How Much Do 1 Million Tokens Cost? Per-Provider Pricing in 2026

The cost of 1 million tokens depends on the model, the direction (input vs output), and where the request runs. In 2026 the spread is enormous: from well under a dollar on small models to tens of dollars on the largest reasoning models.

This guide breaks down per-million-token pricing on the models people actually use, and shows how caching and batch discounts change the effective rate.

Rough price bands per 1 million tokens

• Small/fast models (mini, flash, small): $0.10 - $0.50 input, ~$1-2 output.
• Mid-size models (gpt-4o-mini class, Gemini flash): $0.25 - $0.60 input, $1.50 - $2.50 output.
• Frontier models (GPT-4o, Claude, Gemini Pro): $2.50 - $5 input, $10 - $15 output.
• Reasoning models (o-series, Claude thinking): $3 - $4 input, $12 - $16 output.
• Cached input: often 50-90% cheaper than uncached input on the same model.

Costs that hide inside the per-million number

1. Output tokens cost up to 5x input tokens, so verbose completions dominate the bill.
2. System prompts and tool definitions are input tokens billed on every request.
3. Retries and function-calling loops multiply tokens silently.
4. Context bloat re-sends history every turn in a multi-turn conversation.

How to cut the effective per-million price

Routing small tasks to cheap models, caching stable prefixes, batching background work, and trimming context are the four biggest levers. A LayerFlow-style gateway can apply all four automatically and show the per-million cost per model in your spend analytics.

FAQ