Cheapest LLM API in 2026: Cost-Per-Token Compared
Cost-per-million-token comparison across OpenAI, Anthropic, Google, Mistral, and Groq for 2026, with a use-X-when verdict for each.
The cheapest capable LLM APIs in 2026 are the small models: Gemini 2.5 Flash-Lite and GPT-5.4-nano at roughly $0.10 to $0.20 per million input tokens, with Groq’s open-model hosting going lower still on speed-sensitive work. For frontier reasoning you pay 20x to 50x more. The right answer is almost never one model, it is routing cheap models for easy work and expensive ones only when needed. Prices below are updated 2026 and pulled from each provider’s own pricing page; they drop fast, so treat the ranking as a snapshot, not a contract.
Cost-per-1M-token comparison (updated 2026)
| Model | Input / 1M | Output / 1M | Use it when |
|---|---|---|---|
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | Highest-volume, cost-sensitive work where quality is good enough |
| GPT-5.4-nano | $0.20 | $1.25 | Cheap OpenAI option for classification and simple tasks |
| Groq Llama 3.3 70B | $0.59 | $0.79 | You need very high throughput and low latency on an open model |
| Gemini 2.5 Flash | $0.30 | $2.50 | Strong price-to-quality for everyday production traffic |
| Mistral Small | $0.10 | $0.30 | EU data preference, open-weight friendly, cheap |
| Mistral Medium 3 | $0.40 | $2.00 | Mid-tier quality without frontier pricing |
| Claude Haiku 4.5 | $1.00 | $5.00 | Fast Claude for high-volume support and extraction |
| Gemini 2.5 Pro | $1.25 | $10.00 | Long-context reasoning at a lower price than GPT-5.4/Claude Sonnet |
| GPT-5.4 | $2.50 | $15.00 | General-purpose OpenAI workhorse |
| Claude Sonnet 4.6 | $3.00 | $15.00 | Most production agent and coding workloads |
| Claude Opus 4.8 | $5.00 | $25.00 | Hardest reasoning, complex agents |
| GPT-5.5 | $5.00 | $30.00 | OpenAI’s top tier for the hardest tasks |
Gemini 2.5 Pro input rises to $2.50/1M above a 200k-token prompt. Note that Gemini 2.0 Flash was shut down on June 1, 2026, so older “cheapest API” lists that name it are out of date.
The verdict by use case
- Absolute cheapest at acceptable quality: Gemini 2.5 Flash-Lite or Mistral Small.
- Best everyday price-to-quality: Gemini 2.5 Flash.
- Fastest open-model hosting: Groq, which runs Llama and similar models on custom hardware at hundreds of tokens per second.
- Best frontier value: Claude Sonnet 4.6 for agents and coding; Gemini 2.5 Pro when you need a long context window cheaply.
- Hardest reasoning: Claude Opus 4.8 or GPT-5.5, used sparingly.
Prices change, so design for it
Industry prices fell roughly 80% from 2025 to 2026. The lesson is not to memorize a number, it is to build so you can switch models without rewriting your app. A gateway or proxy in front of your providers lets you route by cost, fail over, and renegotiate without touching code. That is the angle behind a setup like a self-hosted LiteLLM gateway: one OpenAI-compatible endpoint, many providers behind it, with cost tracking baked in. See LLM API cost control for the techniques and AI integration if you want it wired up for you.
Drill down by provider
Related
Background: what an LLM is and how tokens drive cost. The single biggest lever after model choice is prompt caching.