Claude Sonnet 5.5 Ships 30 Percent Faster and Cheaper
Anthropic's own numbers put Sonnet within three points of Opus 5.5 on seven of eight comparison items, which narrows the reason to pay Opus prices for everyday coding and document work.
Key Facts
- Anthropic released Claude Sonnet 5.5 on September 28 US time, the second model in the Claude 5.5 family after Opus 5.5 launched on September 22.
- Claude Sonnet 5.5 scored 70.6 percent on Terminal-Bench 4.0, up from Claude Sonnet 5's 10.3 percent, and 80.1 percent on OSWorld 2.1, up from 57.0 percent.
- On Anthropic's GDPval-AA v2.1 evaluation Claude Sonnet 5.5 scored 1844, two points behind Claude Opus 5.5 at 1846 and 357 points above OpenAI's GPT-6 Sol at 1487.
- Claude Sonnet 5.5 is available on Claude apps, the API, Amazon Web Services, Google Cloud, and Microsoft Azure, with Claude Haiku 5.5 scheduled within the next few weeks.
Reporting from 2 sources: GameBusiness.jp, GIGAZINE.
Anthropic released Claude Sonnet 5.5 on September 28 US time, the second model in the Claude 5.5 family after Opus 5.5 arrived on September 22. Output generation is more than 30 percent faster than Claude Sonnet 5 and cost per task is up to 30 percent lower, with API pricing unchanged at 2 dollars per million input tokens and 10 dollars per million output tokens. On Terminal-Bench 4.0, which measures command-line work, the score rose from 10.3 percent for Sonnet 5 to 70.6 percent. OSWorld 2.1 reached 80.1 percent and Chartography 61.6 percent. On Anthropic's GDPval-AA v2.1 table, Sonnet 5.5 scored 1844, two points behind Opus 5.5 at 1846 and roughly 400 above Sonnet 5. Against OpenAI's GPT-6 Sol, which shares the same API unit price, the knowledge-work gap is large: 1844 to 1487 on GDPval-AA v2.1 and 1811 to 1483 on AA-Briefcase v1.1. The model is available across Claude apps, the API, Amazon Web Services, Google Cloud, and Microsoft Azure. Claude Haiku 5.5 is planned within weeks.
Sonnet 5.5 beats Sonnet 5 on every benchmark Anthropic published, and on Terminal-Bench 4.0 the jump is the headline number: 10.3 percent to 70.6 percent. OSWorld 2.1 climbed from 57.0 percent to 80.1 percent. Chartography went from 15.6 percent to 61.6 percent. The model is also the first Sonnet to clear Pokemon Red using screenshots only.
Anthropic does not sell it as an Opus 5.5 replacement. On GDPval-AA v2.1, which covers 44 occupations, Sonnet 5.5 scored 1844 against Opus 5.5's 1846. On AA-Briefcase v1.1 it scored 1811 against 1822. On OSWorld 2.1 the gap is 1.7 points, on Chartography 2.8 points, on Humanity's Last Exam 3.2 points. Sonnet 5.5 wins Terminal-Bench outright, beating Opus 5.5's 66.4 percent at the xhigh setting. The guidance is to use Opus for complex work requiring sustained judgment and Sonnet for scoped tasks: bug fixes, documents, slides, spreadsheets.
Thinking effort has five levels: low, medium, high, xhigh, max. The default is medium for Claude Code and the Claude app, and high for the API. At low or medium, Sonnet 5.5 beat Sonnet 5's top score at roughly one tenth the cost. At max the advantage narrows. Artificial Analysis measured the Intelligence Index v4.3.2 at 56, three points above Claude Fable 5.1, and flagged high token use at max reasoning, where per-task cost can exceed Opus 5.5.
Early testers include Base44, which built 118 apps at 3.6 iterations per item against Opus 5's 7.7, and Lovable, which saw tool calls fall by one third. Balyasny Asset Management ran 2,441 financial tasks at about 121,000 tokens per response against Sonnet 5's 497,000.
Sonnet 5.5 is the first Sonnet to adopt model switching for high-risk cyber requests, routing exploit generation and penetration testing to Sonnet 5. Normal development and vulnerability checks stay on 5.5. The classifier detecting distillation attempts is new to the Sonnet line. The Cyber Verification Program does not support Sonnet 5.5 at launch.
Synthesized by Yomimono from the 2 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.