The cheapest LLM API depends on input/output mix, caching, reasoning tokens, tool retries, and task success. Rank models against one workload snapshot and publish the lookup date. Token price alone cannot identify the lowest cost per completed task.
A practical evaluation method
- Use representative prompt and output distributions.
- Calculate cached and uncached input separately.
- Track validator pass rate and human-review time alongside spend.
Cheapest by the Numbers
If cost per token is the priority, a handful of models are dramatically cheaper than the flagships — often 20–100x on output, where most agent and codegen spend concentrates. Ranked by list price (per 1M tokens, input / output):
| Model | Input | Output | Note |
|---|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.28 | 98% cache discount, 1M context |
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | fast, high-volume |
| MiniMax M2.7 | $0.30 | $1.20 | agent-trained, open-weight |
| Gemini 2.5 Flash | $0.30 | $2.50 | balanced workhorse |
| DeepSeek V4 Pro | ~$0.44 | ~$0.87 | promo; reasoning-grade |
| Kimi K2.6 | $0.95 | $4.00 | cache-hit $0.16 |
| GLM-5.1 | ~$1.40 | ~$4.40 | 8-hr agentic |
Winner on raw price: DeepSeek V4 Flash. Gemini 2.5 Flash-Lite edges it on input but costs more on output; for output-heavy work, DeepSeek wins decisively.
Sticker Price Isn't the Whole Bill
The cheapest effective cost depends on your workload:
- Cache-heavy (agents, RAG, chatbots)? DeepSeek's automatic prefix caching cuts repeated input ~98% (to ~$0.0028/1M); Kimi's cache-hit is $0.16 vs $0.95. A model with great cache economics can beat a lower sticker price.
- Output-heavy (codegen, long answers)? Output rates dominate — DeepSeek V4 Flash's $0.28 output is the killer number.
- Reasoning "thinking" modes bill at the same rate but burn 3–5x more tokens. Keep them off by default.
The Two Tricks That Beat Any Sticker Price
- Route by difficulty. Send the bulk to a cheap model and escalate only hard tasks to a flagship — a 5–25x spread most teams never capture.
- Cache + batch. Prompt caching (up to ~90–98%) plus Batch API (50% off) can drop effective spend 95%+ on the right workload.
Full playbook in how to reduce AI API costs.
The Honest Caveats
The cheapest models are mostly open-weight Chinese labs (DeepSeek, MiniMax, GLM, Kimi). Two things to know: several store data in China on their official APIs (a GDPR/compliance consideration — self-hosting or a zero-retention gateway sidesteps it), and some carry Modified-MIT licenses that restrict commercial self-hosting. Using them through a gateway like LinkModel (retention terms that should be verified before use) handles both the data and billing side with one key.
Bottom Line
- Absolute cheapest, output-heavy → DeepSeek V4 Flash.
- Cheapest input for classification/routing → Gemini 2.5 Flash-Lite.
- Cheapest capable agent model → MiniMax M2.7.
- Then layer routing + caching + batch and the flagship-everything baseline disappears.
Compare against the premium tier in best LLM API and the full field in the AI API pricing comparison.
Start free with a $1 credit and see your real per-call cost.
