Claude Sonnet 5.5 Cuts Task Costs by 30% Without Losing Power

Anthropic launched Claude Sonnet 5.5, positioning its mid-tier model as a direct alternative to higher-priced systems. According to TechCrunch, the company designed the release specifically for everyday office tasks and coding workflows.
How does Sonnet 5.5 compare to Opus 5.5 on performance?
The Decoder points out that Sonnet 5.5 nearly matches the flagship Opus 5.5 on knowledge-work benchmarks like GDPval-AA, scoring 1,844 points against Opus's 1,846. In coding evaluations such as Terminal-Bench 4.0, the model jumps from its predecessor's 10.3% to 70.6%.
What drives the 30% cost reduction in practice?
While token prices per million remain identical to Sonnet 5 at $2 for inputs and $10 for outputs, The Decoder notes that the model burns significantly fewer tokens per task. This efficiency cuts effective operating expenses by up to 30% while increasing generation speed by over 30%.
Are there any downsides to the new effort settings?
The Decoder highlights an odd quirk at maximum reasoning effort. On FrontierCode, Sonnet 5.5 actually scored lower at the 'Max' setting than at 'Xhigh' because sub-agents triggered unnecessary code reviews that led to timeouts.
Sources
Frequently asked questions
- What is the price per million tokens for Claude Sonnet 5.5?
- Anthropic charges $2 for input tokens, $10 for output tokens, and $0.20 for cache reads, keeping the nominal rates equal to Sonnet 5.
- Where is Claude Sonnet 5.5 available?
- The model is available across Amazon Web Services, Google Cloud, and Microsoft Azure, as well as directly through the Claude Platform.
- Why does the 'Max' effort setting perform worse on some tests?
- At maximum reasoning effort, the model frequently spawns sub-agents for code reviews, which can cause timeouts and lower benchmark scores.
AI Briefing in 3 Minutes
Get what actually matters in AI & tech delivered free to your inbox. Zero spam, 1-click unsubscribe.
Comments
0 commentsDeixe seu comentário
Be the first to comment.
Continue Reading

Meta Enterprise Platform: What to Expect from Meta's B2B Pivot
Meta is moving into enterprise AI. Here is what the new platform means for your business operations and customer support.

GPT-6 Astra: visual diagnostics leap to 80% accuracy
GPT-6 Astra reaches 80% accuracy in visual error diagnostics, jumping past previous benchmarks while facing a three-minute processing bottleneck.

Microsoft Unifies Copilot Into a Single Work App
Microsoft combines chat, coding, and autonomous agents into a single Copilot app, introducing usage-based billing for enterprise tasks.