I Calculated What 1M Tokens Costs Across 50+ LLM Models

I spent weeks compiling pricing for every major LLM model while building a cost tracking tool. Here's the complete breakdown—50+ models across OpenAI, Anthropic, Google, Mistral, and more.

Most developers guess their AI costs. They pick a model, ship to production, and hope for the best. But the price difference between models is staggering—o1-pro costs 1,500x more than GPT-5-nano for output tokens.

This guide covers every major model as of February 2026, with real pricing data and practical recommendations.

The Complete Pricing Table

Prices are per 1 million tokens. Most models charge differently for input (what you send) and output (what the model generates).

OpenAI Models

GPT-5 Series (Latest)

GPT-4.1 Series

GPT-4o Series

O-Series (Reasoning Models)

Anthropic Models

Claude 4.5 (Latest)

Claude 4 Series

Claude 3.x Series

Google Gemini Models

Gemini 3 (Latest)

Gemini 2.5

Gemini 2.0 (Deprecating March 2026)

Mistral Models

Surprising Findings

1. The Price Range is Insane

Output tokens range from $0.04/1M (Ministral 3B) to $600/1M (o1-pro). That's a 15,000x difference.

2. The "Mini" Model Wars

Every provider now has a mini model competing for the budget tier:

• GPT-5-nano: $0.05 input / $0.40 output
• Gemini 2.5 Flash Lite: $0.10 input / $0.40 output
• Ministral 3B: $0.04 input / $0.04 output
• GPT-4o-mini: $0.15 input / $0.60 output

For simple classification and extraction, these models are practically free at scale.

3. Reasoning Models are 10-100x More Expensive

The o-series models (o1, o3, o4) cost significantly more—and that's just the sticker price. They also generate hidden "thinking" tokens that inflate your actual costs.

4. Claude 4.5 Opus is Surprisingly Cheap

At $5/1M input and $25/1M output, Claude 4.5 Opus is cheaper than Claude 4 Opus ($15/$75). Anthropic is getting more aggressive on pricing.

5. Google Offers the Best Value for High-Volume

Gemini 2.0 Flash at $0.10/$0.40 is hard to beat for volume. And Gemini 2.5 Flash Lite at $0.10/$0.40 maintains quality while staying cheap.

Real Cost Comparisons

Let's calculate actual costs for common tasks:

Customer Support Bot (1M conversations/month)

Average: 500 input tokens, 300 output tokens per conversation

Document Summarization (100K docs/month)

Average: 4,000 input tokens, 500 output tokens per document

AI Agent (50 steps per task, 10K tasks/month)

Average: 1,000 input tokens, 500 output tokens per step

My Recommendations

For Startups (Cost-Sensitive)

1. Default to GPT-5-nano or GPT-4o-mini — Handle 80% of tasks
2. Use Gemini 2.0/2.5 Flash for volume — Best price/performance
3. Route complex queries to GPT-5 or Claude 4.5 Sonnet
4. Avoid reasoning models unless necessary — 10x+ cost increase

For Enterprise (Quality-Focused)

1. Claude 4.5 Opus or GPT-5.2 for critical tasks
2. Claude 4.5 Sonnet or GPT-5 for general production
3. o3-mini for reasoning tasks — Good balance of capability and cost
4. Track everything — Know your cost-per-feature

For AI Agents

1. Start with GPT-4o-mini — Test your agent logic cheaply
2. Use o3-mini for reasoning steps — Not o1 or o3-pro
3. Batch and cache aggressively — Agents make many similar calls
4. Set per-task budgets — Runaway agents can cost hundreds per task

How to Track All This

Knowing prices is one thing. Tracking actual costs in production is another. Provider dashboards show totals, but not which features drive costs.

I built Orbit to solve this. One-line SDK integration, and you get per-feature cost breakdowns across all providers:

import { Orbit } from '@with-orbit/sdk';
import OpenAI from 'openai';
import Anthropic from '@anthropic-ai/sdk';

const orbit = new Orbit({ apiKey: process.env.ORBIT_API_KEY });

// Track OpenAI costs by feature
const openai = orbit.wrapOpenAI(new OpenAI(), {
  feature: 'chat-assistant'
});

// Track Anthropic costs by feature
const anthropic = orbit.wrapAnthropic(new Anthropic(), {
  feature: 'document-analysis'
});

// All calls automatically tracked with cost, tokens, latency