Anthropic's Cheaper AI Now Nearly Matches Its Best Model

Anthropic has released Claude Sonnet 5.5, the middle-tier version of its AI assistant. On an independent scoreboard called the Artificial Analysis Intelligence Index, it landed at number two — just two points behind Anthropic's flagship Opus 5.5. In plain terms: the "good enough and cheaper" option now performs almost like the expensive one. The price per word stayed the same as the previous Sonnet, which is unusual when performance jumps this much.

The headline skill is doing computer chores on its own. On a test where the AI operates a command-line terminal like a junior engineer, Sonnet 5.5 scored 64%, up from 14% for the previous Sonnet and slightly ahead of its own flagship. It also matched Opus on office-style knowledge work. That matters because "AI that takes actions" (called agents) is what turns a chatbot into something that actually files your reports, updates your spreadsheets, or runs your research.

Here's the catch, and it's a money one. To hit those scores, Sonnet 5.5 is the chattiest model ever measured by Artificial Analysis — about 193,000 output words per task, roughly seven times what rival GPT-6 Astra uses for the same job. Since you pay per word, a typical task costs about $7.60, around 50% more than the previous Sonnet. Cheap-sounding prices can hide an expensive habit.

There is a second trade-off: facts. Sonnet 5.5 answers factual questions correctly 54% of the time versus 66% for Opus, and trails on scientific reasoning. Interestingly, it makes up fewer things when unsure — a 47% hallucination rate against Opus's 59%. The practical takeaway: this is a strong, affordable choice for multi-step computer tasks where you can check the work, but not your final authority on facts. Anthropic also fixed a pre-release bug affecting structured responses.