We put 20 frontier large language models — those with real, at-scale API traffic, image/video models excluded — onto one ruler, measured on two axes: price (USD per 1M tokens) and token structure (the in : cache : out : reasoning mix of real usage, normalized to output = 1).
1 · Two price bands
US flagships (Claude Opus, GPT-5.5) sit at $5 in / $25–30 out. Chinese models cluster an order of magnitude below — the cheapest, Tencent Hy3, is $0.063 in / $0.21 out, and DeepSeek V4 Flash undercuts it on output at $0.18.
2 · The full table
| Model | Input | Output | Cache | In : Cache : Out : Think |
|---|---|---|---|---|
| US · avg in $2.93 / out $15.43 / cache $0.30 | ||||
 Claude Opus 4.7 Claude · 2026-04 · 1M | $5 | $25 | $0.5 | 114.6 : 94 : 1 : 0.05 |
Claude Opus 4.8 Claude · 2026-05 · 1M | $5 | $25 | $0.5 | 86.2 : 68.3 : 1 : 0.06 |
 GPT-5.5 GPT · 2026-04 · 1.05M | $5 | $30 | $0.5 | 67.1 : 55.9 : 1 : 0.51 |
GPT-5.4 GPT · 2026-03 · 1.05M | $2.5 | $15 | $0.25 | 38.3 : 22.8 : 1 : 0.23 |
 Gemini 3.1 Flash Lite Gemini · 2026-05 · 1M | $0.25 | $1.5 | $0.025 | 11.6 : 3.5 : 1 : 0.31 |
Gemini 3.5 Flash Gemini · 2026-05 · 1M | $1.5 | $9 | $0.15 | 17.1 : 10.4 : 1 : 0.7 |
 Grok 4.3 Grok · 2026-04 · 1M | $1.25 | $2.5 | $0.2 | 8.8 : 4.4 : 1 : 0.68 |
| China · avg in $0.50 / out $1.73 / cache $0.08 | ||||
 DeepSeek V4 Flash DeepSeek · 2026-04 · 1M | $0.09 | $0.18 | $0.02 | 23.5 : 13.4 : 1 : 0.52 |
DeepSeek V4 Pro DeepSeek · 2026-04 · 1M | $0.44 | $0.87 | $0.0036 | 23.9 : 19.5 : 1 : 0.7 |
 MiMo-V2.5 小米 MiMo · 2026-04 · 1M | $0.14 | $0.28 | $0.0028 | 111.4 : 104.2 : 1 : 0.36 |
MiMo-V2.5-Pro 小米 MiMo · 2026-04 · 1M | $0.44 | $0.87 | $0.0036 | 82.6 : 73.2 : 1 : 0.38 |
 MiniMax M3 MiniMax · 2026-05 · 1M | $0.3 | $1.2 | $0.06 | 53 : 44.6 : 1 : 0.53 |
 Hy3 preview 腾讯混元 · 2026-04 · 262K | $0.063 | $0.21 | $0.021 | 60 : 53.9 : 1 : 0.37 |
 GLM 5.2 GLM · 2026-06 · 1M | $0.98 | $3.08 | $0.18 | 61.2 : 48.3 : 1 : 0.56 |
GLM 5.1 GLM · 2026-04 · 203K | $0.98 | $3.08 | $0.18 | 52.9 : 45.2 : 1 : 0.62 |
 Step 3.7 Flash 阶跃星辰 · 2026-05 · 256K | $0.2 | $1.15 | $0.04 | 50.2 : 35.3 : 1 : 0.02 |
 Kimi K2.6 Kimi · 2026-04 · 262K | $0.66 | $3.41 | $0.14 | 31.3 : 23.7 : 1 : 0.78 |
Kimi K2.7 Code Kimi · 2026-06 · 262K | $0.61 | $3.07 | $0.13 | 58.2 : 49.3 : 1 : 0.64 |
 Qwen3.7 Plus 通义千问 · 2026-06 · 1M | $0.32 | $1.28 | $0.064 | 20.5 : 15.1 : 1 : 0.71 |
Qwen3.7 Max 通义千问 · 2026-05 · 1M | $1.25 | $3.75 | $0.25 | 28.1 : 21.1 : 1 : 0.69 |
3 · Token structure: input dwarfs output
In real usage every output token rides on tens of input tokens. MiMo-V2.5 runs 111 : 104 : 1 — 111 input tokens per output token, of which 104 are cache hits (94%). Even the leanest, Grok 4.3, is 8.8 : 4.4 : 1. Reasoning tokens stay small (mostly under 0.7 of output); Claude and Step barely meter them.
4 · What you actually pay
Because cached input bills at the discounted cache price, and input outweighs output by tens of times, the real bill is dominated by how much of your input hits cache — not by the headline output price. A pricier model with aggressive cache discounts can end up cheaper in production.
Method: API list prices as of June 2026 (USD / 1M tokens); token structure from real-usage statistics of mainstream models, normalized to output = 1; 20-model sample with at-scale traffic, image/video models excluded.