Anthropic 发布 Claude Opus 5：以 Fable 5 一半的 token 价格实现接近其性能

Anthropic的新旗舰模型Claude Opus 5在编码和知识工作方面获得了最高分，同时在使用一半的令牌费率下接近Claude Fable 5的表现。

100万令牌的上下文窗口和令牌费率保持不变：https://platform.claude.com/docs/en/about-claude/pricing。与其前身Opus 4.8一样，Opus 5每百万输入令牌收费5美元，每百万输出令牌收费25美元。新推出的快速模式提高了2.5倍的速度，但价格翻倍。广告

如果不考虑令牌效率，令牌费率并不能说明全部情况。Opus 4.7每个任务的成本比Opus 4.6高出30%到40%：https://the-decoder.com/first-token-counts-reveal-opus-4-7-costs-significantly-more-than-4-6-despite-anthropics-flat-pricing/，尽管两者的基础费率相同。最近Claude Sonnet 5也出现了类似的情况：https://the-decoder.com/claude-sonnet-5-continues-anthropics-pattern-of-hiding-price-increases-behind-unchanged-token-rates/。更新：早期独立基准测试支持Anthropic的声明：https://the-decoder.com/anthropics-claude-opus-5-costs-well-below-fable-5-while-matching-or-beating-it-across-most-benchmarks/。Opus 5成本低于Fable 5，同时在很多情况下性能更佳，也在价格和性能上超越了Opus 4.8和Sonnet 5。广告 DEC_D_Incontent-1

用户可以通过五种名为低、中、高、极高和最大努力的设置在性能和令牌使用之间进行权衡。Anthropic表示，Opus 5在每个努力级别上都比其前身更具价值。

在其提示指南中：https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5，Anthropic建议广泛使用“低”和“中”设置。公司表示，这些设置能在使用较少令牌和降低延迟的情况下提供良好结果，同时超越了早期Opus模型的相同设置。在编码和代理任务中，Anthropic仍建议从“极高”设置开始。广告

在两个基准测试中，Opus 5在最大努力设置下的得分略低于第二高努力设置，尽管成本更高。这一下降出现在Frontier-Bench v0.1和人工分析编码代理指数上。

更新：根据 FrontierCode 开发者 Cheng-Yuan Lee：https://x.com/cl571128/status/2080783750456311836，这种行为是预期的。基准测试不仅评估正确性，还评估模型是否仅对代码进行最小必要修改。在更高的努力等级下，Opus 5 即使只需要小修，也可能重构周围的代码。在一个示例中，Opus 5 在最低设置下仅修改了一个字符，而在“xhigh”设置下它重构了整个周围代码。基准测试将这种非必要的更改计为错误。Ad DEC_D_Incontent-2

根据 Anthropic 自己的基准测试，Opus 5 在多个评估中创下记录。在 Frontier-Bench v0.1 上，该模型在自主终端编码中达到 43.3%，远超 Fable 5（33.7%）、GPT-5.6 Sol（34.4%）和其前身 Opus 4.8（21.1%）。在知识工作基准 GDPval-AA：https://the-decoder.com/openai-says-top-ai-models-are-reaching-expert-territory-on-real-world-knowledge-work/ v2 上，Opus 5 以 1861 的 Elo 分领先，领先 Fable 5（1747）和 GPT-5.6 Sol（1736）。Ad

Opus 5 并不是无处取胜。在通过 DeepSWE v1.1 进行自主编码时，GPT-5.6 Sol 以 72.7% 居首，其次是 Fable 5（69.7%）和 Opus 5（68.8%）。在健康任务和法律基准测试中，Fable 5 和 Mythos 5 分别表现优于 Opus 5。

ARC-AGI-3 的结果可能是 Anthropic 基准测试中最大惊喜和异常值。该基准测试衡量没有记忆模式的新问题解决能力：https://the-decoder.com/even-the-latest-ai-models-make-three-systematic-reasoning-errors-arc-agi-3-analysis-shows/，Opus 5 得分为 30.2%。Opus 4.8 仅为 1.5%，GPT-5.6 Sol 为 7.8%。这比下一个最佳模型高出近 4 倍。该基准测试没有 Fable 5 的结果，尚不清楚这种大幅领先是否会在实际使用中体现。

在网络安全任务中，Opus 5 落后于 Mythos 5。Anthropic 表示，它故意没有对 Opus 5 进行网络任务训练，其前身也是如此。该模型在发现漏洞方面接近 Mythos 5，但在利用这些漏洞上表现明显较差。

Anthropic还表示，Opus 5在生成视觉输出和分析视觉内容（如图表和示意图）方面有了改进。

Anthropic将Opus 5描述为在自我检查和通过迭代改进工作方面能力大幅提升。在Frontier-Bench的一项任务中，Opus 5收到了一个机器零件的图纸，并且必须在FreeCAD中创建一个三维模型。难点在于，该模型故意没有直接查看图纸的方式。Opus 5通过编写自己的计算机视觉流程，从原始像素中提取几何信息，然后重建了完整的机器零件。Anthropic表示，在五次尝试后，没有其他模型能完成该任务。

Opus 5还处理了一个流行开源包管理器的实际漏洞。据Anthropic称，它找到根本原因并修复了社区补丁遗漏的边缘情况。而一个竞争模型仅修复了表面症状，然后报告漏洞已解决。

据报道，一家交易公司的工程师使用Opus 5在一次会话中为一个新交易所构建了市场数据馈送。之前的模型即使有详细计划，也无法完成该任务。

Anthropic表示，Opus 5的安全设置允许进行源代码漏洞研究，但会阻止基于二进制的漏洞扫描、渗透测试和漏洞利用生成。网络分类器触发频率比Fable 5低约85%。Fable 5的频繁干预曾受到严重批评：https://the-decoder.com/claude-fable-5-anthropic-admits-wrong-tradeoff-after-invisibly-throttling-rival-ai-researchers/。在Claude.ai、Claude Code和Claude Cowork中被阻止的请求默认回退至Opus 4.8，这与Fable 5使用的方法相同。

Anthropic称Opus 5是当前最强大的通用科学研究模型。它在所有生命科学评估中均优于Opus 4.8，尤其在有机化学（从光谱数据推导分子结构提高了10.2个百分点）和蛋白质相关任务（提高了7.7个百分点）方面表现突出。

与 Opus 5 一起发布，Anthropic 正在推出两个测试功能。Claude 平台上的“对话中工具更换”允许开发者在对话过程中更换可用工具，而不会使提示缓存失效。API 路径上的“自动回退”会将被阻止的请求自动切换到不同的模型。

保持对 AI 的了解。清晰、有用，没有废话。

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 new flagship model, Claude Opus 5, posts top scores in coding and knowledge work while approaching Claude Fable 5's performance at half its token rates.

The 1 million-token context window and token rates remain unchanged：https://platform.claude.com/docs/en/about-claude/pricing. Like its predecessor Opus 4.8, Opus 5 costs $5 per million input tokens and $25 per million output tokens. A new Fast Mode increases speed by 2.5x but doubles the price. Ad

Token rates don't tell the full story without factoring in token efficiency. Opus 4.7 ended up costing 30 to 40 percent more per task than Opus 4.6：https://the-decoder.com/first-token-counts-reveal-opus-4-7-costs-significantly-more-than-4-6-despite-anthropics-flat-pricing/, even though both models had the same base rates. A similar pattern showed up recently with Claude Sonnet 5：https://the-decoder.com/claude-sonnet-5-continues-anthropics-pattern-of-hiding-price-increases-behind-unchanged-token-rates/. Update: Early independent benchmarks back up Anthropic's claims：https://the-decoder.com/anthropics-claude-opus-5-costs-well-below-fable-5-while-matching-or-beating-it-across-most-benchmarks/. Opus 5 costs less than Fable 5 while often performing better, and it also beats both Opus 4.8 and Sonnet 5 on price and performance. Ad DEC_D_Incontent-1

Users can trade off performance against token use through five effort settings called low, medium, high, xhigh, and max. Anthropic says Opus 5 offers better value than its predecessor at every effort level.

In its prompting guide：https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5, Anthropic recommends making broad use of the "low" and "medium" settings. The company says they deliver good results with a fraction of the token use and latency while beating the same settings on earlier Opus models. For coding and agentic tasks, Anthropic still recommends starting with "xhigh." Ad

Opus 5 scores slightly worse at the max effort setting than at the second-highest setting on two benchmarks, despite costing more. The drop appears on Frontier-Bench v0.1 and the Artificial Analysis Coding Agent Index.

Update: According to FrontierCode developer Cheng-Yuan Lee：https://x.com/cl571128/status/2080783750456311836, this behavior is expected. The benchmark doesn't just grade correctness but also whether a model makes only the minimum necessary changes to the code. At higher effort levels, Opus 5 tends to refactor surrounding code even when only a small fix is needed. In one example, Opus 5 on the lowest setting changed exactly one character, while on "xhigh" it refactored the entire surrounding code. The benchmark counts such unsolicited changes as errors. Ad DEC_D_Incontent-2

According to Anthropic's own benchmarks, Opus 5 sets records across several evaluations. On Frontier-Bench v0.1, the model hits 43.3 percent on agentic terminal coding, beating Fable 5 (33.7 percent), GPT-5.6 Sol (34.4 percent), and its predecessor Opus 4.8 (21.1 percent) by wide margins. On the knowledge work benchmark GDPval-AA：https://the-decoder.com/openai-says-top-ai-models-are-reaching-expert-territory-on-real-world-knowledge-work/ v2, Opus 5 leads with an Elo score of 1,861, ahead of Fable 5 (1,747) and GPT-5.6 Sol (1,736). Ad

Opus 5 doesn't win everywhere. On agentic coding via DeepSWE v1.1, GPT-5.6 Sol leads with 72.7 percent, followed by Fable 5 (69.7 percent) and Opus 5 (68.8 percent). On health tasks and legal benchmarks, Fable 5 and Mythos 5 outperform Opus 5, respectively.

The ARC-AGI-3 result is likely the biggest surprise and outlier in Anthropic's benchmarks. The benchmark measures novel problem-solving without memorized patterns：https://the-decoder.com/even-the-latest-ai-models-make-three-systematic-reasoning-errors-arc-agi-3-analysis-shows/, and Opus 5 scores 30.2 percent. Opus 4.8 managed just 1.5 percent, GPT-5.6 Sol hit 7.8 percent. That's nearly a 4x gap over the next-best model. There's no Fable 5 result for this benchmark, and it's unclear whether such a large lead on a test will show up in actual use.

Opus 5 falls behind Mythos 5 on cybersecurity tasks. Anthropic says it deliberately didn't train Opus 5 on cyber tasks, as was also the case with its predecessor. The model comes close to Mythos 5 at finding vulnerabilities but performs much worse when asked to exploit them.

Anthropic also says Opus 5 has improved at generating visual outputs and analyzing visual content such as charts and diagrams.

Anthropic describes Opus 5 as much better at checking its own work and improving it through iteration. In a Frontier-Bench task, Opus 5 received a drawing of a machine part and had to create a 3D model in FreeCAD. The catch was that the model intentionally had no way to view the drawing directly. Opus 5 responded by writing its own computer vision pipeline to extract the geometry from raw pixels, then reconstructed the complete machine part. No other model solved this task after five attempts, Anthropic says.

Opus 5 also worked on a real bug in a popular open-source package manager. According to Anthropic, it found the root cause and fixed an edge case that the community patch had missed. A competing model fixed only the surface symptom before reporting the bug as resolved.

An engineer at a trading firm reportedly used Opus 5 to build a market data feed for a new exchange in one session. Previous models couldn't complete the task, even with detailed plans.

Opus 5's safety setup allows source code vulnerability research but blocks binary-based vulnerability scanning, penetration testing, and exploit generation, according to Anthropic. The cyber classifiers trigger about 85 percent less often than on Fable 5. Fable 5's frequent interventions drew heavy criticism：https://the-decoder.com/claude-fable-5-anthropic-admits-wrong-tradeoff-after-invisibly-throttling-rival-ai-researchers/. Blocked requests in Claude.ai, Claude Code, and Claude Cowork default to Opus 4.8 as a fallback, the same approach used with Fable 5.

Anthropic calls Opus 5 the most capable generally available model for scientific research. It shows gains over Opus 4.8 across all life sciences evaluations, with standout improvements in organic chemistry (plus 10.2 percentage points on deriving molecular structures from spectroscopy data) and protein-related tasks (plus 7.7 percentage points).

Alongside Opus 5, Anthropic is releasing two beta features. Mid-Conversation Tool Changes on the Claude Platform let developers swap available tools during a conversation without invalidating the prompt cache. Automatic Fallbacks on the API route blocked requests to a different model automatically.

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/