Anthropic 发布 Claude 3 家族

到 2024 年初，大模型的世界已经分成两派：GPT-4 霸占着性能王座，而 Anthropic 的 Claude 虽然以安全、负责任著称，却总被质疑「性能是不是差一截」。对一家把「对齐」写进使命的公司来说，这种质疑是一种难言的憋屈。Anthropic 需要一次证明自己的机会。

2024 年 3 月 4 日，这个机会来了。Claude 3 家族一次发布三档：旗舰 Opus、均衡 Sonnet、轻量 Haiku。三档全部支持视觉输入和 20 万 token 的上下文。最让人意外的是 Opus——它在多项基准测试上超过了 GPT-4，这是 Claude 系列第一次在性能上真正杀进第一梯队。

Claude 3 带来的改变不只是性能数字。它第一次把「分级」做成了产品哲学：同一个模型家族，按能力、速度、成本分成三档，用户按自己的场景和预算选。Haiku 便宜到可以大量调用，Opus 强到可以承担旗舰任务。这种「一代多档」的设计，后来成为整个行业的通用做法。

对 Anthropic 而言，Claude 3 是一次自我证明。它证明一家把安全放在首位的公司，同样能做出顶级的模型；也证明「安全」和「能力」不必互斥。发布后，Claude 在企业市场，尤其是对安全合规敏感的行业里迅速站稳，Anthropic 也从「有理想的挑战者」变成「能打的正规军」。

回看 Claude 3，它的分量在于双重意义：对外，它让 Anthropic 第一次在性能上赢得尊重；对内，它用「分档」重新定义了模型产品线的玩法。此后每家公司发旗舰，几乎都要带上几个档位。那个 2024 年春天，GPT-4 不再是唯一的选择——而 Anthropic 的这场翻身仗，也成了 AI 竞争史上最值得回味的故事之一。

By early 2024, the large-model world had split into two camps: GPT-4 held the performance throne, while Anthropic's Claude, praised for being safe and responsible, was constantly questioned with "but is it a step behind on capability?" For a company that wrote alignment into its mission, that doubt was an irksome burden. Anthropic needed a chance to prove itself.

On March 4, 2024, the chance arrived. The Claude 3 family launched all at once in three tiers: flagship Opus, balanced Sonnet, and lightweight Haiku. All three supported vision input and 200K-token context. The biggest surprise was Opus—it beat GPT-4 on many benchmarks, the first time a Claude model truly broke into the first tier on performance.

Claude 3 changed more than the numbers. It made "tiering" a product philosophy for the first time: one family, divided by capability, speed, and cost, so users could choose by scenario and budget. Haiku was cheap enough for massive calling; Opus was strong enough for flagship work. This "one generation, many tiers" design later became standard across the industry.

For Anthropic, Claude 3 was self-validation. It proved a company that put safety first could still build top-tier models, and that "safety" and "capability" need not be mutually exclusive. After the release, Claude rapidly took hold in the enterprise market, especially in safety- and compliance-sensitive industries, turning Anthropic from an "idealistic challenger" into a serious force.

Looking back, Claude 3's weight is twofold: externally, it won Anthropic respect on performance for the first time; internally, it redefined how model product lines are designed with its tiering. Since then, nearly every flagship launch comes with tiers. In that spring of 2024, GPT-4 was no longer the only choice—and Anthropic's comeback became one of the most memorable stories in AI competition.