Chinese AI API Pricing Landscape — July 2026

Chinese model vendors have spent the last quarter undercutting each other, and the result is the cheapest frontier-adjacent API market in the world. This analysis currently covers 60 models live on our gateway: 27 LLMs and 33 image, video, and speech models. Every price below is generated from our public pricing dataset rather than copied into this article by hand. Dataset revision: 2026-11-11; pricing fingerprint: c8ce28b249cdc4e9.

One honest caveat up front: price is not quality. This article tells you what things cost; our benchmark series (task completion rates per model, measured through this same gateway) is the companion that tells you what they're worth.

The sub-$0.20 input tier

5 models charge less than $0.20 per million input tokens. The important comparison inside this tier is output cost: two models with the same cheap input rate can diverge sharply once they start generating long answers.

Two things stand out. First, input rates inside this tier only run from $0.1 to $0.15 per million tokens, so the input price barely separates these models. Second, watch output prices: they range from $0.3 for step-3.5-flash to $0.528 for hy3, and that gap dominates generation-heavy workloads.

The 1M-context club

13 of the 27 LLMs in the current catalog publish a 1M-token context window: qwen3.7-plus, qwen3.8-flash, qwen3.8-max, deepseek-flash, deepseek-v4-pro, LongCat-2.0, MiniMax-M3, kimi-k3, step-5-preview, hy4-preview, glm-5.2, glm-5.3, glm-5.3-flash. A year ago 1M context was a premium feature; in the Chinese market it is becoming table stakes.

The cheapest published input rate among those million-context models belongs to qwen3.8-flash at $0.15 per million input tokens ($0.47 output). The most interesting long-context entry is Meituan's LongCat-2.0: an open MoE reasoning model at $0.3 input / $1.2 output with the full 1M window — flagship-class specs at flash-tier pricing.

Flagship pricing by vendor

Taking each vendor's highest published output price in the current text catalog as a consistent proxy—not a quality ranking—produces this comparison:

Moonshot's kimi-k3 currently carries the highest output rate in this table at $15 per million tokens, and Meituan's LongCat-2.0 sits at the other end at $1.2. That spread is a pricing claim by the vendors, not evidence of the same gap in quality; only task-completion benchmarks can answer whether the premium holds.

Tencent's hy3 deserves a flag: a major-vendor model at $0.132 input / $0.528 output per million tokens. Access friction outside China is exactly the gap this gateway exists to close.

Video: per-second vs per-clip, with real math

Video pricing comes in two shapes, and comparing them requires arithmetic vendors don't do for you:

For short drafts, per-generation MiniMax-Hailuo-2.3-Fast is $0.1875 in the representative catalog configuration. Per-second billing (HappyHorse, Wan) can win when you need very short clips or precise duration control. ByteDance's Seedance line has configuration-specific pricing, so the public dataset intentionally marks rows without one representative unit price and points to live pricing instead.

What to pick, by workload

Our stance, from pricing and specs alone (quality benchmarks follow separately):

• Bulk extraction, classification, RAG pre-processing: deepseek-flash ($0.3 input / $1.2 output) — with a 1M window.
• Long-document agents on a budget: LongCat-2.0 — 1M context and reasoning-tuned at $0.3 input / $1.2 output.
• Multimodal agent loops (image and video input): start with qwen3.8-flash ($0.15 input / $0.47 output), the lowest output rate among the million-context models that take image and video input in this catalog: qwen3.8-flash, kimi-k3, step-5-preview, glm-5.3-flash.
• Low-cost video drafts among representative per-generation rows: MiniMax-Hailuo-2.3-Fast, currently $0.1875 per generation.
• Premium coding: Moonshot's kimi-k3 carries the catalog's highest output rate at $15/M. Benchmark it on your own repository before committing spend.

What this article doesn't tell you

• Quality. A cheap model that fails your task twice can cost more than a premium model that succeeds once. Our task-completion benchmark is the companion piece.
• Stability under load. Chinese vendor APIs vary in rate limits and peak-hour behavior; we route around some of this at the gateway layer, but haven't published numbers yet.
• Tomorrow's prices. Vendors reprice aggressively (usually downward, occasionally promo-then-revert). We now snapshot every price change, so future editions of this article will show trend lines, not just a snapshot.

All 60 models run behind one OpenAI-compatible endpoint with a single key. Download the JSON or CSV, read the methodology, and use the $2 free credit to verify a published price with a real request.