Cheapest Chinese LLM API for Coding Agents

"Which coding model is cheapest" is usually answered with a per-million-token headline, which hides the thing that matters: coding agents emit far more output than input per useful unit of work. The table below applies one realistic shape — 1M input tokens and 200K output tokens — to every model, so the numbers are comparable.

Cost per large coding run

Output price is the lever, not input price

Two models illustrate the trap:

• Doubao-Seed-2.0-Code has cheap input ($0.448) but expensive output ($2.240) — the run costs $0.90.
• DeepSeek V4 Pro has higher input ($0.570) but cheap output ($1.150) — the same run costs $0.80.

Coding agents spend most of their tokens on completions, diffs and test output, so a model with a 5x output rate is 3-5x more expensive in practice even if its input looks competitive.

The 50x spread, and when it is worth paying

The range from DeepSeek V4 Flash ($0.11) to Kimi K3 ($5.80) is roughly 50x for the same token volume. In practice teams pay for:

• Long-context consistency — refactors over 100K+ token repos punish models that lose track mid-file.
• Tool-call reliability — a failed edit plus retry can cost more than the model's price difference.
• Latency — HighSpeed variants charge 2-3x for shorter wall-clock time on interactive work.

A practical split: run DeepSeek V4 Flash or GLM-5.3-Flash for search, summarisation and diff review, and reserve DeepSeek V4 Pro or Kimi K2.7 Code for the edits that need to land first time.

Calling it

curl https://aiapi-pro.com/v1/chat/completions \
  -H "Authorization: Bearer $NOVAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"refactor src/app.py"}]}'

Every model in the table is reachable through the same endpoint, with the same OpenAI SDK settings — switching is a one-line model id change.

FAQ

What is the cheapest model for a coding run?

DeepSeek V4 Flash at $0.08 per 1M input and $0.17 per 1M output: a 1M-input / 200K-output run costs about $0.11. GLM-4.7-Flash is free but capped at 10 calls/day on the free tier.

Why is output pricing more important than input pricing for coding?

Coding agents emit long completions. A model with cheap input but a $2.24 output rate (Doubao-Seed-2.0-Code) costs $0.90 per run, while DeepSeek V4 Pro with higher input ($0.57) but low output ($1.15) costs $0.80.

Do these prices include a platform fee?

No. NovAI passes through the provider list price with a 1.0 multiplier and zero markup, so the numbers equal the upstream rate card.

How much does a month of agent usage cost?

A heavy agent month of 20M input and 4M output tokens costs about $2.28 on DeepSeek V4 Flash, $10.80 on MiniMax M3, $16 on DeepSeek V4 Pro and $116 on Kimi K3.

Prices reflect provider rate cards as of 2026-10-22 and are published machine-readably at /china-ai-leaderboard.json. Cost per run = 1M input tokens + 200K output tokens.