The cheapest LLM APIs in 2026 — with real numbers
"Cheapest" depends on your input/output mix, but for a typical 60/40 input-heavy workload, this is the August 2026 budget leaderboard among widely available models:
| Model | Vendor | Input | Output | 1M in + 1M out |
|---|---|---|---|---|
| Llama 3.1 8B | Meta (via providers) | $0.02 | $0.05 | $0.07 |
| DeepSeek V4 Flash | DeepSeek | $0.14 | $0.28 | $0.42 |
| Llama 4 Scout | Meta (via providers) | $0.10 | $0.30 | $0.40 |
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | $0.50 | |
| GPT-4o mini | OpenAI | $0.15 | $0.60 | $0.75 |
| Mistral Small 4 | Mistral | $0.15 | $0.60 | $0.75 |
| DeepSeek V4 Pro | DeepSeek | $0.435 | $0.87 | $1.31 |
| GPT-5.6 Luna | OpenAI | $0.20 | $1.20 | $1.40 |
The catch: cheap per token ≠ cheap per task
A small model that needs three retries, or produces answers you must post-process, costs more than a mid-tier model that nails it once. The honest procedure is: build a 50-example eval set from your real traffic, run it down the price ladder, and pick the cheapest model that passes. Most teams discover their classification, routing, extraction and short-summary traffic — often 70%+ of volume — passes on models under $1/M output.
Where each budget model shines
Llama 3.1 8B is unbeatable for high-volume, low-stakes transforms (tagging, dedupe, normalization) and its per-provider pricing keeps falling. DeepSeek V4 Flash punches far above its price on reasoning-lite tasks, and its cache-hit input price of $0.0028/M is effectively free for repeated prefixes. Gemini 2.5 Flash-Lite pairs a 1M-token context with rock-bottom rates, making it the budget choice for long-document work. GPT-5.6 Luna is the strongest raw model of the four — the safe default when quality matters but budgets are tight.
Don't forget the two discounts
Prompt caching (~90% off repeated input) and batch APIs (50% off everything, async) can make a mid-tier model cost less than a budget model's sticker price. If your workload has a big stable system prompt or can tolerate async processing, run the numbers both ways — our caching guide walks through the math.
At sustained high volume, renting GPUs undercuts every API above. Community marketplaces like Vast.ai and serverless GPU clouds like RunPod rent an RTX 4090-class card for well under $1/hour — enough to serve a quantized 8B model at effectively $0.01–0.02 per million tokens if you keep it busy. The trade-off is ops work: cold starts, scaling, and monitoring are on you.
Enter your token mix once — see cost per request and per month, cheapest first.
Prices compiled from vendor pricing pages, last verified August 21, 2026. Open-weight model prices vary by inference provider. Verify before committing.