Model pricing moved a lot this year, and most "cheapest API" lists still quote numbers from spring. This one uses the live rate cards as of August 25, 2026 — all official-direct prices, all callable through a single NovAI key.
| Model | Input | Output | Context |
|---|---|---|---|
| qwen3.5-flash | $0.063 | $0.336 | 1M |
| minimax-text-01 | $0.14 | $1.13 | 1M |
| deepseek-v4-flash | $0.21 | $0.63 | 1M |
| qwen-plus | $0.26 | $0.78 | 128K |
| minimax-m2.5 | $0.282 | $1.128 | 200K |
| qwen3-max | $0.33 | $1.32 | 256K |
| deepseek-v4-pro | $0.63 | $1.90 | 1M |
| glm-5 | $0.671 | $3.021 | 128K |
| qwen3.8-max (flagship) | $1.67 | $5.00 | 256K |
And one entry that breaks the scale: glm-4.6v-flash is free — a vision-capable model at $0 per token, no daily cap.
Take a support-bot pipeline: 2,000 requests/day, ~800 input tokens and ~300 output tokens each. On qwen3.5-flash that is 1.6M input + 0.6M output per day = $0.30/day, about $9/month. The same traffic on a $1.67/$5.00 flagship runs ~$55/month. For most routing/classification/extraction workloads the budget tier passes the same evals — run your own before paying flagship rates.
1. Ignoring context window. A cheap model with 8K context forces chunking pipelines that multiply token usage. Every model above has at least 128K.
2. Comparing input price only. Summarization and agent workloads are output-heavy; minimax-text-01 looks cheap until you multiply output volume by $1.13.
3. Paying for a relay markup. Aggregators add 5%+ on top of list price. NovAI bills at provider list rates with no platform fee, and USDT (TRC20) are accepted.
Starting point for most teams: qwen3.5-flash for volume work, deepseek-v4-pro when quality matters, and reserve the flagships for tasks nothing cheaper passes.
Related: Qwen 3.8 Max guide · Kimi K3 pricing