The Cheapest Ways to Use LLMs in 2026

AI — by Joe Jurczyk, CLIKT | November 13, 2026

A practical 2026 guide to the cheapest ways to use LLMs: free tiers, budget APIs, open-model hosts, self-hosting, GPU rental and the tactics that cut AI costs by 90%.

LLM prices now run from about $0.02 to $50 per million tokens, so the same AI task can cost more than 100 times as much depending on how you run it. The cheapest setup for you depends on three things: how much you use it, how private your data needs to be, and how smart the model has to be. This guide covers the main options, from free to fully self-hosted, plus the habits that cut costs no matter which option you pick.

TL;DR

  • Just experimenting? Start with free tiers: Google AI Studio (Gemini Flash), OpenRouter's free models, or Groq.
  • Running automations at low to moderate volume? Use a small API model (Claude Haiku 5.5, Gemini Flash-Lite, GPT-5 nano) and turn on batch processing and prompt caching.
  • High volume or sensitive data? Run open-weight models on a cheap host, or self-host them.
  • The biggest savings come from how you use a model, not which provider you pick. Right-sizing the model, batching and caching can cut a bill by 90% or more.

1. Free tiers

Several providers give you real API access at no cost, with rate limits.

ProviderWhat's freeThe catch
Google AI Studio (Gemini API)Flash and Flash-Lite models, 1M-token contextPro models have been paid-only since April 2026. Free-tier prompts may be used to improve Google's products. Rate limits were cut sharply in 2025–26.
OpenRouterAbout 25–30 free open models (DeepSeek, Llama, Qwen, Gemma) through one OpenAI-compatible API keyAround 20 requests per minute plus daily caps. The free lineup changes often.
GroqFree tier for open models, with very fast responsesPer-minute and per-day token limits
OthersGitHub Models, Cerebras, Mistral, Cloudflare Workers AI, NVIDIA NIMLimits vary; good for prototyping

Consumer chat plans. The free tiers of Claude, ChatGPT and Gemini are fine for manual work. A $20/month subscription is often cheaper than the API for heavy back-and-forth chat, but you can't plug a chat subscription into automations. For that you need the API.

2. Cheap pay-per-token APIs

APIs charge per million tokens (about 750,000 English words), with separate rates for input (what you send) and output (what the model writes). Output usually costs about 5 times as much as input, so long answers cost more than long prompts.

Formula: cost = (input tokens × input rate + output tokens × output rate) ÷ 1,000,000

Budget models worth knowing (per 1M tokens, November 2026)

ModelInputOutputNotes
GPT-5 nano (OpenAI)$0.05$0.40Cheapest major API
Gemini 2.5 Flash-Lite (Google)$0.10$0.40Free tier also available
Claude Haiku 5.5 (Anthropic)$0.10$0.50Strong small model, 1M context
DeepSeek V3.2 (direct)~$0.14~$0.28Low-cost open-weight model
Claude Sonnet 5.5 (mid-tier, for comparison)$2.00$10.00Use only for harder tasks

Prices change often, and some Gemini models are scheduled to double in price on January 1, 2027. Check each provider's pricing page before you budget.

Worked example: daily email summaries

Say you summarize 1,000 emails a day, each with 2,000 tokens in and 300 tokens out. That's about 60M input and 9M output tokens a month.

  • Claude Haiku 5.5: (60 × $0.10) + (9 × $0.50) = $10.50/month
  • Haiku 5.5 with the Batch API: about $5.25/month
  • Claude Sonnet 5.5: (60 × $2) + (9 × $10) = $210/month

Same job, 40 times the price, just from picking a bigger model than the task needs.

Two discounts you should always use

  • Batch API: about 50% off both input and output for jobs that can wait up to 24 hours. Anthropic, OpenAI and Google all offer it.
  • Prompt caching: when you resend the same instructions or documents, cached input is billed at about 10% of the normal rate. Caching and batch stack, cutting input costs by up to about 95%.

3. Third-party hosts for open-weight models

Open-weight models (DeepSeek, Qwen, Llama, Gemma, Mistral, GPT-OSS) are hosted by many competing providers, which pushes prices down. The cheapest small models now run around $0.02 per million tokens.

  • OpenRouter: one API key for hundreds of models, including free ones. It passes through provider pricing and charges about a 5.5% fee when you buy credits. Good for testing several models and having a fallback.
  • Groq and Cerebras: extremely fast responses, which helps for real-time chat and voice.
  • Together AI, Fireworks, DeepInfra, SiliconFlow: low per-token prices on popular open models.

Tradeoffs: rate limits, uneven uptime, and different data-retention policies. Read each host's policy before sending client data, especially with offshore providers.

4. Self-hosting on your own machine

Once you own the hardware, every request is free apart from electricity.

  • Software: Ollama (easiest, one command), LM Studio (desktop app with a chat window), llama.cpp (most control).
  • Hardware: an Apple Silicon Mac with 32GB+ of memory, or a PC with a used RTX 3090/4090 (24GB of video memory). Either runs 7B–30B models comfortably.
  • Quantization: 4-bit versions of models use roughly a quarter of the memory with only a small loss in quality. This is what makes larger models fit on home hardware.

Best for: private data, unlimited drafting, offline work, and learning. Limits: local models are still behind frontier models on hard reasoning, and you handle setup and updates yourself. A used RTX 4090 costs about $2,500, roughly the same as 3,000–8,000 hours of renting one in the cloud.

5. Renting cloud GPUs

Rent a GPU by the hour when you need more power than you own, or want to fine-tune a model.

GPUVast.ai (median)RunPod CommunityRunPod Secure
RTX 4090 (24GB)$0.36/hr$0.34/hr$0.74/hr
L40S (48GB)$0.60/hr$0.79/hr$0.99/hr
H100 SXM (80GB)$3.07/hr—$3.29/hr
  • Watch idle time: a 4090 left running 24/7 costs about $248/month, and an H100 about $2,110/month. Storage keeps billing even when a machine is stopped.
  • Serverless GPU (Modal, Replicate) bills by the second and shuts down when idle. Modal includes $30/month of free credit.
  • Verdict: renting usually beats APIs only at steady, high volume or for fine-tuning. For most small-business work, a cheap API is less expensive and much less hassle.

6. Tactics that cut costs on any option

  1. Right-size the model. Send simple tasks (classifying, extracting, summarizing) to a small model, and route only the hard ones to a bigger one.
  2. Batch anything that isn't urgent: overnight reports, bulk content, data cleanup.
  3. Cache stable context like system prompts, style guides and reference documents.
  4. Trim prompts and cap output length. Output is the expensive part.
  5. Use retrieval (RAG) to send only the relevant passages, not whole documents.
  6. Cache answers to questions that repeat.
  7. Watch reasoning tokens. "Thinking" models bill their hidden reasoning as output. Lower the effort setting for simple tasks.
  8. Set spending caps and alerts on every provider account.

7. Which option fits you?

SituationBest choiceTypical cost
Trying ideas, personal useFree tiers or a chat subscription$0–$20/month
Small-business automations (email summaries, content drafts, tagging)Haiku / Flash-Lite / nano with batch and caching$5–$50/month
Customer-facing app, moderate trafficSmall model plus routing to a mid-tier model$50–$500/month
High volume or sensitive dataOpen-weight model on a host, or self-hostedVaries; $1,500–$3,000 one-time for hardware
Fine-tuning and experimentsRented GPU by the hour$0.30–$3.30/hour

Videos worth watching

What is Ollama? Running Local LLMs Made Simple

(7 min) — A quick explainer on running open models on your own machine, for free and with your data kept private.

Learn Ollama in 15 Minutes: Run LLM Models Locally for FREE

(Tech With Tim, 14 min) — A hands-on crash course covering install, running models and calling them from code.

Complete Ollama Tutorial (2026): LLMs via CLI, Cloud & Python

An up-to-date walkthrough of local, cloud and Python use.

Could This Gemini Trick Finally Replace RAG?

(Prompt Engineering) — Explains context caching and how it can cut API costs by up to 90%.

Bottom line

For most businesses, the cheapest reliable setup in 2026 is a small API model with batch processing and prompt caching turned on, plus free tiers for testing. Self-hosting and GPU rental pay off when volume is high, data is sensitive, or you need a custom model.

Need help setting this up? CLIKT builds AI automations, content pipelines and web apps on WordPress, Shopify and Lovable. Contact us.

Sources

Back to the CLIKT.com blog