slower than Sonnet 5.5, a gap of 0.6 seconds.
Neither tool had a single failure. Sonnet 5.5 came back with 0 failures in 329 calls and Sonnet 5 with 0 in 36. That is the failure rate on our side, not the vendor's uptime, and 36 calls is a small sample to hold against 329.
What the token figures show
The jobs were not the same size. Sonnet 5.5 averaged 9,059 total input tokens per call and Sonnet 5 averaged 4,414. Read as written, the record shows Sonnet 5.5 handling a bigger input per call. Output ran at 143 tokens per call on Sonnet 5.5 against 163 on Sonnet 5.
Cache use was low on both. On Sonnet 5.5, 1,097 of those input tokens per call came from the prompt cache, 12.1 percent, leaving 7,811 billed at full rate. On Sonnet 5, 219 came from the cache, 5.0 percent, leaving 4,196 billed at full rate.
Most of the input was a cache miss on both tools, so most of what we paid for was input billed at full rate. Sonnet 5.5 had the higher cache share, but it still billed 7,811 input tokens per call at full rate against 4,196 on Sonnet 5.
The record does not say why the input per call is larger on Sonnet 5.5, so we make no claim about it. The higher cost per call of Sonnet 5.5 goes with the larger input recorded for it, and the figures alone do not separate what the model charges from what the job sent.
Which one we kept
We kept Sonnet 5.5. Its last call was on 2026-10-26, while Sonnet 5 last ran on 2026-09-29, the day before Sonnet 5.5 first appears.
The record does not say why we moved. It shows the switch and the cost per call on each side, and nothing about a reason.
Sonnet 5 ran for 5 days and Sonnet 5.5 for 27 days, so the two tools did not overlap at any point in the window.
The limits of this comparison
This is not a benchmark. It is our own prompts, our own job and our own volumes, measured over the window in the facts. A reader with a different job may see different numbers.
The two tools also did not run side by side. Sonnet 5 had 5 days of use and 36 calls, and Sonnet 5.5 had 27 days and 329 calls. The input per call differs between them, so the cost and speed figures are not a like-for-like test, and nothing here speaks to output quality.
The failure counts are on our side only, and with no failures on either tool they do not separate the two.
Today we would pick Sonnet 5.5 for this job, because it is the one we kept running, it had no failures in 329 calls, and its median of 2.0 seconds was the faster of the two. We would pick Sonnet 5 instead if cost per call is the deciding factor, since it came in at $0.01007 against $0.01755. Whichever you choose, check how much of your input comes from the prompt cache first, because on both tools most of it was billed at full rate.
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.


