Closed AI models now offer the best value at almost every budget. Village's LLM Atlas found that an open-weight model was the best buy at just 3% of price points on October 31, 2026, down from 41% in June. The researchers offer three explanations: distillation, broader product lineups and lower serving costs at large closed labs.

Are closed AI models better value than open models?

At almost every budget, the best AI model you can buy is closed. An open-weight model wins at just 3% of price points, according to Village's LLM Atlas, down from 41% in June.

The LLM Atlas maps the price-performance frontier, which shows the strongest AI model available at each price. The analysis compares every model released by June 30 and by October 31, 2026, measuring capability against estimated cost per task relative to GPT-5 at high effort. In June, eight of the 19 points on the frontier were open-weight models, including DeepSeek-V4-Flash, GLM-5.2, Gemma 4 E4B IT, Hy3 preview and Llama 3 8B Instruct. By October 31, only five of 21 were, and all of them sat in the cheaper half of the frontier.

Those remaining open models are concentrated at the bottom of the price range. Google's Gemma 4 E4B and Xiaomi's MiMo-V2.6-Flash hold most of that 3% share.

Two OpenAI releases took much of the ground back. GPT-6 Luna arrived in September and GPT-6.1 Terra in October. Together they now cover the budgets open models once won, alongside Anthropic's Claude Haiku 5.5.

Open models have led on value before

How has the open-model share of budgets changed since 2024?

The shift is not a straight line. Open models have held the top spot at certain points since 2024. Their share of budgets at the end of each quarter ran as follows: 6% in the first quarter of 2024, 42% in the second, 19% in the third and 32% in the fourth. It dipped to 7% in the first quarter of 2025 and 7% in the second, then jumped to 52% in the third quarter of 2025. The share stood at 30% in the fourth quarter of 2025, 26% in the first quarter of 2026, 41% in the second and 20% in the third. On October 31, 2026, it was 3%.

The researchers said today's figure is the lowest they have measured. Open models held at least 6% of budgets at the end of every quarter since 2024.

The peaks tended to come from single releases. DeepSeek-V3.2-Exp was the best buy at nearly half of all budgets during the week it launched in September 2025. None of those gains lasted. After each peak, the open share fell by about half or more by the next quarter as closed labs released cheaper tiers.

Open models, in other words, win on value in bursts. A strong, inexpensive release, usually from DeepSeek, claims a wide band of budgets until a closed model undercuts it. The latest closed releases have been both cheaper and stronger than anything open, which has squeezed that band down to the bottom of the price range.

Why closed labs may be winning on value

Why are closed AI models winning on value?

The researchers offer three explanations that fit the pattern. The first is distillation. A lab with the strongest model can train cheaper versions on that model's outputs in-house. Today's cheap closed tiers followed their flagships within weeks. Claude Haiku 5.5 arrived two weeks after Claude Opus 5.5. Open labs, by contrast, have to start from their own weaker models, because closed labs' terms generally bar training competitors on their outputs.

The second is product breadth. OpenAI's GPT-6 and GPT-6.1 families, which run from Luna and Terra up to Sol and Astra, hold 11 of the frontier's 21 points between them. One generation covers nearly every price.

The third is serving cost. OpenAI, Anthropic and Google run inference at very large scale, and Google does so on its own chips. That can lower what each task costs. Most open-model labs serve far less traffic.

The analysis relies on capability scores that Atlas estimates for each model and reasoning setting, based on public benchmark results. Costs come from what evaluators paid through hosted APIs, which may not reflect the cost of running open models on your own hardware. Restricting the analysis to models scored on at least 10 benchmark families, or narrowing the budget range to 3% to one times GPT-5's cost, produces the same pattern. Widening the range to 0.3% to ten times GPT-5's cost raises today's open-model share to 16%, because very small open models still win at the cheapest end. The scores reflect data as of October 31, 2026, and Atlas updates as new results come in.

Questo articolo è stato scritto con l'assistenza dell'IA.
News Factory APP - notizie agentiche per potenziare il tuo SEO e AEO.