AI model prices and names change too quickly for an annual list to stay accurate without a dated snapshot. The table below was checked on August 24, 2026 against Requesty’s live catalog and official provider pages. Prices are per million tokens and use list or direct-provider rates where noted.
Current low-cost models
API price snapshot
| DeepSeek V4 Flash | $0.44 input; $0.01 cached input; $1.32 output at direct peak rate |
|---|---|
| Step 3.7 Flash | $0.20 input; $1.15 output in Requesty's catalog before displayed discount |
| MiniMax M3 | $0.60 input; $2.40 output list rate in Requesty's catalog |
| Qwen3.7 Plus | $0.40 input; $1.60 output list rate in Requesty's catalog |
| BSeed 2.0 Mini | $0.10 input; $0.40 output list rate in Requesty's catalog |
| BSeed 2.0 Code | $0.50 input; $3.00 output list rate in Requesty's catalog |
| GLM-5.3 | $1.40 input; $4.40 output list rate in Requesty's catalog |
| Kimi K2.7 Code | $0.75 input; $3.50 output list rate in Requesty's catalog |
The cheapest row is not automatically the best coding choice. BSeed 2.0 Mini costs less than its Code sibling, but the Code model is intended for development work. Step 3.7 Flash adds vision and selectable reasoning. MiniMax M3 and GLM-5.3 offer 1 million-token context windows, while Kimi K2.7 Code is built for coding workflows with a 256K context window.
DeepSeek’s time-based pricing
DeepSeek introduced peak and off-peak pricing in August. Its direct V4 Flash peak rate is already low, and the published off-peak period charges half price. V4 Pro costs more and is intended for harder work. See DeepSeek V4 pricing for the current schedule.
Context and caching change the bill
Coding agents resend repository context. A model with a higher input price can cost less when its cache-hit rate is much lower and the prompt prefix remains stable. Output is usually more expensive than input, so long explanations and repeated retries can dominate a coding bill.
Use a smaller working set, keep stable instructions at the beginning of the prompt, and check usage after an actual agent session. Marketing examples based on a one-line chat do not represent repository work.
Subscription plans
Z.AI’s Coding Plan uses points and applies a half-points off-peak rate. Alibaba’s current Token Plan uses credits and seats. Neither should be described as unlimited. Compare the live quota with a week of measured API usage before subscribing.
For setup guides, see GLM-5.3 with Claude Code, MiniMax M3 with Claude Code, Qwen3.8-Max with Claude Code, and Kimi K3 with Claude Code.
