LLM API Pricing Comparison
Normalize Before You RankLLM API 价格对比：先统一口径，再进行排序
A price table becomes useful only after routes, units, workloads and acceptance criteria match. Keep current official source data separate from measured workload evidence.
只有路由、单位、负载与验收标准一致后，价格表才有意义。应将当前官方来源数据与实测负载证据分开保存。
TL;DR
Fix prompt distribution, context, tools, modalities, latency and quality gates.
Map units for comparison but keep every provider-native billing field.
Store source URL, retrieval time, currency, region, model and route.
Calculate total delivered cost after retries, tools and quality failures.
固定提示词分布、上下文、工具、模态、延迟与质量门禁。
为比较映射单位，同时保留每个供应商原生计费字段。
保存来源 URL、获取时间、币种、区域、模型与路由。
在纳入重试、工具与质量失败后计算交付总成本。
Why simple price tables fail为什么简单价格表会失真
Providers may separate input, output, cached input, cache writes or storage, batch, flex, priority, long context, media and tool charges. Access paths can add platform fees, commitments, taxes or regional differences.
供应商可能分别计费输入、输出、缓存输入、缓存写入或存储、Batch、Flex、Priority、长上下文、媒体与工具。不同接入路径还会增加平台费、承诺、税费或区域差异。
The billing words are not always equivalent. OpenAI currently separates short- and long-context rates and multiple service tiers. Claude distinguishes base input, two cache-write durations, cache reads, Batch, inference geography, and server-side tools. Gemini prices models, modalities, processing tiers, grounding, and tools under different rules. DeepSeek separates cache-hit and cache-miss input. Normalize each native field, but do not erase what it means.
各家使用的计费术语并不完全等价。OpenAI 当前区分长短上下文价格与多个 Service Tier；Claude 区分基础输入、两种缓存写入时长、缓存读取、Batch、推理地域与 Server-side Tool；Gemini 对模型、模态、处理层级、Grounding 与工具采用不同规则；DeepSeek 则区分缓存命中和未命中输入。比较时可以统一单位，但不能抹掉字段含义。
A model row is also not a quality-equivalence claim. Different models may require different prompt lengths, reasoning settings, output tokens, tool retries, or human correction to finish the same task. A candidate that needs more attempts or review can cost more per accepted outcome despite a lower listed token price.
同一价格表中的模型也不代表质量等价。完成同一任务时，不同模型可能需要不同 Prompt 长度、Reasoning Setting、输出 Token、工具重试或人工修正。即使标价更低，若需要更多尝试或复核，每个有效结果的成本仍可能更高。
Comparison layers比较层次
| Layer层次 | Best fit最适合 | Verify before choosing选择前验证 |
|---|---|---|
| Official list price官方标价 | Build a timestamped reference for exact models and routes.为准确模型与路由建立带时间戳参考。 | Units, thresholds, currencies, tiers, regions and special modes.核对单位、阈值、币种、层级、区域与特殊模式。 |
| Measured usage实测用量 | Estimate each route with representative workload evidence.用代表性负载证据估算每条路由。 | Native usage, cache, tools, media, retries and failed calls.核对原生用量、缓存、工具、媒体、重试与失败调用。 |
| Accepted-output cost有效输出成本 | Compare delivered work after a shared quality gate.在共享质量门禁后比较已交付工作。 | Sample size, evaluator version, review cost and uncertainty.核对样本量、Evaluator 版本、审查成本与不确定性。 |
| Total commercial cost商业总成本 | Support procurement and long-term operating decisions.支持采购与长期运营决策。 | Support, SLA, commitments, credits, taxes, data terms and engineering.核对支持、SLA、承诺、抵扣、税费、数据条款与工程成本。 |
Comparison worksheet fields比较工作表字段
Provider, platform, model version, endpoint, region and processing mode.
Input, output, cache, batch, tools, modalities, storage and extras.
Latency, retry rate, valid schema, quality acceptance and review effort.
Official URL, timestamp, owner, currency, tax basis and expiry review.
供应商、Platform、模型 Version、端点、Region 与 Processing Mode。
输入、输出、缓存、批处理、工具、模态、存储与附加项。
延迟、重试率、结构定义有效性、质量通过率与审查工作量。
官方 URL、时间戳、Owner、币种、税基与到期复核。
Create a repeatable comparison建立可重复比较
Use a workload manifest that records the task mix, prompt and context distribution, expected schema, permitted tools, latency target, factuality and safety rubric, and human-review rule. Freeze that manifest for one comparison run. Otherwise a prompt improvement or easier sample can be mistaken for a price advantage.
建立工作负载 Manifest，记录任务组合、Prompt 与上下文分布、预期 Schema、允许工具、延迟目标、事实性与安全 Rubric，以及人工复核规则。一次比较运行期间应冻结该 Manifest，否则 Prompt 改进或样本变简单都可能被误判成价格优势。
- Version the workload, acceptance rubric, prompt, tool schemas, SDK, and provider configuration.
- Fetch current official pricing and terms, retaining source URL, retrieval time, currency, region, route, model ID, context band, and service tier.
- Run candidates against the same sample and capture every native request ID, usage field, cache event, tool charge, retry, timeout, and failed attempt.
- Apply the shared quality gate before calculating cost per accepted result and per completed user workflow.
- Reconcile measured usage to invoices and report discounts, credits, taxes, platform fees, and uncertainty separately.
- Publish the date, scope, assumptions, excluded features, sample size, confidence, owner, and next review trigger.
- 版本化负载、验收 Rubric、Prompt、工具 Schema、SDK 与 Provider 配置。
- 获取当前官方价格和条款，并保留来源 URL、获取时间、币种、区域、路由、模型 ID、上下文区间与 Service Tier。
- 让候选运行相同样本，记录每个原生 Request ID、用量字段、缓存事件、工具费用、重试、超时与失败尝试。
- 先应用统一质量门禁，再计算每个有效结果与每个完整用户工作流的成本。
- 将实测用量与账单对齐，并分开报告折扣、抵扣、税费、平台费与不确定性。
- 发布时说明日期、范围、假设、排除功能、样本量、置信度、Owner 与下次复核触发条件。
Join price snapshots to workload runs连接价格快照与负载运行
A source registry stores official pricing snapshots. A workload runner creates traceable native-usage events and acceptance results. The comparison layer normalizes units, adds retry and tool costs, applies the quality gate and reports cost per accepted output.
Source 注册表保存官方价格快照；工作负载 Runner 生成可追踪的原生用量事件与验收结果；比较层标准化单位、加入重试与工具成本、应用质量门禁，并报告每个有效输出的成本。
Production rule: never mix price data from different dates, model versions or commercial routes without labeling the difference.
生产规则：绝不能在未标记差异的情况下混合不同日期、模型版本或商业路由的价格数据。
Compare external capability costs on their own ledger在独立账本中比较外部能力成本
QVeris discovers and calls external APIs, tools, services and live data. Their fees may use requests, records, compute time or subscription units rather than tokens. Keep the native units, then join end-to-end cost at the workflow level.
QVeris 发现并调用外部 API、工具、服务与实时数据。其费用可能按请求、记录、计算时间或订阅单位计算，而非 Token。保留原生单位，再在工作流层合并端到端成本。
A dated API price snapshot for workload tests用于工作负载测试的定价快照
Verified against official pricing pages on July 21, 2026. Prices are USD per 1 million tokens and are not quality-equivalent recommendations. Use them to seed a workload calculator, then refresh before procurement or publication because providers can change models and rates.
以下数据于 2026 年 7 月 21 日根据官方定价页核验，单位为美元/百万 Token，并不表示模型质量等价。可用来初始化工作负载计算器，但采购或发布前必须再次刷新。
| Model or route模型或路由 | Input / cost输入/成本 | Output输出 | Scope适用范围 |
|---|---|---|---|
| OpenAI GPT-5.6 Luna | $1.00 | $6.00 | Cached input $0.10/MTok缓存输入 $0.10/百万 Token |
| Claude Haiku 4.5 | $1.00 | $5.00 | Cache read $0.10/MTok缓存读取 $0.10/百万 Token |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | Standard text/image/video input标准文本/图像/视频输入 |
| DeepSeek V4 Flash | $0.14 | $0.28 | Cache-miss input; cache hit $0.0028未命中缓存输入；命中为 $0.0028 |
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
Only as a dated snapshot with an owner and review process; prices change.
Cost per accepted workload result, supported by full billing and quality evidence.
Yes when they materially affect production feasibility or total cost.
只能作为带日期、Owner 与复核流程的快照；价格会变化。
每个有效工作负载结果的成本，并有完整计费与质量证据支持。
当其显著影响生产可行性或总成本时，应纳入。