Open models are the best buy at only 3% of budgets, down from 41%

The LLM Atlas index shows that an open model is the best buy at only 3% of budgets. The scores are from 31 October 2026. In June the share was 41%. This is the lowest share the index has measured. OpenAI released GPT-6 Luna in September and GPT-6.1 Terra in October.

How it works

The index finds the strongest model at 201 price points, from 1% to three times the cost of GPT-5. It uses public test results. The index gives three causes for the success of closed models:

• Closed labs train cheaper models on the output of their own best models.
• OpenAI holds 11 of the 21 points on the frontier, the best models.
• Closed labs can possibly use models at a lower price, at a larger scale.

The numbers

DeepSeek V4 Flash alone was the best buy at almost a third of budgets. Each peak came from one release. After each peak, the open share fell by half or more by the next quarter. Claude Haiku 5.5 came two weeks after Claude Opus 5.5.

What is not known

The costs are the price that testers paid through hosted APIs, which are paid services that vendors operate. Open models can cost less on your own hardware. With a wider price range, the open share is 16%. Tiny open models win at the lowest prices.

Insight: Open models lead in use, but not in best value

Vercel says open models handled more than 7 in 10 tokens on its gateway in September. Dell says 80% of tokens worldwide, but gave no data. But the index gives open models 3% of best buys. This is maybe a divide: open models win on volume and control, while closed labs set the price ladder. Look for the next cheap DeepSeek release. Test open models on your own hardware.