Claude Opus 5 发布

2026 年 7 月 24 日，Anthropic 发布 Claude Opus 5。官方给出的定位翻译过来是：接近旗舰智能，价格腰斩。措辞本身值得玩味——"接近"不是"达到"，发布方谨慎地把旗舰位置留给了更贵的产品，而把 Opus 5 放在"最强预算"的位置上。Claude Max 的新默认模型是它，Claude Pro 的最强选项也是它；API 名 claude-opus-5，定价每百万输入 5 美元、输出 25 美元，与 Opus 4.8 持平。

评测数字需要放在厂商语境里读：Frontier-Bench v0.1 约为 Opus 4.8 的两倍，OSWorld 2.0 以约旗舰三分之一的成本取得更好成绩，ARC-AGI 3 约为第二名模型的三倍。这些是 Anthropic 自己发布的结果，不是独立认证；它们说明方向——能力与成本的组合——而不是放之四海的标准答案。同一家族内还给出了标准与 Fast 两档：Fast 约 2.5 倍速度、2 倍价格，把"要不要多花钱换时间"变成每次调用都能做的选择。

更值得注意的反而是两个 beta 级工程能力。一是对话中途切换工具：长流程 Agent 不必在开头就押定工具集，模型可以边推进边换。二是安全拦截后的自动回退：当安全分类器拦截某个请求时，API 可以按配置把它路由到另一个模型。过去，被拒答的请求通常直接失败，开发者只能自行处理；现在它成为一个可配置的接口行为。这不等于安全问题的解决，而是把"模型说不"之后的流程控制权交还给了工程团队。

System Card 同期发布，其中有一段容易被忽略的自我设限：Opus 5 没有做网络攻防训练，漏洞利用能力仍明显落后于同期专攻该方向的模型。在普遍高调的宣传环境里，这样一句话的分量不亚于任何评测分数——它划出了这条产品线的能力边界，也提醒买家：旗舰档的选择从来不是"哪个更强"，而是"强在哪里、弱在哪里、代价几何"。

Opus 5 因此像一场关于定价权的宣言：前沿竞争的重心，正在从单点能力向"能力与成本的组合曲线"移动。对用它跑长流程编码任务的团队来说，最直接的变化是账单；对行业来说，变化的是默认假设——最强的模型，不一定再是最贵的选择。

On July 24, 2026, Anthropic released Claude Opus 5. The official positioning, translated plainly: near-frontier intelligence at roughly half the price. The wording deserves attention—"near" rather than "at." Anthropic kept the top slot for its more expensive line and placed Opus 5 at the strongest-budget position. It became the default model for Claude Max and the strongest option in Claude Pro; the API name is claude-opus-5, priced at $5 per million input tokens and $25 per million output, unchanged from Opus 4.8.

Benchmark numbers should be read inside their vendor context: roughly double Opus 4.8 on Frontier-Bench v0.1, better results than the flagship at about one-third the cost on OSWorld 2.0, and about three times the runner-up on ARC-AGI 3. These are Anthropic's own results, not independent certification. They indicate a direction—the combination of capability and cost—rather than a universal ranking. The family also offers standard and Fast tiers: Fast runs about 2.5x faster at 2x price, turning "pay more to save time" into a per-call decision.

Two beta-level engineering features deserve more attention. First, mid-conversation tool switching: a long agent loop no longer commits to its toolset up front. Second, automatic fallback after safety interception: when a safety classifier blocks a request, the API can route it to another model by configuration. Previously, a declined request simply failed and developers handled it themselves; now it is a configurable interface behavior. This does not solve safety—it hands the engineering team control over what happens after a model says no.

The System Card contains a sentence easy to skip: Opus 5 received no cyber-offense training, and its exploitation capability still trails models specialized in that area. In a loud launch season, such a line carries as much weight as any score. It draws the boundary of this product line and reminds buyers that flagship selection was never "which is stronger" but "strong where, weak where, at what cost."

Opus 5 reads like a statement about pricing power. The center of frontier competition is moving from peak capability toward the capability-cost curve. For teams running long coding loops on it, the immediate change is the bill. For the industry, the change is the default assumption: the strongest model no longer has to be the most expensive one.