Cheapest Claude API
Optimize the Workload, Not the Sticker最便宜的 Claude API：优化实际负载，不要只看标价
Direct API, cloud marketplace and gateway routes can differ in model availability, features, data terms and commercial treatment. Compare only equivalent workload outcomes.
直连 API、云市场与网关路由可能在模型可用性、功能、数据条款与商业处理上不同。只比较等价的工作负载结果。
TL;DR
Record model version, provider path, region, processing mode and date.
Caching economics depend on stable prefixes, reuse, timing and cache policy.
Use asynchronous batch only when completion windows and failure handling fit.
Include retries, tools, long context and output quality in the denominator.
记录模型版本、供应商路径、区域、处理模式与日期。
缓存经济性取决于稳定前缀、复用、时间与缓存 Policy。
只有完成窗口与失败处理适配时才使用异步批处理。
把重试、工具、长上下文与输出质量纳入分母。
Claude cost changes with workload shapeClaude 成本随负载形态变化
Input, output, cache writes, cache reads and batch processing can have different commercial treatment. Tool use, long context, regional inference and platform access add further dimensions.
输入、输出、缓存写入、缓存读取与批处理采用不同计费方式；工具使用、长上下文、区域推理与平台接入还会增加其他成本维度。
Prompt caching is not simply “discounted input.” Current first-party pricing distinguishes a 5-minute write, a more expensive 1-hour write, and cache reads at a fraction of base input. The result depends on prefix size, reuse count, reuse timing, and whether requests actually hit the same cacheable content. Model the initial write and later reads separately.
Prompt Caching 并不只是“更便宜的输入”。当前第一方价格区分 5 分钟写入、价格更高的 1 小时写入，以及按基础输入一小部分计费的缓存读取。最终是否省钱取决于前缀大小、复用次数、复用时间以及请求是否真正命中相同可缓存内容，因此必须分别计算首次写入与后续读取。
A batch route can fit offline evaluation but be unsuitable for an interactive assistant. US-only inference and some regional or multi-region cloud endpoints can carry premiums. Server-side tools may add usage charges, while tool definitions and results also consume tokens. Compare the exact model, route, geography, processing mode, and tool set instead of treating every Claude endpoint as one price.
Batch 适合离线评测，却可能不适合交互助手。仅限美国的推理以及部分区域或多区域云端点可能存在溢价；Server-side Tool 可能产生额外使用费，工具定义与结果也会消耗 Token。应比较准确模型、路由、地域、处理模式与工具集合，而不是把所有 Claude 端点视为同一个价格。
Claude access route comparisonClaude 接入路由对比
| Route路由 | Best fit最适合 | Verify before choosing选择前验证 |
|---|---|---|
| Direct Claude APIClaude 直连 API | First-party features, documentation and workspace controls are required.需要一方功能、文档与 Workspace 控制。 | Current models, limits, support, data terms and official pricing.核对当前模型、限制、支持、数据条款与官方价格。 |
| Cloud marketplace云市场 | Cloud procurement, network, identity or region is decisive.云采购、网络、身份或区域是关键。 | Model parity, feature availability, platform terms and regional price.核对模型等价、功能可用性、平台条款与区域价格。 |
| Managed gatewayManaged 网关 | Multi-provider policy and unified operations justify a gateway.多供应商策略与统一运营需要网关。 | Markup, caching behavior, request translation, evidence and data path.核对加价、缓存行为、请求转换、证据与数据路径。 |
| Batch processing批处理 Processing | Latency is flexible and requests can run asynchronously.延迟灵活且请求可异步执行。 | Supported features, result lifetime, failures and current batch terms.核对支持功能、结果生命周期、失败与当前批处理条款。 |
Claude cost dimensionsClaude 成本维度
Measure stable prefix length, writes, reads, TTL and invalidation behavior.
Capture context thresholds, thinking or intermediate usage and output mix.
Count every model turn and tool-related iteration in agent workflows.
Review support, commitments, residency, taxes, credits and invoice path.
测量稳定前缀长度、写入、读取、TTL 与失效行为。
记录上下文阈值、Thinking 或中间用量与输出结构。
计算智能体工作流中的每次模型轮次与工具相关迭代。
检查支持、承诺、驻留、税费、抵扣与账单路径。
Compare with a workload experiment用工作负载实验比较
Build a replay set from real traffic segments: short support turns, long-document analysis, repeated system or document prefixes, structured extraction, tool-using agents, and offline batch jobs. Apply the same factuality, schema, latency, policy, and human-review criteria before counting an output as accepted.
回放集应来自真实流量分层，例如短客服对话、长文档分析、重复系统或文档前缀、结构化抽取、使用工具的 Agent 与离线批处理任务。只有通过相同的事实性、Schema、延迟、策略与人工复核标准，输出才计为有效结果。
- Capture current official pricing, exact model, route, geography, processing mode, and terms.
- Replay identical prompts, tools, output schemas, and acceptance gates with fixed versions.
- Record 5-minute and 1-hour cache writes, cache hits, uncached input, output, tool fees, batch mode, retries, and rejected outputs.
- Measure end-to-end completion and correction work; a lower-priced model may need more retries or human edits.
- Reconcile provider usage to the invoice and calculate cost per accepted result.
- Review the result with engineering, finance, security, data-governance, and legal owners.
- 记录当前官方价格、准确模型、路由、地域、处理模式与条款。
- 使用固定版本回放相同 Prompt、工具、输出 Schema 与验收门禁。
- 记录 5 分钟与 1 小时缓存写入、缓存命中、未缓存输入、输出、工具费用、Batch、重试与未通过输出。
- 衡量端到端完成情况与返工；较低价模型可能需要更多重试或人工修改。
- 将 Provider 用量与账单对齐，并计算每个有效结果的成本。
- 由工程、财务、安全、数据治理与法务负责人共同审查结果。
Treat price as versioned input data把价格视为版本化输入数据
A pricing snapshot records official source, retrieval time, model, route and dimensions. A workload runner emits native usage and acceptance results. A ledger combines them without overwriting historical snapshots, enabling repeatable comparisons when terms change.
价格快照记录官方来源、获取时间、模型、路由与维度；负载 Runner 生成原生用量与验收结果；账本组合二者且不覆盖历史快照，从而在条款变化后可重复比较。
Production rule: do not infer current Claude pricing from an old article, screenshot or cached table.
生产规则：不要从旧文章、截图或缓存表格推断当前 Claude 价格。
Account for external capability calls计入外部能力调用
QVeris gives Claude workflows governed access to external APIs, tools, services and live data. Attribute those calls to the same workflow trace, but keep their pricing source, credentials and ledger dimensions separate.
QVeris 为 Claude 工作流提供受治理的外部 API、工具、服务与实时数据访问。可将调用归属到同一工作流调用链，但价格来源、凭证与账本维度应独立。
Current Claude list-price ladder当前 Claude 目录价阶梯
Official Claude API prices verified July 21, 2026, in USD per 1 million tokens. Claude Sonnet 5 has a temporary rate through August 31, 2026; refresh the source before using this table after that date. Regional and data-residency choices can add premiums.
以下 Claude API 官方价格于 2026 年 7 月 21 日核验，单位为美元/百万 Token。Claude Sonnet 5 的临时价格持续至 2026 年 8 月 31 日，之后使用必须刷新；区域与数据驻留选项可能产生溢价。
| Model or route模型或路由 | Input / cost输入/成本 | Output输出 | Scope适用范围 |
|---|---|---|---|
| Claude Haiku 4.5 | $1.00 | $5.00 | Cache read $0.10缓存读取 $0.10 |
| Claude Sonnet 5 (through Aug 31) | $2.00 | $10.00 | Time-limited price限时价格 |
| Claude Sonnet 4.6 | $3.00 | $15.00 | Global first-party base rate第一方全球基础价 |
| Claude Opus 4.8 | $5.00 | $25.00 | Standard mode标准模式 |
Always verify the linked official pricing source immediately before a buying decision. Model quality, output length, cacheability, retries, tools, service tier, region, taxes, and discounts can reverse a token-price comparison.
采购决策前务必立即核验页面所链接的官方定价来源。模型质量、输出长度、可缓存性、重试、工具、服务层、区域、税费与折扣都可能逆转 Token 单价比较。
FAQ
No. Route, region, workload, contract and platform fees all matter.
No. It depends on eligible stable prefixes, reuse timing and current terms.
Only if the workload tolerates asynchronous completion and batch constraints.
不是。路由、区域、负载、合同与平台费用都会影响结果。
不能。取决于合格稳定前缀、复用时机与当前条款。
只有负载能接受异步完成与批处理约束时才适合。