LLM API Pricing: 375 Models Compared Across 137 Providers

本文为原文前 6,000 字符的节选翻译，完整内容请查看原文。

Oct 20, 2026 You would expect that when 375 models are each served by at least two providers, the market would have found some pricing equilibrium. It has not. The median input price spread across those 375 models is 0.0%—meaning that for the typical model, every provider charges exactly the same per-million-token rate. Nobody is undercutting anybody. In a market with 137 providers and 4,033 total models, that level of uniformity is not competition; it is coordination by default.

2026年10月20日，你或许会认为，当375个模型每个都由至少两家供应商提供服务时，市场应该已经找到了某种定价平衡。但事实并非如此。这375个模型的输入价格中位数价差为0.0%——这意味着对于典型模型而言，每家供应商收取的每百万token费率完全相同。没有人试图压低价格。在一个拥有137家供应商和4033个模型的市场中，这种统一性并非竞争，而是默认的协调。

But that median hides a long tail of extreme divergence. Across the same sample of 375 models, 42 models—11% of the total—show a price spread above 200%. Another 13 sit between 100% and 200%, and 25 more fall in the 50–100% range. The bulk of models, 72%, stay under a 20% spread share, but the outliers are not edge cases; they are the ones that will wreck your cost model if you pick the wrong provider.

但这一中位数掩盖了极端差异的长尾效应。在同样的375个模型样本中，有42个模型（占总数的11%）的价格价差超过200%。另有13个模型处于100%至200%之间，还有25个模型落在50%至100%的区间。虽然大部分模型（72%）的价差保持在20%以下，但这些异常值并非边缘案例；如果你选择了错误的供应商，它们足以摧毁你的成本模型。

The worst case is openai/gpt-oss-120b. It is offered by 13 providers, and the cheapest charges $0.03 per million input tokens (wandb) while the priciest charges $0.80 (crusoe). That is a spread of +2566.7%—a 26.7x difference in cost for the same model, same input, same tokens. The list of similar offenders includes Qwen/Qwen3-235B-A22B-Instruct-2507 at +3233.3% and deepseek-ai/DeepSeek-R1-Distill-Qwen-14B at +2185.7%. If you are choosing an inference provider based on a single price quote, you are gambling.

最糟糕的情况是openai/gpt-oss-120b。它由13家供应商提供，最便宜的收费为每百万输入token 0.03美元（wandb），而最贵的收费为0.80美元（crusoe）。这产生了+2566.7%的价差——对于同一个模型、同样的输入和同样的token，成本差异高达26.7倍。类似的违规者名单还包括Qwen/Qwen3-235B-A22B-Instruct-2507（+3233.3%）和deepseek-ai/DeepSeek-R1-Distill-Qwen-14B（+2185.7%）。如果你仅凭单一报价来选择推理供应商，那无异于赌博。

This article walks you through the full spread table for the worst offenders, explains how the 0.0% median coexists with the 2566.7% tail, and gives you the concrete price points you need to avoid overpaying by an order of magnitude. Read on for the numbers that matter.

本文将带你查看最严重违规者的完整价差表，解释0.0%的中位数如何与2566.7%的尾部并存，并为你提供避免支付高出一个数量级费用的具体价格点。请继续阅读以获取关键数据。

The spread distribution Input price spread | Models | Share

200% | 42 | 11% 100-200% | 13 | 3% 50-100% | 25 | 7% 20-50% | 25 | 7% <20% | 270 | 72% Of 375 models with comparable provider coverage, 42 (11%) cost more than 3x more at the priciest provider than at the cheapest.

价差分布 输入价格价差 | 模型数量 | 占比

200% | 42 | 11% 100-200% | 13 | 3% 50-100% | 25 | 7% 20-50% | 25 | 7% <20% | 270 | 72% 在375个具有可比供应商覆盖范围的模型中，有42个（11%）在最昂贵供应商处的成本是最低供应商处的3倍以上。

Why the median is zero The median input price spread across comparable models is exactly 0.0%, which is not an artifact of rounding—it is the structural norm. Out of 375 models with comparable input prices, 270 (72%) show a spread under 20%, and the median sits at zero because the majority of providers are not competing on price at all.

为什么中位数是零 可比模型之间的输入价格中位数价差恰好为0.0%，这并非四舍五入的结果，而是结构性的常态。在375个具有可比输入价格的模型中，270个（72%）的价差低于20%，中位数之所以为零，是因为大多数供应商根本不在价格上进行竞争。

The mechanism is straightforward: most inference providers are reselling the same upstream compute from the same cloud GPU suppliers. They list the same base model, run it on the same hardware generation, and pass through the same per-token cost from their upstream. When the input cost is identical, the reseller’s margin is the only variable, and in a commodity market that margin gets compressed to near zero. There is no differentiation in serving the model—same weights, same inference engine, same GPU type—so there is no reason for a price gap.

其机制很简单：大多数推理供应商都在转售来自同一云GPU供应商的上游算力。他们列出相同的基准模型，在相同的硬件代际上运行，并转嫁来自上游的相同每token成本。当输入成本相同时，转售商的利润率是唯一的变量，而在大宗商品市场中，这一利润率会被压缩至接近零。在模型服务方面没有差异化——相同的权重、相同的推理引擎、相同的GPU类型——因此没有理由出现价格差距。

The second mechanism is list-price resale. Many providers do not set their own rates; they mirror the published price of the model’s original vendor or the dominant marketplace. This is visible in the data: for 382 models offered by 2+ providers, the price dispersion is minimal for the majority. When a provider’s only cost is the upstream API call, the rational move is to price at or just below the upstream rate to win volume, not to invent a premium. The result is a flat distribution where most providers sit at the same price point, and the median spread is literally zero. This is not a sign of a healthy competitive market; it is a sign of a pass-through market where no one owns the compute.

第二种机制是标价转售。许多供应商并不设定自己的费率；他们直接镜像模型原始供应商或主导市场发布的定价。这一点在数据中显而易见：对于由两家以上供应商提供的382个模型，大多数模型的价格离散度极小。当供应商唯一的成本是上游API调用时，理性的做法是以等于或略低于上游费率的价格来赢得交易量，而不是凭空增加溢价。结果就是一种平坦的分布，大多数供应商处于相同的价格点，中位数价差确实为零。这并非健康竞争市场的标志，而是无人拥有算力的转手市场的标志。

Where the spreads actually come from The large spreads do not come from the commodity resellers—they come from two distinct business models that sit at opposite ends of the pricing spectrum. The first cluster is GPU-hour hosts. These providers (e.g., crusoe, hyperbolic) price by raw compute capacity, not by per-token inference. They rent you a GPU for a fixed hourly rate and let you run any model on it. Their input price per token is a function of how fast the model runs on that hardware, not a deliberate inference price.

价差的真正来源 巨大的价差并非来自大宗商品转售商，而是来自处于定价谱系两端的两种截然不同的商业模式。第一类是GPU小时租赁商。这些供应商（如crusoe、hyperbolic）按原始算力容量定价，而非按每token推理定价。他们以固定的每小时费率向你出租GPU，并允许你在上面运行任何模型。他们每token的输入价格是模型在该硬件上运行速度的函数，而非刻意设定的推理价格。

When a model like openai/gpt-oss-120b runs on a rented GPU, the per-token cost can be wildly different depending on the host’s utilization, power cost, and hardware generation. That is why the worst case in the dataset is openai/gpt-oss-120b at $0.03/M on wandb versus $0.8/M on crusoe—a +2566.7% spread. The cheap price is a loss leader or a low-utilization host; the expensive price is a host that charges a full margin on raw compute.

当像openai/gpt-oss-120b这样的模型在租赁的GPU上运行时，每token的成本会因主机的利用率、电力成本和硬件代际而产生巨大差异。这就是为什么数据集中最糟糕的情况是openai/gpt-oss-120b在wandb上为0.03美元/百万token，而在crusoe上为0.8美元/百万token——价差高达+2566.7%。低价是亏本引流或低利用率主机所致；高价则是对原始算力收取全额利润的主机。

The second cluster is specialised inference providers. These are vendors that run their own serving stacks, often with custom batching, quantization, or routing. They price to cover their engineering cost, not just the GPU. For example, mistral-large-latest is $0.5/M on mistral’s own platform but $8/M on azure—a +1500.0% spread. Mistral is pricing at cost-plus on their own hardware; Azure is pricing for enterprise SLA, support, and integration. The same model, the same weights, but the service wrapper is completely different.

第二类是专业推理供应商。这些供应商运行自己的服务堆栈，通常带有自定义批处理、量化或路由功能。他们的定价旨在覆盖工程成本，而不仅仅是GPU成本。例如，mistral-large-latest在Mistral自己的平台上为0.5美元/百万token，但在Azure上为8美元/百万token——价差为+1500.0%。Mistral是在自己的硬件上按成本加成定价；而Azure的定价则包含了企业级SLA、支持和集成服务。同一个模型，同样的权重，但服务包装完全不同。

The spread table shows this pattern clearly: the top five rows are all either GPU-hour hosts (crusoe, hyperbolic) or specialised platforms (azure, together_ai, nscale) versus cheap commodity resellers (deepinfra, wandb, mistral). The 42 models with spread >200% (11% of n) are almost entirely in this bucket. These are not price errors; they are two different products sold under the same model name.

价差表清楚地显示了这种模式：前五行要么是GPU小时租赁商（crusoe, hyperbolic），要么是专业平台（azure, together_ai, nscale），与廉价的大宗商品转售商（deepinfra, wandb, mistral）形成对比。价差超过200%的42个模型（占总数的11%）几乎全部属于这一类。这些并非价格错误；它们是在同一个模型名称下销售的两种不同产品。

What the spread table hides Three caveats are easy to miss when reading the spread table. First, the prices are per-token input rates at a short context length. Long-context requests—where the model must read a large prompt—are priced differently by many providers. The $0.03/M rate on wandb for openai/gpt-oss-120b likely assumes a standard context length; a long-context request can push that price up by an order of magnitude. The spread table does not capture that, so the 2566.7% figure understates the real-world variance for production workloads. Second, the dataset’s price

价差表掩盖了什么 阅读价差表时，有三个注意事项容易被忽略。首先，价格是短上下文长度下的每token输入费率。长上下文请求（模型必须读取大型提示词）在许多供应商处的定价方式不同。wandb上openai/gpt-oss-120b的0.03美元/百万token费率可能假设的是标准上下文长度；长上下文请求可能会使该价格上涨一个数量级。价差表并未体现这一点，因此2566.7%的数字低估了生产工作负载中的实际差异。其次，数据集的价格