🔶 Claude Opus 5 定价与发布日单任务成本测试结果
英文标题：🔶 Claude Opus 5 pricing and launch-day cost-per-task result
Anthropic 于 2026 年 7 月 24 日发布 Claude Opus 5，并以与 Opus 4.8 相同的标准 API Token 价格提供更强能力。Anthropic 称，Opus 5 在多项基准测试中显著提升性能，并通过减少推理 Token、工具调用和失败重试降低每个完成任务的实际成本。AI Pricing Guru Labs 在发布日的 49 项确定性任务测试中测得 Opus 5 获得 49/49、零 API 错误，但其输出 Token 数明显高于 Opus 4.8，导致实测运行成本更高。开发者应以真实业务中的成功任务成本、重试率、工具调用和人工修改量为依据进行评估，并谨慎使用价格为基础费率两倍的 Fast mode。
568 B
中文
Anthropic 于 2026 年 7 月 24 日发布 Claude Opus 5，现已可在 Claude 产品和 Claude API 中使用。对买家而言，核心信息非常明确：Opus 5 在标准 Token 定价与 Opus 4.8 相同的情况下，实现了能力升级。
上方的实时表格由我们的定价数据集生成，因此费率变化时会保持更新。如需查看完整 Claude 产品线，请访问 Anthropic 定价页面；如需进行工作负载成本测算，请打开 AI Token 计算器。
发生了什么变化
Opus 5 现已成为 Claude Max 的默认模型，也是 Claude Pro 可用的最强模型。API 开发者可通过模型 ID claude-opus-5 调用它。
Anthropic 将此次发布定位于日常 Opus 使用场景与定价更高的 Fable 层级之间。在 Frontier-Bench v0.1 上，Anthropic 表示，Opus 5 的表现超过 Opus 4.8 的两倍，同时降低了每个完成任务的成本。在 CursorBench 3.2 的最高努力等级下，它以大约一半的单任务成本，达到距离 Fable 5 峰值分数 0.5% 以内的水平。
其他发布声明包括：在 ARC-AGI 3 上的得分达到下一最佳模型的三倍；在 Zapier AutomationBench 上，以相同任务成本实现约为下一最佳模型 1.5 倍的通过率；以及在 OSWorld 2.0 上，以略高于三分之一的成本取得优于 Fable 5 峰值的结果。
这些均为厂商报告的评测结果，并不保证适用于每一种生产工作负载。在调整路由策略之前，应使用自己的已接受补丁、任务完成率、审核时间和重试率等指标测试 Opus 5。
定价影响
从 Opus 4.8 升级至 Opus 5，标准 API 价格没有上涨。对于已经支付 Opus 层级费用的团队而言，这是一项直接的价值升级。
更重要的定价因素是任务效率。如果 Opus 5 能以更少轮次完成复杂工作、减少工具调用，或避免陷入失败的 Agent 循环，即便单 Token 费率不变，其实际成本也可能下降。Anthropic 引述的早期访问结果包括：在 Lovable 最困难的 Agent 编程测试中，性能比 Opus 4.7 高 22%；在一项法律工作流评测中，最大推理设置下生成的 Token 数比 Opus 4.8 少 26%。
Fast mode 是买家需要单独预算的例外。Anthropic 表示，该模式的运行速度约为普通模式的 2.5 倍，但在 Claude Platform 上的价格为 Opus 5 基础费率的 2 倍。当低延迟具有业务价值时可以使用它；对于后台工作而言，将其作为默认选项成本较高。
我们的首个 Opus 5 Labs 测试结果
我们在发布当天将 Opus 5 加入 AI Pricing Guru Labs，并运行了全部 49 项确定性的推理、提取、编程和定价任务。Opus 5 获得 49/49 的成绩，且 API 错误为零，准确率与 Opus 4.8 持平。
在这一有限测试集中，新模型的输出更为冗长：完成 Token 数为 1,989，而 Opus 4.8 为 630。尽管标价相同，这带来了更高的实测运行成本。这是一个有价值的提醒，而非普适性的排名。该测试集未衡量 Anthropic 报告提升幅度最大的长周期编程和 Agent 任务；团队应在自身工作中同时测试任务成功率和输出 Token 行为。
请在我们的 OpenAI 定价页面 比较其他高端模型路线，并阅读此前的 Fable 5 和 Mythos 5 定价发布信息，了解 Anthropic 顶级产品层级的背景。
谁会受益，谁会受损
现有 Opus 4.8 API 用户受益最大。他们可以评估能力更强的模型，而无需接受更高的标准价格表。若报告中的重试次数和推理 Token 减少效果能够在实际任务中成立，编程 Agent 团队、金融研究人员、法律平台、科研工作流和重度工具调用产品也将从中获益。
Fable 5 作为通用高端路由将更难证明其合理性。Anthropic 仍将 Fable 定位为前沿智能选项，但 Opus 5 现在已能在重要的编程和知识工作测试中接近其表现，同时单任务成本显著更低。
不需要 Opus 级推理能力的高吞吐量团队获益较少。Sonnet 5 仍然是常规生产流量更经济的默认选择；而 Opus 5 应负责那部分更好的判断能力或更高完成率足以覆盖溢价的任务。
如果不加区分地启用 Fast mode，对延迟敏感的买家反而可能蒙受损失。为交互式 Agent、紧急调试或收入关键型工作流支付两倍基础费率是合理的，但不适用于夜间批处理任务。
开发者现在应该做什么
在变更生产环境默认模型之前，先将 claude-opus-5 加入评估套件。使用真实代码仓库、文档、工具调用和失败条件，将它与 Opus 4.8 和 Sonnet 5 直接比较。
跟踪每个被接受结果的成本，而不仅仅是 Token。记录总输入、缓存输入、输出、重试次数、工具调用次数、实际耗时、人工修改量和最终任务成功率。这些指标才是所宣称的 Opus 5 效率应当体现的地方。
在部署期间保留备用路由。Anthropic 已推出测试版自动回退功能，可将被 Opus 5 安全分类器标记的请求路由至另一模型。在 Claude.ai、Claude Code 和 Claude Cowork 中，被标记的请求默认会回退至 Opus 4.8。
重新验证提示词和安全敏感型工作流。Opus 5 针对一小部分网络安全任务提供了更强的防护措施，尽管 Anthropic 表示，其分类器的干预频率应远低于 Fable 5。安全团队应在迁移前测试具有代表性的防御性工作。
结论
Claude Opus 5 是 Opus 4.8 的同价替代品，具备更有说服力的购买理由：在不提高标准 API 费率的前提下，提供更好的编程、Agent、研究和专业工作表现。
如果 Opus 4.8 已能通过你最困难的任务，Opus 5 显然是最应优先测试的模型。如果 Sonnet 可以处理大部分流量，则应继续将其作为默认模型，仅将困难任务升级至 Opus。应选择性使用 Fast mode，在日志中清晰记录回退情况，并依据每个成功任务的成本作出最终决策。
来源：Anthropic 的 Claude Opus 5 公告、Claude Opus 5 系统卡，以及 AI Pricing Guru 的实时定价数据集。
English / 原文
Anthropic released Claude Opus 5 on July 24, 2026 , making it available across Claude products and the Claude API. The buyer headline is unusually clean: Opus 5 delivers a capability upgrade at the same standard token price as Opus 4.8.
The live table above is generated from our pricing dataset, so it stays current if rates change. For the complete Claude lineup, use the Anthropic pricing page; for workload math, open the AI token calculator.
What changed
Opus 5 is now the default model on Claude Max and the strongest model available on Claude Pro. API developers can call it with the model ID claude-opus-5.
Anthropic positions the release between everyday Opus use and its higher-priced Fable tier. On Frontier-Bench v0.1, Anthropic says Opus 5 more than doubled Opus 4.8’s performance while lowering cost per completed task. At maximum effort on CursorBench 3.2, it finished within 0.5% of Fable 5’s peak score at roughly half the cost per task.
Other launch claims include a score three times higher than the next-best model on ARC-AGI 3, around 1.5× the next-best pass rate at the same task cost on Zapier AutomationBench, and better OSWorld 2.0 results than Fable 5’s peak at just over one-third of the cost.
These are vendor-reported evaluations, not guarantees for every production workload. Run Opus 5 against your own accepted-patch, task-completion, review-time, and retry-rate metrics before changing routing.
Pricing impact
There is no standard API price increase from Opus 4.8 to Opus 5. That makes this a direct value upgrade for teams already paying for the Opus tier.
The bigger pricing story is task efficiency. If Opus 5 closes difficult jobs in fewer turns, uses fewer tool calls, or avoids a failed agent loop, its real cost can fall even though the per-token rate stays flat. Anthropic cites early-access results including 22% higher performance than Opus 4.7 on Lovable’s hardest agentic coding tests and 26% fewer generated tokens than Opus 4.8 at maximum reasoning in one legal-workflow evaluation.
Fast mode is the exception buyers need to budget separately. Anthropic says it runs at around 2.5× normal speed and costs 2× the base Opus 5 rate on the Claude Platform. Use it when latency has business value; it is an expensive default for background work.
Our first Opus 5 Labs result
We added Opus 5 to AI Pricing Guru Labs on launch day and ran all 49 deterministic reasoning, extraction, coding, and pricing tasks. Opus 5 scored 49/49 with zero API errors , matching Opus 4.8 on accuracy.
The new model was more verbose in this narrow suite: 1,989 completion tokens versus 630 for Opus 4.8, producing a higher measured run cost despite identical list prices. That is a useful warning, not a universal ranking. The suite does not measure the long-horizon coding and agent tasks where Anthropic reports the largest gains; teams should test both task success and output-token behavior on their own work.
Compare alternative premium routes on our OpenAI pricing page and read the earlier Fable 5 and Mythos 5 pricing launch for context on Anthropic’s top tier.
Who benefits—and who loses
Existing Opus 4.8 API users benefit most. They can evaluate a stronger model without accepting a higher standard rate card. Coding-agent teams, financial researchers, legal platforms, scientific workflows, and tool-heavy products also stand to gain if the reported reductions in retries and reasoning tokens hold on their tasks.
Fable 5 becomes harder to justify as a blanket premium route. Anthropic still positions Fable as the frontier-intelligence option, but Opus 5 now gets close on important coding and knowledge-work tests at materially lower task cost.
High-volume teams that do not need Opus-level reasoning gain less. Sonnet 5 remains the more economical default for routine production traffic, while Opus 5 should handle the smaller set of jobs where better judgment or a higher completion rate pays back the premium.
Latency-sensitive buyers can lose if they enable Fast mode indiscriminately. Paying twice the base rate makes sense for interactive agents, urgent debugging, or revenue-critical workflows—not overnight batch jobs.
What developers should do now
Add claude-opus-5 to your evaluation suite before changing the production default. Compare it directly with Opus 4.8 and Sonnet 5 using real repositories, documents, tool calls, and failure conditions.
Track cost per accepted result, not only tokens. Record total input, cached input, output, retries, tool calls, wall-clock time, human edits, and final task success. That is where the claimed Opus 5 efficiency should appear.
Keep a fallback route during rollout. Anthropic launched automatic fallbacks in beta, allowing requests flagged by Opus 5 safety classifiers to route to another model. In Claude.ai, Claude Code, and Claude Cowork, flagged requests fall back to Opus 4.8 by default.
Revalidate prompts and safety-sensitive workflows. Opus 5 has stronger safeguards for a narrow set of cyber tasks, although Anthropic says its classifiers should intervene far less often than Fable 5’s. Security teams should test representative defensive work before migration.
Bottom line
Claude Opus 5 is a same-price replacement for Opus 4.8 with a stronger buyer case: better reported coding, agent, research, and professional-work performance without a higher standard API rate.
If Opus 4.8 already passes your hardest tasks, Opus 5 is the obvious model to test first. If Sonnet handles most traffic, keep it as the default and escalate only difficult jobs. Use Fast mode selectively, keep fallbacks visible in logs, and make the final decision on cost per successful task.
Sources: Anthropic’s Claude Opus 5 announcement, the Claude Opus 5 system card, and AI Pricing Guru’s live pricing dataset.