LLM Costs & Pricing in 2026
How much does each model really cost? A full price comparison of GPT-5, Claude, Gemini and Grok — cost per million tokens, which model is worth it for each task, and how to cut the bill.
How LLMs are priced — the basics
Language models are priced by tokens (roughly 3–4 characters per word in English; a bit more in other languages). The price splits into input (what you send the model — the prompt and context) and output (what the model generates). Output is almost always more expensive. The prices in the table are per million tokens ($ / 1M).
An average chat ≈ 1K–3K tokens. A million tokens ≈ hundreds of chats. So a model at $1/1M input costs a fraction of a cent per chat — until you reach high volume, and then the differences add up to real money.
Price table — August 2026
Price per 1M tokens (input / output). Figures per the official API pricing as of August 2026:
| Provider | Model | Input | Output | Best for |
|---|---|---|---|---|
| OpenAI | GPT-5.6 Sol | $5.00 | $30.00 | reasoning, heavy code & agents |
| GPT-5.6 Terra | $2.00 | $12.00 | balanced day-to-day work | |
| GPT-5.6 Luna | $0.20 | $1.20 | high volume, minimal cost | |
| GPT-5 (original) | $0.625 | $5.00 | cost-efficient reasoning | |
| Anthropic | Claude Opus 4.8 | $5.00 | $25.00 | complex code, long-running agents |
| Claude Sonnet 4.6 | $3.00 | $15.00 | balanced — an excellent default | |
| Claude Haiku 4.5 | $1.00 | $5.00 | fast & cheap, high volume | |
| Gemini 3 Pro | $2.00 | $12.00 | long context, document analysis | |
| Gemini 3.7 Flash | $0.75 | $3.75 | a fast, cheap workhorse | |
| Gemini 3.5 Flash-Lite | ~$0.10 | ~$0.40 | classification, huge volume, low latency | |
| xAI | Grok 4.6 | $2.00 | $6.00 | agents, code, real-time X |
| Grok 4.5 | $2.00 | $6.00 | previous gen, cheaper when cached |
LLM prices change frequently. Some models charge a higher rate for context above a certain threshold (e.g. Grok above 200K, and Gemini Flash is expected to rise in 2027). Always verify against the official pricing before budgeting.
Which model for which task
- Massive volume / classification / simple extraction: Gemini 3.5 Flash-Lite or GPT-5.6 Luna — a fraction of a cent per call
- Chatbot / day-to-day work: Gemini 3.7 Flash, Claude Haiku 4.5 or GPT-5.6 Terra — a price/quality balance
- Complex code & reasoning: Claude Opus 4.8, GPT-5.6 Sol or Grok 4.6
- Very long context (documents/codebase): Gemini 3 Pro
- Real-time information from X: Grok 4.6
- High-quality non-English: GPT-5 and Claude lead; Gemini is close
You don't need a single model. Route simple tasks to a cheap model (Haiku/Flash/Luna) and only the complex ones to the expensive model. That alone can cut 50–80% of the cost without hurting quality.
How to save — 4 big levers
- Prompt Caching: a recurring fixed context (system prompt, documents) is cached — a discount of up to ~90% on that input
- Batch API: non-urgent processing (up to 24 hours) at a 50% discount with most providers
- Model Routing: the cheap model for easy tasks, the expensive one only for the hard ones
- Tightening prompts: fewer input tokens, a precise
max_tokens— every wasted token costs money
For a full guide with code examples for each technique — Cutting LLM costs.
Don't want to pay at all? Run locally (free)
You can run language models locally on your machine — completely free, with no API and no account. The data also stays with you (full privacy), and it works offline too. The catch: you need decent hardware (especially RAM/GPU), and open models are usually a bit weaker than the commercial flagships (GPT-5, Claude, Gemini) — but in 2026 the gap has narrowed a lot, and it's more than enough for most tasks.
Install Ollama (or LM Studio for a GUI), and run a model with one command — e.g. ollama run qwen3 or ollama run llama3.3. That's it. The model downloads and runs locally.
Leading open models for local install (2026)
| Model | Developer | Best for | Size / hardware | License |
|---|---|---|---|---|
| Qwen 3 | Alibaba | general default + code | 8B ≈ 5GB · larger sizes too | Apache 2.0 |
| DeepSeek V4 | DeepSeek | reasoning & complex logic | large (strong GPU / lots of RAM) | MIT |
| Llama 3.3 | Meta | easy start, general chat | 8B ≈ 4.9GB · runs in 8GB RAM | Llama License |
| Gemma 3 | weak hardware + multimodal | 4B ≈ 4.2GB (most efficient) | Gemma Terms | |
| Mistral | Mistral AI | code & speed | 7B ≈ 4–5GB | Apache 2.0 |
| Phi-4 | Microsoft | especially weak hardware | small, easy to run | MIT |
Which local model to choose?
- A safe all-purpose pick: Qwen 3 — strong, multilingual and openly licensed
- Reasoning & logic: DeepSeek V4 — but it needs strong hardware
- A weak machine / 8GB RAM: Gemma 3 4B or Llama 3.3 8B
- Code: Qwen Coder or Mistral
- Fully flexible (commercial) license: Qwen and Mistral (Apache 2.0), DeepSeek (MIT)
A ~8B-parameter model runs comfortably with 8–16GB RAM (or a GPU with 6GB+ VRAM). Models of 70B and up need a strong GPU or lots of memory. If your machine is weak — start with Gemma 3 4B or Phi-4.
For the full guide on running locally, DeepSeek and the open models — including step-by-step installation with Ollama and LM Studio — see DeepSeek & local models.
Sources
The figures were verified against the providers' official API pricing (OpenAI, Anthropic, Google, xAI) as of August 2026. Because pricing changes frequently, this guide is included in an automated freshness check that periodically re-verifies the numbers against the sources.
Want to cut costs?
After you've picked a model — the practical cost-reduction guide, and the full guides for each model.