AI API prices span 100-fold, but the token rate is only half the bill

A survey of 18 AI API offers found output prices ranging from $0.50 to $50 per million tokens, a 100-fold spread. But caching, batch discounts, regional rates and tokenisation can make the price on a rate card a poor guide to the cost of a real job.

Watch Desk analysis

What happened

ActuIA surveyed public vendor and cloud rate cards between midnight and 01:00 UTC on 6 October. Its comparison covers OpenAI, Anthropic, Google, Mistral AI, DeepSeek and xAI, with prices converted into euros using the European Central Bank’s 5 October reference rate. Vendors bill in dollars, and the figures can change after the survey.

The headline spread is striking, but the report’s worked examples show why a single per-token figure is not a budget. For 10,000 support tickets with repeated instructions, ActuIA calculates that caching cuts the bill from $80 to $51.50 on GPT-6.1 Sol and $53 on Claude Sonnet 5.5. The estimate excludes cache-write fees.

Key findings

• Repeated text can be cheaper to process
  Caching discounts apply only to prompt material repeated between calls, such as instructions and tool definitions; writes can carry a fee.
• Batch processing can halve eligible bills
  OpenAI, Anthropic, Google and Mistral AI offer batch discounts in the configurations surveyed, in exchange for slower responses.
• The same text can produce different token counts
  Anthropic says its newer tokenizer can produce about 30% more tokens for the same text, though the increase varies by content.
• A cheaper token does not guarantee a cheaper task
  The survey’s prices do not measure how many tokens a model needs, or whether it completes the task successfully.

Why it matters

For teams choosing a model or estimating API spend, the practical question is not simply “What does a million tokens cost?” It is how much of each prompt can be cached, whether the job can wait for batch processing, where inference must take place, and how many tokens the task actually consumes. Regional endpoints can add a 10% premium in the surveyed examples, while DeepSeek’s weekday peak hours affect its rates.

That turns pricing into an engineering and procurement question as much as a model-selection one. A low list price can lose its shine if a workload generates more tokens or needs more retries; a higher rate may be offset by efficient use. ActuIA’s examples reproduce rate cards at stated volumes, not model success rates or a universal cost per task.

Our read

This is a useful comparison precisely because it resists the neat but misleading leaderboard of cheapest tokens. Start with your own workload: measure its token use, account for cache writes and location requirements, then compare the cost of getting the job done. The rate card sets the price of a token; your architecture decides how many you buy.

What to watch

• Whether providers change public rates or discount rules after the 6 October survey.
• How real workloads compare once token counts, retries and task success are included.
• Whether regional availability and data-location requirements change the apparent bargain.

Discussion spark: When choosing an AI API, should teams optimise first for the lowest price per token or the lowest cost for a successfully completed task?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.