2026 大模型 API 价格横评：20 个模型的单价怎么读

The short version

Picking the wrong model can inflate the same workload's bill by close to 30x, and that has little to do with which model is smarter. An LLM API invoice is set by three interacting variables: the input-to-output price ratio, whether long contexts are surcharged, and how efficiently a tokenizer compresses your language. This piece unpacks all three using the 20 model rows in our models.json dataset, verified on 2026-09-15. Every figure is US dollars per million tokens, so you can check it against the vendors' pricing pages.

The price sheet itself

The most expensive row (GPT-5.6 Sol at 30 dollars for output) and the cheapest hosted row (DeepSeek V4 Flash at 0.28) differ by roughly two orders of magnitude. The formula below, not any single cell, is what decides your invoice.

From unit price to invoice: a hand-checkable example

Take an internal summarisation job: 8,000 input tokens per call (system prompt plus document), 1,200 output tokens, 2,000 calls per day, 30 days a month. Cost per call = input tokens / 1,000,000 x input price + output tokens / 1,000,000 x output price.

• Claude Sonnet 4.6: 8,000 / 1,000,000 x 3 = 0.024 dollars, 1,200 / 1,000,000 x 15 = 0.018 dollars, 0.042 per call, about 2,520 dollars a month.
• DeepSeek V4 Flash: 8,000 / 1,000,000 x 0.14 = 0.00112 dollars, 1,200 / 1,000,000 x 0.28 = 0.000336 dollars, 0.001456 per call, about 87 dollars a month.

That is a factor of roughly 29, not the 100 the table header suggests, because this workload is input-heavy and DeepSeek prices output at only 2x input instead of the more common 6x. Our LLM API cost calculator runs the same arithmetic on the same dataset if you want to substitute your own volumes.

The second variable: long-context surcharges

A million-token window does not mean a flat million-token price:

• Claude Opus 4.6 charges one rate across its whole 1,000,000 window, with no long-context premium.
• Gemini 3 Pro roughly doubles the input price beyond 200,000 input tokens, so a 300K-token retrieval question is billed at 4 dollars per million, not 2.
• Grok 4.1 Fast offers 2,000,000 tokens of context with both rates under one dollar, which suits large-context, short-output jobs.
• Llama 4 Scout's 10,000,000 window is a hosted reference price; self-hosting bills you in GPU hours instead of tokens.

Our token counter tells you whether a request crosses a surcharge threshold in the first place; it estimates with the same tokenizer parameters and labels its output as approximate.

The third variable: how much text fits in one token

Per-million-token prices are only comparable when a token carries comparable content. Our dataset records 1.3 Chinese characters per token for Claude, 1.4 for Gemini and 1.5 for the OpenAI family, so the same 3,000-character Chinese document is about 2,308 tokens on Claude and 2,000 on GPT-5.6, a 15 percent gap. Fifteen percent is noise in a benchmark and real money in a high-volume pipeline. These are flagged estimates: good enough for routing decisions, not for contracts.

Open weights and self-hosting

Three rows in the table are marked as locally runnable (Llama 4 Scout, Qwen3 32B, Gemma 3 27B). Their 0.10 dollar input price comes from third-party hosts; self-hosting trades tokens for VRAM and throughput. Qwen3 32B's 32,768 window and Gemma 3 27B's 128,000 window rule out long-context work, while Llama 4 Scout's ten-million window makes KV cache the first thing that breaks. The local LLM hardware matcher converts parameter counts and quantisation into a VRAM range instead of a vague "you need 24GB".

Three procurement patterns

1. High-volume classification, extraction and routing: weight output price and latency. GLM 5.3 Flash prices input and output identically (0.19/0.19), and Grok 4.1 Fast (0.2/0.5) works on the same logic.
2. Long-document QA and contract review: check the surcharge tiers before the headline price. Gemini 3 Pro and Claude Sonnet 4.6 are two different philosophies, and only your own token distribution decides which wins.
3. Chinese-language work with compliance constraints: Qwen 3.7 Max and Kimi K3 sit in the directly purchasable tier, DeepSeek V4 Pro (1.74/3.48) is the cheapest comparable row, and anything with data-residency requirements has to be compared against self-hosted cost.

To line up capabilities beside the prices, generate a side-by-side sheet with the model comparison table builder; for sustained product comparisons we keep separate write-ups on ChatGPT vs Claude and DeepSeek R1 vs ChatGPT.

Sources and limits

Every price here comes from our model dataset, last verified on 2026-09-15 against vendor pricing pages and three aggregator articles. API prices change often, sometimes twice in a month, so treat this as a framework and confirm the numbers before you sign anything. The monthly figures are linear extrapolations from a hypothetical workload: they exclude prompt caching, batch discounts, retries and network cost, and they are not budget advice.