Claude Sonnet 5 发布:性能接近 Opus 4.8,删除采样参数,分词器变更致实际费用上涨约30%
英文摘要
Anthropic released Claude Sonnet 5 on June 30, 2026, claiming performance close to Opus 4.8 at lower prices. The model no longer supports sampling parameters temperature, top_p, and top_k; it offers a 1 million token context window and 128,000 max output tokens. Adaptive thinking is on by default. While official pricing is unchanged at $3/$15 per million input/output tokens (with a $2/$10 introductory discount until August 31), a new tokenizer produces approximately 30% more tokens for the same text than Sonnet 4.6—effectively a 30% price increase, with variation across languages. The system card notes that Sonnet 5’s lower cyber capabilities compared to Mythos 5 allow safeguards similar to those for Opus 4.7 and 4.8.
中文摘要
Anthropic 于 2026 年 6 月 30 日发布 Claude Sonnet 5,声称性能接近 Opus 4.8 且价格更低。该模型不再支持 temperature、top_p、top_k 等采样参数;提供 100 万 token 上下文窗口和 12.8 万最大输出 token。自适应思考默认开启。虽然官方定价维持在输入/输出每百万 token 3/15 美元(8 月 31 日前优惠价 2/10 美元),但新的分词器使相同文本产生的 token 数比 Sonnet 4.6 多约 30%,实际相当于涨价约 30%,且不同语言涨幅各异。系统卡说明 Sonnet 5 在网络任务上的能力低于 Mythos 5,因此可沿用 Opus 4.7 和 4.8 的安全措施。
关键要点
Claude Sonnet 5 released with performance close to Opus 4.8 but at lower official prices.
发布 Claude Sonnet 5,性能接近 Opus 4.8,官方定价更低。
Sampling parameters temperature, top_p, and top_k are no longer supported; context window is 1M tokens, max output 128k tokens.
不再支持 temperature、top_p、top_k 采样参数;上下文窗口 100 万 token,最大输出 12.8 万 token。
A new tokenizer produces ~30% more tokens for the same input, effectively increasing costs despite unchanged per-token prices.
新分词器使相同输入产生约 30% 更多 token,尽管单 token 价格不变,实际成本上升。
The system card attributes Sonnet 5’s lower cyber capabilities to allowing safeguards similar to Opus 4.7/4.8.
系统卡称 Sonnet 5 网络能力较弱,可沿用 Opus 4.7/4.8 级别安全措施。