Anthropic 发布 Claude Opus 5：前沿级智能体编程与计算机使用能力，Opus 定价不变

今天，Anthropic 发布了 Claude Opus 5：https://www.anthropic.com/news/claude-opus-5。它取代了 Claude Opus 4.8，成为 Opus 级旗舰。定价维持不变：每百万输入令牌 5 美元，每百万输出令牌 25 美元。

Anthropic 团队将 Opus 5 定位为以一半价格接近 Claude Fable 5 的智能水平。它现在是 Claude Max 的默认模型，也是 Claude Pro 上最强的模型。

在进行任何基准测试前，三个变化非常重要：

模型 ID 为 claude-opus-5。上下文为 100 万令牌，既是默认值也是最大值，没有更小的版本。在同步 Messages API 上最大输出为 128k 令牌。Message Batches API 在 output-300k-2026-03-24 测试版标题下可达到 300k 令牌。可缓存的最小提示从 1024 令牌降至 512 令牌。

在 FrontierBench v0.1：https://www.frontierbench.ai/ 上，一个 74 任务的 Terminal-Bench 2.1 继任者，Opus 5 在最大努力下得分为 43.3%。Opus 4.8 得分为 18.7%。Fable 5 达到 33.7%，GPT-5.6 Sol 达到 37.5%。在 xhigh 努力下，Opus 5 达到 44.4% 的平均奖励，这是它的最佳成绩。

该运行中的一个细节值得注意。Opus 5 的安全分类器标记并拒绝了 5% 的 API 调用，占 4% 的试验。Fable 5 的分类器标记了 42% 的调用，占 26% 的试验。

Opus 5 在 SWE-bench Verified 上得分 96.0%，在 SWE-bench Pro 上得分 79.2%：https://arxiv.org/abs/2509.16941。Fable 5 在 Pro 上仍略胜一筹，为 80.0%。在 SWE-bench Multimodal 上提升更大，从 38.4% 提升到 59.4%。

自主代理的数字是最明显的胜利。Opus 5 在 OSWorld 2.0：https://arxiv.org/abs/2606.29537 上达到 70.57%，而 Opus 4.8 为 55.7%。在 Zapier AutomationBench：https://arxiv.org/abs/2604.18934 上，它得分 26.0%，而 Opus 4.8 为 17.0%，Fable 5 为 17.4%。在中等努力下，它仍以每任务 0.89 美元得分 24%。

在 Artificial Analysis GDPval-AA v2：https://artificialanalysis.ai/evaluations/gdpval-aa 上，Opus 5 占据排行榜前两名，ELO 分别为 1861 和 1827。xhigh 设置击败了所有其他模型，同时使用的输出令牌比最大值少 25%。

Anthropic 在没有工具或代理辅助的情况下，让 Opus 5 对 IMO 2026 的全部六个问题进行了提示。一个由三名模型组成的评审小组对生成的 24 个解答全部评分为正确。人类专家独立对每个问题的一个预先指定解答评分为 7/7。最终的 42/42 是金牌水平，高于 29/42 的及格线。

ARC 奖基金会报告称，在高强度努力下，Opus 5 在 ARC-AGI-3 测试中的验证分数为 30.16%。这大约是此前排行榜上最好成绩的四倍。GPT-5.6 Sol 的成绩为 7.78%，Opus 4.8 的成绩为 1.52%。Opus 5 在最大努力下的成绩在发布时尚不可用。

在人类的最后考试中：https://arxiv.org/abs/2501.14249，Opus 5 在没有工具的情况下得分为 56.3%，使用工具后得分为 64.7%。

多模态部分带来了一个实际的教训。代理工具的使用比单纯的适应性思考更具成本效益地扩展了测试时的计算能力。

在 Chartography 上，Opus 5 无工具得分为 29.6%，使用一个容器和图像裁剪工具得分为 83.0%。在 BenchCAD：https://arxiv.org/abs/2605.10865 Vision2Code 上，体素 IoU 从 0.366 升至 0.821。使用工具后，这一成绩远超 Claude Mythos 5 的 0.678。

Anthropic 并未让 Opus 5 参与网络安全任务训练。然而，该能力依然提升，作为通用能力提升的副产品。

在 ExploitBench：https://arxiv.org/abs/2605.14153 上，Opus 5 在 AutoNudge 模块捕获了平均 10.14 个能力标志。它生成了 99 个完整任意代码执行漏洞。Mythos 5 生成了 132 个。在 OSS-Fuzz 上，Opus 5 对 79.4% 的目标得分非零。Mythos 5 大约达到 80%。但 Opus 5 完成了 4 个完整漏洞，而 Mythos 5 完成 13 个。

这种差距定义了安全保护设计。Opus 5 在发现漏洞方面几乎和 Mythos 5 一样强，但在利用漏洞方面明显落后。因此，Anthropic 在源代码漏洞发现方面解锁了，但基于二进制的扫描、渗透测试和漏洞生成依然被阻止。预期分类器介入的频率比 Fable 5 少约 85%。防御者可以申请网络验证计划：https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude-opus-and-sonnet。

英国AISI在三个网络演练场上对早期检查点进行了测试，每次尝试处理1亿个令牌。Opus 5在10次尝试中，有8次端到端解决了《最后一批》。它未能解决更高难度的《终身监禁》演练。它达到了第23步中的第22步，比任何测试过的模型都更进一步。

根据RSP，Anthropic将Opus 5视为具有CB-1能力，但不具备CB-2能力。它采用了用于Opus 4.8的相同ASL-3保护措施。AI研发门槛未被突破。

在Gray Swan间接提示注入基准测试中，攻击者在15次尝试内的成功率从Opus 4.8的5.5%降至2.0%。Mythos 5的成功率为2.6%，而GPT-5.6 Sol为20.0%。

在通过Claude Cowork运行的浏览器环境中，攻击成功率从Opus 4.8的31.5%降至3.70%。该数据是在未应用任何安全措施的情况下统计的。启用自动模式后，在所有129个环境中达到0%的成功率。

来源：Anthropic发布文章：https://www.anthropic.com/news/claude-opus-5, Claude Opus 5系统卡：https://www.anthropic.com/claude-opus-5-system-card, Claude Opus 5新特性：https://platform.claude.com/docs/en/about-claude/models/whats-new-opus-5, TechCrunch：https://techcrunch.com/2026/07/24/anthropic-launches-opus-5/, 以及CodeRabbit独立评测：https://www.coderabbit.ai/blog/opus-5-model-review

Asif Razzaq是Marktechpost Media Inc.的首席执行官。作为一位具有远见的企业家和工程师，Asif致力于利用人工智能的潜力造福社会。他最近的项目是推出人工智能媒体平台Marktechpost，该平台以对机器学习和深度学习新闻的深入报道而脱颖而出，既技术可靠又易于广大受众理解。该平台每月浏览量超过200万，显示出其在观众中的受欢迎程度。

Today, Anthropic released Claude Opus 5：https://www.anthropic.com/news/claude-opus-5. It replaces Claude Opus 4.8 as the Opus-tier flagship. Pricing is unchanged at $5 per million input tokens and $25 per million output tokens.

The Anthropic team positions Opus 5 as approaching the intelligence of Claude Fable 5 at half the price. It is now the default model on Claude Max and the strongest model on Claude Pro.

Three changes are quite important before any benchmark does:

The model ID is claude-opus-5 . Context is 1M tokens as both default and maximum, with no smaller variant. Maximum output is 128k tokens on the synchronous Messages API. The Message Batches API reaches 300k with the output-300k-2026-03-24 beta header. The minimum cacheable prompt drops to 512 tokens, down from 1,024.

On FrontierBench v0.1：https://www.frontierbench.ai/, a 74-task successor to Terminal-Bench 2.1, Opus 5 scored 43.3% at max effort. Opus 4.8 scored 18.7%. Fable 5 reached 33.7% and GPT-5.6 Sol reached 37.5%. At xhigh effort Opus 5 reaches 44.4% mean reward, its best result.

One detail from that run is worth noting. Opus 5 safety classifiers flagged and refused 5% of API calls, across 4% of trials. Fable 5 classifiers flagged 42% of calls across 26% of trials.

Opus 5 scored 96.0% on SWE-bench Verified and 79.2% on SWE-bench Pro：https://arxiv.org/abs/2509.16941. Fable 5 still edges it on Pro at 80.0%. On SWE-bench Multimodal the jump is larger, from 38.4% to 59.4%.

The agentic numbers are the clearest wins. Opus 5 reached 70.57% on OSWorld 2.0：https://arxiv.org/abs/2606.29537 against 55.7% for Opus 4.8. On Zapier AutomationBench：https://arxiv.org/abs/2604.18934 it scored 26.0%, against 17.0% for Opus 4.8 and 17.4% for Fable 5. At medium effort it still scores 24% at $0.89 per task.

On Artificial Analysis GDPval-AA v2：https://artificialanalysis.ai/evaluations/gdpval-aa, Opus 5 takes the top two leaderboard spots at ELO 1861 and 1827. The xhigh setting beats every other model while using 25% fewer output tokens than max .

Anthropic prompted Opus 5 on all six IMO 2026 problems without tools or an agent harness. A three-model judge panel scored all 24 generated solutions correct. Human experts independently graded one pre-specified solution per problem at 7/7. The final 42/42 is gold-medal level, above the 29/42 cutoff.

The ARC Prize Foundation reports a verified 30.16% on ARC-AGI-3 at high effort. That is roughly four times the best previously reported leaderboard score. GPT-5.6 Sol reached 7.78% and Opus 4.8 reached 1.52%. Opus 5 results at max effort were unavailable at release.

On Humanity's Last Exam：https://arxiv.org/abs/2501.14249, Opus 5 scored 56.3% without tools and 64.7% with them.

The multimodal section carries a practical lesson. Agentic tool use scales test-time compute more cost-effectively than adaptive thinking alone.

On Chartography, Opus 5 scored 29.6% without tools and 83.0% with a container and an image-cropping tool. On BenchCAD：https://arxiv.org/abs/2605.10865 Vision2Code, voxel IoU moves from 0.366 to 0.821. With tools, that beats Claude Mythos 5 at 0.678 by a wide margin.

Anthropic did not train Opus 5 on cybersecurity tasks. Capability rose anyway, as a byproduct of general capability gains.

On ExploitBench：https://arxiv.org/abs/2605.14153, Opus 5 captured 10.14 mean capability flags in the AutoNudge arm. It produced 99 full arbitrary-code-execution exploits. Mythos 5 produced 132. On OSS-Fuzz, Opus 5 scored non-zero on 79.4% of targets. Mythos 5 reached roughly 80%. But Opus 5 completed 4 full exploits to Mythos 5's 13.

That gap defines the safeguard design. Opus 5 is nearly as strong as Mythos 5 at finding vulnerabilities, and substantially behind at exploiting them. So Anthropic unblocked vulnerability finding in source code. Binary-based scanning, penetration testing and exploit generation stay blocked. Classifiers are expected to intervene around 85% less often than on Fable 5. Defenders can apply to the Cyber Verification Program：https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude-opus-and-sonnet.

UK AISI tested early checkpoints on three cyber ranges at 100M tokens per attempt. Opus 5 solved 'The Last Ones' end-to-end in 8 of 10 attempts. It did not solve the harder 'Doing Life' range. It reached step 22 of 23, further than any model tested.

Under the RSP, Anthropic treats Opus 5 as having CB-1 capabilities but not CB-2. It applies the same ASL-3 protections used for Opus 4.8. The AI R&D threshold is not crossed.

On the Gray Swan indirect prompt injection benchmark, attacker success within 15 attempts fell from 5.5% on Opus 4.8 to 2.0%. Mythos 5 sits at 2.6% and GPT-5.6 Sol at 20.0%.

In browser environments run through Claude Cowork, attack success dropped from 31.5% on Opus 4.8 to 3.70%. That figure is with no safeguards applied. With auto mode enabled, it reached 0% across all 129 environments.

Sources: Anthropic launch post：https://www.anthropic.com/news/claude-opus-5, Claude Opus 5 System Card：https://www.anthropic.com/claude-opus-5-system-card, What's new in Claude Opus 5：https://platform.claude.com/docs/en/about-claude/models/whats-new-opus-5, TechCrunch：https://techcrunch.com/2026/07/24/anthropic-launches-opus-5/, and CodeRabbit independent review：https://www.coderabbit.ai/blog/opus-5-model-review

Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.