Just Now, Anthropic Released Sonnet 5, Performance Close to Opus 4.8, but Not Necessarily Cheaper
Anthropic has officially released Claude Sonnet 5, describing it as the most "agentic" Sonnet model to date. It can plan, use tools like browsers and terminals, and autonomously perform tasks at a level previously requiring larger, more expensive models. Performance in reasoning, tool use, programming, and knowledge work has significantly improved compared to Sonnet 4.6, now approaching that of Opus 4.8.
Evaluation results indicate that Sonnet 5, at medium "effort" levels, offers better cost efficiency than its predecessor. At higher effort levels, its performance in some tasks can match Opus 4.8. In terms of safety, Sonnet 5 shows improved rates of refusing malicious requests and resisting prompt injection attacks compared to Sonnet 4.6, though it has a slightly higher rate of policy-violating behavior than Opus 4.8 and Mythos Preview. Its cybersecurity capabilities remain weaker than those models.
Notably, Sonnet 5 uses a new tokenizer. The same text input now results in approximately 1.0 to 1.35 times more tokens, depending on content. To offset this, Anthropic offers a promotional launch price until August 31, 2026, at $2 per million input tokens and $10 per million output tokens. The standard pricing will be $3/$15 per million tokens thereafter. However, some external analysis suggests that due to increased token usage, the actual cost per task for Sonnet 5 may be higher than both Sonnet 4.6 and Opus 4.8.
marsbit9 хв тому
