Real cost of OpenAI and Anthropic models: the gap narrows

Tests on four business tasks show that the price displayed per million tokens poorly reflects the final bill. Depending on usage, the gap between OpenAI and Anthropic varies greatly, with a small model often much more economical.

The price displayed per million tokens is not enough to estimate the real cost of an AI model in business. Based on four concrete tasks, OpenAI and Anthropic result in very different bills, sometimes with a gap ranging from one to two times, or even more. In some cases, the cheapest model in the catalog is not really the cheapest in use. And the “million tokens” sometimes has the elasticity of an old tape measure.

Quand le tarif catalogue ne raconte pas toute l’histoire

The three OpenAI and Anthropic models tested here are not compared on a simple lab benchmark, but on business tasks where the answer must be unique and verifiable. The authors used the official APIs, with the same level of reasoning (“high”), three passes per model, and caching enabled on both sides to match production use. This method matters, as the final cost depends as much on the face price as on the number of tokens actually consumed. This number varies according to the model, the type of task, and the tokenizer, that is, the tool that splits the text into billable units.

On paper, Anthropic is cheaper on several pricing lines. GPT-6 Astra is listed at $10.00 per million input tokens versus $4.00 for Claude Opus 5.5, and $50.00 output versus $20.00. However, GPT-6.1 Sol and Claude Sonnet 5.5 are aligned at $2.00 input and $10.00 output. The same goes for GPT-6 Luna and Claude Haiku 5.5, both at $0.10 input and $0.50 output. At first glance, the ranking seems simple. It is not.

Des écarts qui changent selon la tâche

On the Flask bug-fixing test, GPT-6 Astra ends at $0.4278 versus $0.2534 for Claude Opus 5.5, a cost 1.7 times higher despite a catalog price 2.5 times greater. GPT-6.1 Sol and Claude Sonnet 5.5, though at the same theoretical price, come out at $0.0730 versus $0.1217, giving the advantage to OpenAI. At the low end, GPT-6 Luna costs $0.0048 while Claude Haiku 5.5 rises to $0.0398, a gap of 8.4 times. Anthropic’s models here consumed more tokens for a comparable result, notably because they tokenize the text differently.

The logic repeats on invoice data extraction. GPT-6 Astra remains more expensive, at $0.4023 versus $0.2405 for Claude Opus 5.5. In the mid-range, GPT-6.1 Sol costs $0.0791 and Claude Sonnet 5.5 $0.1284. The smallest model changes the equation again: GPT-6 Luna comes to $0.0041, while Claude Haiku 5.5 reaches $0.0088. Here, Anthropic is only winning at the high end thanks to Claude Opus 5.5’s lower listed price and slightly lower token consumption on this test.

Le long document favorise surtout les petits modèles

The pre-IPO SpaceX prospectus, 400 pages long, creates more marked gaps. GPT-6 Astra is billed $2.9275, versus $1.0770 for Claude Opus 5.5. In the mid-range, the two models are almost equal, with $0.5369 for GPT-6.1 Sol and $0.5358 for Claude Sonnet 5.5. The low end becomes OpenAI’s territory again: GPT-6 Luna drops to $0.0294, versus $0.1458 for Claude Haiku 5.5. On this type of document, Anthropic’s models counted many more input tokens, which increases the bill.

Deduplication of a 300-line client file, including 40 duplicates, confirms the pattern. GPT-6 Astra costs $0.6184 against $0.3451 for Claude Opus 5.5. GPT-6.1 Sol drops to $0.1126, below Claude Sonnet 5.5’s $0.1673. At the low end, GPT-6 Luna goes to $0.0061, where Claude Haiku 5.5 shows $0.0105. On this task, OpenAI’s models consume fewer tokens in output, input, and reasoning, which makes the difference when paying.

Ce que ces mesures disent du marché

Overall, the four tests converge: at comparable listed prices, Anthropic almost always proves more expensive in use. According to the authors, Claude’s tokenizer explains a large part of the gap, with 46 to 63% more tokens consumed depending on the case. In other words, one million Anthropic tokens contains about 35% less text than one million OpenAI tokens. The observation also applies to reasoning: at the same level, Anthropic’s models often produce more tokens.

One point stands out clearly. On each task, a small model completed the work without error for a cost 30 to 90 times lower than the high-end model. The right choice is therefore not necessarily the one with the highest price — nor the biggest ego, if one wants to be polite. For companies making multiple API calls, this difference can quickly change the monthly bill, especially when volumes rise and expected accuracy remains modest.

By the numbers

  • 2.5 times — catalog gap between GPT-6 Astra and Claude Opus 5.5.
  • 8.4 times — gap observed on the bug test, at the low end.
  • 46 to 63% — surplus tokens consumed by Anthropic depending on the case.
  • 30 to 90 times — cost gap between small models and high-end.
  • 400 pages — length of the SpaceX prospectus used for the test.

To read