AI Model Pricing 2026: What Every Model Actually Costs Per Million Tokens (Complete Comparison)
The Only AI Pricing Table You Need in 2026
Every major model's cost, normalized to one number: dollars per million tokens. No hunting through five different pricing pages with five different units.
This guide is part of AIGetFree's ongoing coverage of AI pricing and access. For a multi-model subscription that bundles chat, image, video, and agents into one meter, read our Abacus AI Review 2026. For a developer-first single API key across 300+ models, see our OpenRouter Review 2026, and compare it against our API Gateway Showdown.
The short version: Anthropic, OpenAI, and Google each price differently per million tokens, OpenRouter passes those rates through with zero markup, and 26+ open models cost nothing at all. Below, every model normalized to the same unit so you can actually compare them.
What "Per Million Tokens" Actually Means
AI pricing is a mess of different units and fine print, so this guide normalizes everything to one measure: cost per million tokens, split into input (what you send) and output (what the model generates). As a rule of thumb, 1 million tokens is roughly 750,000 words, or about 3,000 pages of text. A typical chat exchange runs 500–2,000 tokens, a long-form article runs 2,000–5,000 tokens, and a complex agent task with tool calls can burn 10,000–50,000 tokens in a single run.
Anthropic (Claude)
| Model | Input / 1M | Output / 1M | Context | Best For | Notes |
|---|---|---|---|---|---|
| Claude Opus 4.8 | $5.00 | $25.00 | 1M | Complex reasoning, multi-file refactors | 88.6% SWE-bench · intro pricing through Aug 2026 |
| Claude Sonnet 4.6 | $3.00 | $15.00 | 1M | Best all-around for content, coding | Best Value Claude |
| Claude Haiku 4.5 | $1.00 | $5.00 | 200K | Fast, cheap Claude tasks | Fastest Claude model |
OpenAI (GPT)
| Model | Input / 1M | Output / 1M | Context | Best For | Notes |
|---|---|---|---|---|---|
| GPT-4.1 | $2.00 | $8.00 | 1M | Structured output, JSON, SEO | Flagship model |
| GPT-4.1 Mini | $0.40 | $1.60 | 1M | Budget OpenAI tasks | Best Value OpenAI |
| GPT-4.1 Nano | $0.10 | $0.40 | 1M | Ultra-cheap classification | Cheapest OpenAI model |
| o3 | $2.00 | $8.00 | 200K | Complex reasoning, math | Price dropped from $10/$40 |
| o4-mini | $1.10 | $4.40 | 200K | Budget reasoning | Latest reasoning model |
Google (Gemini)
| Model | Input / 1M | Output / 1M | Context | Best For | Notes |
|---|---|---|---|---|---|
| Gemini 2.5 Pro | $1.25 | $10.00 | 1M | Best value frontier model | Best Value Frontier |
| Gemini 2.5 Flash | $0.30 | $2.50 | 1M | Fast classification, routing | Price increased from $0.075/$0.40 |
| Gemini 2.5 Flash Lite | $0.10 | $0.40 | 1M | Ultra-cheap, high volume | NEW — cheapest Google model |
Open Source & Independent (via OpenRouter)
| Model | Input / 1M | Output / 1M | Context | Best For | Notes |
|---|---|---|---|---|---|
| Llama 4 Maverick | $0.15 | $0.60 | 1M | Ultra-cheap fallback | Best Cheap Model |
| Llama 4 Scout | $0.10 | $0.30 | 10M | Long-context tasks | 10M context — largest available |
| DeepSeek V3 | $0.20 | $0.77 | 164K | Technical/code tasks | Price dropped slightly |
| DeepSeek R1 | $0.70 | $2.50 | 164K | Open-source reasoning | Chain-of-thought visible |
| Mistral Large 2 | $2.00 | $6.00 | 128K | European AI, multilingual | Strong in French, German, Spanish |
| Qwen3 235B | $0.46 | $1.82 | 131K | Chinese open-weight leader | Strong multilingual |
| NVIDIA Nemotron 3 | $0.09 | $0.45 | 1M | Cheapest 1M context model | NEW — ultra-cheap |
Free Models (via OpenRouter — $0/M Tokens)
| Model | Context | Rate Limit | Best For | Quality |
|---|---|---|---|---|
| Llama 3.3 70B | 131K | 20 req/min, 50/day | General purpose, drafting | ★★★★ |
| Gemma 4 31B | 128K | 20 req/min, 50/day | Lightweight content, outlines | ★★★ |
| Mistral 7B | 32K | 20 req/min, 50/day | Quick classification, routing | ★★ |
| DeepSeek V3 (free) | 164K | 20 req/min, 50/day | Technical tasks, code | ★★★★ |
Where Your Token Spend Actually Goes
Input tokens are cheap almost everywhere — it's output tokens that decide the bill, running 4–8× the input rate across every provider above. Long generations, reasoning traces, and agent tool-call loops are what turn a $0.03 article into a $270/month agent budget. The fix is rarely "use a cheaper model for everything" — it's routing the expensive model only at the steps that need it.
Real-World Cost Scenarios
What you actually pay depends on usage pattern, not sticker price. Three realistic profiles, month over month:
How To Choose: The Decision Framework
Use OpenRouter's free tier: Llama 3.3 70B for general tasks, DeepSeek V3 for code, Gemma 4 31B for drafts. 50 requests/day. Quality is decent, not frontier.
Mix Gemini 2.5 Flash for classification, Claude Sonnet 4.6 for writing, GPT-4.1 for structured output. The sweet spot for content automation.
Add Claude Opus 4.8 for the hardest reasoning tasks, with Llama 4 Maverick as fallback — a full multi-model pipeline with automatic failover via OpenRouter.
Go direct to providers and skip OpenRouter's 5.5% fee. At $500/month that fee alone is $27.50 — a whole extra subscription.
FAQ
What does "per million tokens" actually mean in practice?
Which is the cheapest frontier model?
Why would anyone pay for Claude Opus 4.8 at $25/M output tokens?
Are free models actually usable?
How often do these prices change?
The Bottom Line: Mix Models, Save Money
The single biggest mistake in AI pricing is running everything through one expensive model. A multi-model strategy — Gemini Flash for classification, Claude Sonnet for writing, Llama for fallback — typically cuts costs 60–80% versus routing everything through Opus or GPT-4.1.
OpenRouter makes the swap trivial: one API key, and you change models by changing a string. Prices move every few weeks, so this page gets rechecked on the same schedule — bookmark it.