Price duel: Three AI models cost very different amounts per token

If a company or developer builds AI into a product, they typically pay per token. A token is a small chunk of text, often part of a word. You pay both for what you send in (input) and for what the model replies (output). Here are the prices for three models.

DeepSeek: lowest prices, but with ranges

DeepSeek's own pricing page lists two API models, deepseek-flash and deepseek-v4-pro. Both have a context window of 1 million tokens, meaning how much text the model can handle at once.

Flash costs $0.15-0.30 per million input tokens and $0.60-1.20 per million output tokens. V4-Pro costs $0.66-1.32 for input and $1.98-3.96 for output. The ranges reflect peak and off-peak traffic: the off-peak price is half the peak price. Peak hours are 01-04 and 06-10 UTC, Monday through Friday, except Chinese public holidays.

Input that hits the cache (text the model has already seen) is far cheaper: $0.003-0.006 per million tokens for flash and $0.022-0.044 for V4-Pro.

DeepSeek also charges $0.75 per 1,000 successful calls to its separate search and fetch APIs. They are free through DeepSeek Harness, Claude Code or Codex until plan-based billing starts on November 1, 2026.

Haiku 5.5: cheap, but tokenization eats into the discount

According to the developer blog Apidog, Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts of up to 100,000 tokens. Above that limit, the price is $0.50 and $2.50. Cache reads start at $0.01, and the batch API costs half the standard price.

Apidog describes Haiku 5.5 as roughly 75 percent cheaper to run than Haiku 4.5 on average. But according to the source, a new tokenization produces about 30 percent more tokens than before, so the price drop is smaller in practice. Cost also swings with the effort-level setting: on some benchmarks, the max level costs up to 28.5 times more than the medium level.

GPT-5.6 Terra: the most expensive of the three

OpenAI's GPT-5.6 Terra has a context window of 1,050,000 tokens. Via OpenRouter, it costs $2 per million input tokens and $12 per million output tokens. That is OpenRouter's price, not necessarily OpenAI's own list price.

Be careful with direct comparisons

The models belong to different classes: a cheap small model, a flagship model and a Chinese open model. The price per token therefore says nothing about quality. Nor does it say what a task costs in practice, because tokenization and effort level vary.