Roundup noted Anthropic released Claude Sonnet 5.5, priced at $2/$10 input/output with a stated 30% average token reduction.
Following·Day 5·first covered 28 Sep 2026·3 sources
Anthropic Launches Claude Sonnet 5.5 With 30% Token Savings Claim
The gist
The efficiency gains affect cost at scale for agentic workloads, where a model completing a task in fewer tool calls or tokens directly reduces spend. The relationship between token price and token usage matters when the same task triggers very different call patterns across models.
The full picture
Anthropic released Claude Sonnet 5.5, priced at $2 input / $10 output, the same as Sonnet 5, with a stated claim of up to 30% fewer tokens per task. In code-review workloads, CodeRabbit found the model costs roughly 40% of Sonnet 5 for equivalent reviews, a saving of about 60%, because Sonnet 5 repeatedly read the same files and wrote long deliberations, inflating its token count. In agentic testing, Sonnet 5.5 fixed a multi-bug Python project in three tool calls where Sonnet 5 needed twelve. In a single informal coding session, Sonnet 5.5 completed a 3D brick-model designer app in about 29 minutes versus 45 minutes for Opus 5.5, with Opus 5.5 showing slightly higher fidelity. The test was one run and not a formal benchmark. Pricing also includes $2.50 for cache writes and $0.20 for cache reads.
How it developed
Agentic testing reported Sonnet 5.5 fixed a multi-bug Python project in three tool calls versus twelve for Sonnet 5.
CodeRabbit published benchmarks showing Sonnet 5.5 costs roughly 40% of Sonnet 5 in code-review workloads and completed a coding task 1.5x faster than Opus 5.5 in a single informal run.
Sources
Zvi's AI Roundups
dreasays.substack.com
The daily email
Want this in your inbox?
I send one email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free