Models at or under $0.20 per million input tokens ($1.50 output) — for classification, extraction, and high-volume workloads
| Features | |||||
|---|---|---|---|---|---|
| $0.10 | $0.40 | $0.02 | |||
| $0.10 | $0.40 | $0.02 | |||
| $0.15 | $0.60 | $0.07 |
© 2026 LLM Gateway. All rights reserved.
Every model here costs at most $0.20 per million input tokens and $1.50 per million output tokens — some, like Qwen3 4B and Llama 3.2 3B, as little as $0.03. At these prices a million-token workload costs a few cents, which changes what's economical: classify every support ticket, summarize every call, run an LLM check on every commit.
Cheap doesn't mean toy: GPT-OSS 120B, GLM-4.7 Flash, Gemini 2.5 Flash-Lite, and Qwen3.5 9B punch far above their price on everyday tasks. Route high-volume work here and reserve frontier models for the requests that actually need them.
In this catalog, Llama 3.2 3B and Qwen3 4B start around $0.03 per million input tokens, with GPT-OSS 20B at about $0.04 and GLM-4.7 Flash at $0.06. Prices differ per provider, so check the list — the same model is often cheaper through one provider than another.
For classification, extraction, routing, summarization, and simple chat — usually yes. Small models fail mostly on multi-step reasoning and niche knowledge. A common pattern is a cheap model as the default with automatic escalation to a frontier model when confidence is low.
Cache responses for repeated requests, use cached input pricing for long shared prefixes, batch offline work, and set per-project spending limits. Routing through a gateway also lets you switch to whichever provider currently offers the lowest price for the same model with zero code changes.
Yes — free mappings show up at $0.00 in this list. They're rate-limited and best for prototyping; for production traffic, the paid models on this page are the reliable low-cost option.