Claude Opus 5 在智能指数上以 61 分登顶，成本低于 Fable 5

根据多个基准测试，Anthropic 的 Claude Opus 5 是目前最强大的 AI 模型，性能超过 Fable 5，同时成本更低。

Opus 5 在人工分析智能指数（Artificial Analysis Intelligence Index：https://artificialanalysis.ai/）中得分为 61，该指数结合了涵盖知识工作、编码、科学推理和事实准确性的九项测试。这使其略高于 Claude Fable 5（60）、GPT-5.6 Sol（59）、Kimi K3（57）和 Claude Opus 4.8（56）。人工分析与 Anthropic 合作，在模型公开发布前对其进行了测试。

在编码方面，Claude Opus 5 在“xhigh”模式下与 Claude Code 配合，在人工分析编码指数上并列第一，该指数衡量 AI 模型独立处理编程任务的能力，包括发现和修复漏洞。在 Terminal-Bench v2.1 测试中，该测试评估代理在真正终端环境中作为自主工程师的表现，Opus 5 在“max”模式下得分 89%，与之前的领先者 GPT-5.6 Sol 持平。

在科学推理方面，Opus 5 在 Humanity's Last Exam 中得分 53%，这是一项非常困难的涵盖多个学术领域的知识测试，其成绩与 Fable 5 不相上下。在 CritPt 上，这是来自阿贡国家实验室和伊利诺伊大学香槟分校研究人员的物理基准测试，它再次与 Fable 5 持平，但落后于 GPT-5.6 Sol、GPT-5.5 Pro 和 GPT-5.6 Terra。

事实准确性仍然是薄弱环节：在 AA-Omniscience 测试中，该测试评估模型知识声明的准确性，Opus 5 比 Opus 4.8 提高了 7 分，但仍落后于 Fable 5。Opus 5 在不确定时也更容易作答，使其幻觉率上升 14 个百分点，达到 50%。

Epoch AI（https://epoch.ai/benchmarks/eci?subset-view=graph&subset-tab=Software+engineering&view=graph&tab=release-date）也对 Claude Opus 5 进行了测试。该研究机构给它的整体 Epoch 能力指数评分为 159，略低于 Fable 5 的 161。然而，仅在软件工程基准（SWE-ECI）中，Opus 5 与 Fable 5 并列 161，表现优于 GPT-5.6 Terra 和 Claude Opus 4.8。GPT-5.6 Sol 在两个类别中均领先。

总体而言，这证实了前沿模型之间的竞争非常激烈。没有任何单一模型能够脱颖而出或声称拥有明显优势。这增加了AI模型最终将商品化的观点的可信度：https://the-decoder.com/microsoft-ceo-satya-nadella-says-ai-models-are-getting-commoditized/。

使用Opus 5完成平均Intelligence Index任务的成本为2.03美元，低于Claude Fable 5（fallback模式）为2.75美元的成本。它高于Opus 4.8为1.80美元以及Sonnet 5为1.53美元。然而，在“高”层和“超高”层，Opus 5的表现优于Opus 4.8和Sonnet 5，同时成本更低。Ad DEC_D_Incontent-2

Vals.ai：https://vals.ai/ 使用Vibe Code Bench（一种针对编程任务的基准）测试了Claude Opus 5在所有五个推理层的表现。成绩从“低”层的76.7%上升至“中”层的82%，再到“高”层的89.8%。在两个最高层中，表现稍有下降，“超高”层得分为88.3%，“极高”层得分为88.4%，尽管成本显著更高。Ad

Vals.ai发现，最高层往往产生更复杂的解决方案，同时更容易出现错误。“高”层生成的解决方案则更简单，更可靠地满足要求。

Terminal-Bench 2.1：https://x.com/ValsAI/status/2080817015871451500 显示了类似的模式。“高”层的表现优于“极高”，因为模型在最高层的每次尝试耗时更多，在规定时间内尝试次数较少。这与Anthropic的指南一致，同时“高”层被设置为API和Claude Code的默认层。

Token定价保持为每百万输入Token 5美元，每百万输出Token 25美元。缓存写入成本为每百万Token 6.25美元，缓存寿命为五分钟，而缓存命中仅需每百万Token 0.50美元。

Opus 5在AA-Briefcase：https://artificialanalysis.ai/evaluations/aa-briefcase 基准测试中表现尤为出色，该基准评估AI模型处理典型办公任务的能力，如撰写研究报告、制作演示文稿以及基于数千个输入文件分析电子表格。性能从正确性、分析质量和演示质量进行评分，然后汇总为类似于国际象棋排名的Elo评分。

在最高推理水平下，Opus 5 的 Elo 达到 1720，比 Claude Fable 5（1574）高出整整 146 分。它的三个最高等级（max、xhigh、high）占据前三名。结合 Fable 5、Sonnet 5 和 Opus 4.8，Anthropic 现在掌握了前十名的大部分位置。

每项任务的成本下降了 20%，降至 17.79 美元，而 Fable 5 为 22.30 美元。“xhigh”版本的成本为 14.26 美元，“high”仅为 10.41 美元，不到 Fable 5 的一半。两者在 Elo 排名中仍超过 Fable 5。在中等表现下，Opus 5 的 Elo 达到 1470，仅略低于 GPT-5.6 Sol（max，1505）。在低性能下，其 Elo 为 1223，略低于 GLM-5.2（max，1254）。

Opus 5 最大的进步体现在分析质量上。在“max”级别，其分析质量 Elo 达到 2016，比 Fable 5 高出近 300 分。其 Rubric 通过率（跟踪模型达到预定质量标准的频率）为 58%（“max”）、57.2%（“xhigh”）和 56%（“high”）。呈现质量则是另一回事。Opus 5 的呈现 Elo 为 1628，比 GPT-5.6 Sol 在“max”级别的 1666 分低约 40 分。

更高的性能需要更多时间：“max”下，Opus 5 每项任务需要超过 36 分钟，平均完成 103 次，通过率约比 Opus 4.8 的 24 分钟和 55 次高出 50%。

保持对 AI 的了解。清晰、有用，无废话。

关注 The Decoder，获取 AI 新闻、背景故事和专家分析。

The Decoder：https://the-decoder.com/

Anthropic's Claude Opus 5 is the most capable AI model available today, according to several benchmarks, outperforming Fable 5 while costing less.

Opus 5 scored 61 on the Artificial Analysis Intelligence Index：https://artificialanalysis.ai/, which combines nine tests covering knowledge work, coding, scientific reasoning, and factual accuracy. That puts it just ahead of Claude Fable 5 (60), GPT-5.6 Sol (59), Kimi K3 (57), and Claude Opus 4.8 (56). Artificial Analysis worked with Anthropic to test the model before its public release.

In coding, Claude Opus 5 at "xhigh" paired with Claude Code shares first place on the Artificial Analysis Coding Index, which measures how well AI models handle programming tasks on their own, including finding and fixing bugs. On Terminal-Bench v2.1, which tests agents as autonomous engineers in real terminal environments, Opus 5 scored 89 percent at "max," matching the previous leader GPT-5.6 Sol. Ad

For scientific reasoning, Opus 5 scored 53 percent on Humanity's Last Exam, a very difficult knowledge test covering many academic fields. That ties it with Fable 5. On CritPt, a physics benchmark from researchers at Argonne National Laboratory and UIUC, it again matches Fable 5 but trails GPT-5.6 Sol, GPT-5.5 Pro, and GPT-5.6 Terra. Ad DEC_D_Incontent-1

Factual accuracy remains a weak spot: On AA-Omniscience, which tests the accuracy of a model's knowledge claims, Opus 5 improved by 7 points over Opus 4.8 but still trails Fable 5. Opus 5 also answers more often when it's uncertain, pushing its hallucination rate up 14 points to 50 percent.

Epoch AI：https://epoch.ai/benchmarks/eci?subset-view=graph&subset-tab=Software+engineering&view=graph&tab=release-date has also tested Claude Opus 5. The research institute gave it an overall Epoch Capability Index score of 159, just below Fable 5 at 161. When looking solely at the software engineering benchmarks (SWE-ECI), though, Opus 5 ties with Fable 5 at 161, outperforming GPT-5.6 Terra and Claude Opus 4.8. GPT-5.6 Sol leads in both categories. Ad

Overall, this confirms that the race among frontier models is tight. No single model can pull away or claim a clear advantage. That lends weight to the argument that AI models will eventually become commoditized：https://the-decoder.com/microsoft-ceo-satya-nadella-says-ai-models-are-getting-commoditized/.

The average Intelligence Index task costs $2.03 with Opus 5, less than Claude Fable 5 with fallback at $2.75. It costs more than Opus 4.8 at $1.80 and Sonnet 5 at $1.53. At the "high" and "xhigh" tiers, however, Opus 5 beats both Opus 4.8 and Sonnet 5 while costing less. Ad DEC_D_Incontent-2

Vals.ai：https://vals.ai/ tested Claude Opus 5 across all five reasoning tiers using Vibe Code Bench, a benchmark for programming tasks. Scores climb from 76.7 percent at "low" to 82 percent at "medium" and 89.8 percent at "high." Performance dips at the two highest tiers, with "xhigh" scoring 88.3 percent and "max" scoring 88.4 percent despite much higher costs. Ad

Vals.ai found that the highest tiers tend to produce more complex solutions that contain errors more often. The "high" tier produces simpler solutions that meet the requirements more reliably.

Terminal-Bench 2.1：https://x.com/ValsAI/status/2080817015871451500 shows a similar pattern. The "high" tier beats "max" because the model spends more time on each attempt at the top tier, leaving fewer attempts within the time limit. That matches Anthropic's guidance, with "high" set as the default tier in both the API and Claude Code.

Token pricing stays at $5 per million input tokens and $25 per million output tokens. Cache writes cost $6.25 per million tokens with a five-minute lifetime, while cache hits run just $0.50 per million tokens.

Opus 5 performs especially well on the AA-Briefcase：https://artificialanalysis.ai/evaluations/aa-briefcase benchmark, which measures how well AI models handle typical office tasks like writing research reports, building presentations, and analyzing spreadsheets based on thousands of input files. Performance is scored across correctness, analytical quality, and presentation quality, then rolled into an Elo rating similar to chess rankings.

At max reasoning, Opus 5 reaches an Elo of 1720, a full 146 points ahead of Claude Fable 5 (1574). Its three highest tiers (max, xhigh, high) sweep the top three spots. Combined with Fable 5, Sonnet 5, and Opus 4.8, Anthropic now holds the vast majority of top-10 positions.

Cost per task dropped 20 percent to $17.79, down from $22.30 for Fable 5. The "xhigh" variant costs $14.26, and "high" comes in at just $10.41, less than half of Fable 5. Both still beat Fable 5 in the Elo ranking. At medium performance, Opus 5 hits an Elo of 1470, just behind GPT-5.6 Sol (max, 1505). At low performance, it lands at 1223 Elo, slightly below GLM-5.2 (max, 1254).

Opus 5's biggest gains show up in analytical quality. At "max," it reaches an Analytical Quality Elo of 2016, almost 300 points ahead of Fable 5. Its Rubric Pass Rate, which tracks how often the model meets predefined quality standards, sits at 58 percent ("max"), 57.2 percent ("xhigh"), and 56 percent ("high"). Presentation quality is a different story. Opus 5 scores a Presentation Elo of 1628, about 40 points behind GPT-5.6 Sol at "max" (1666).

More performance takes more time: At "max," Opus 5 needs over 36 minutes per task and averages 103 passes, about 50 percent longer than Opus 4.8 at 24 minutes and 55 passes.

Stay in the loop on AI. Clear, useful, no fluff.

Follow The Decoder for AI news, background stories and expert analyses.

The Decoder：https://the-decoder.com/