Databricks AI Index: open-weight models now account for 30.9% of enterprise tokens

Databricks launched the Databricks AI Index on November 16, 2026, a public index showing how more than 10,000 companies use and pay for AI models. The data comes from Unity Gateway, the company's tool for accessing and governing models, and includes traffic routed to external providers such as Anthropic and OpenAI. The first edition covers September 21 to November 7, 2026, and Databricks promises weekly updates.

The standout finding is the rise of open-weight models: their share of enterprise token traffic climbed from 7.3% in July to 30.9% in November — more than quadrupling in under four months. At the same time, companies are moving away from relying on a single vendor: the share using only one provider fell from about 64% in July to 28% on November 2.

Opus 5.5 surges among proprietary models

Among proprietary models, the highlight is Anthropic's Claude Opus 5.5: in just six weeks, its share of proprietary-model tokens jumped from 7.2% to 37.9%. Anthropic's total share rose from 57.6% to 68.8% of the proprietary tokens measured by the Databricks AI Index.

Same accuracy, 32 times cheaper

The index also includes a coding-agent benchmark built from real pull requests in Databricks' own codebase. In it, GPT-6 Luna and Claude Opus 5.5 both solved 89.1% of tasks — but the average cost per task was $0.05 for Luna versus $1.62 for Opus 5.5, a difference of roughly 32 times. At the top of the quality ranking, Claude Fable 5.1 solved 96.0% of tasks at $7.47 each, while Muse Wonder reached 94.9% for just $1.53. Gemini 3.8 Flash, despite its low token price, needed 140 turns per task on average and cost $11.97 per task — one of the most expensive in the test.

Spending concentrated among few users

AI spending remains highly concentrated: the median user spent $2.49 per week — double the $1.10 of six weeks earlier — while a user at the 99th percentile spent $699.74, about 281 times more. Databricks notes that spending figures are estimates based on providers' public list prices, not what it bills, and that benchmark results reflect the workloads evaluated.

Featured image: Databricks logo (Databricks Inc., public domain), via Wikimedia Commons.

Fontes

Transparency: this article was produced with the assistance of artificial intelligence, based on cross-checking public posts on X and sources from the internet. The information was verified against the original sources cited below; we recommend consulting them for further details.

Por Muse