Databricks has unveiled the Databricks AI Index, an observatory dedicated to the adoption, performance, and costs of artificial intelligence models in the enterprise. Based on anonymized data from over 10,000 organizations using Unity Gateway, its study highlights a diversification of providers and an increase in open-weight models, whose parameters are published. It also underscores significant cost disparities between solutions.
Published on November 16, 2026, this index combines consumption data and evaluations of coding agents—AI software capable of chaining actions to accomplish a task—conducted using internal tasks. The goal is to measure model performance as well as their economic efficiency in professional environments.
Multi-Vendor Adoption Advances
According to Databricks, 64% of companies using Unity Gateway had only used a single AI model provider in July 2026. By November 2, this proportion had fallen to 28%, indicating a diversification of the technologies used. However, this indicator counts all providers used since July 20, including when a company tested an additional provider only once.
Open-weight models also show significant growth. Their share of tokens—the units of text processed by a model—sent to models hosted by Databricks via Unity Gateway rose from 7.3% in July to 30.9% in early November. This figure measures a volume of tokens in this environment, not the proportion of companies that have adopted these models.
This growth is nonetheless showing signs of slowing. According to an analysis by Tomasz Tunguz, the share of open-weight models reached about 33% in September 2026 before stabilizing around 31%, in a context of steep price cuts by OpenAI. This trend suggests that pricing strategies can influence companies’ technology choices, without establishing a causal link on their own.
Major Cost Disparities
In its coding evaluations, Databricks ranks three open-weight models among the top ten. DeepSeek V4.1 and GLM 5.3 notably achieve a success rate of 93.1%, a result that illustrates their competitiveness against proprietary models on the tested tasks.
The study also highlights that comparable performance can lead to very different costs. GPT-6 Luna and Claude Opus 5.5 each achieve a success rate of 89.1%, but their average cost per task reaches $0.05 and $1.62 respectively, a gap of about 32 times.
This finding must, however, be nuanced depending on usage. Databricks indicates it has retained Claude Opus 5.5 as the default model for Claude Code internally, after observing an improvement in quality and a 29% reduction in cost per session compared to its predecessor. Token price alone is therefore not enough to determine a model’s true profitability.
Concentrated Spending
The study also reveals significant disparities in AI spending among employees. The median weekly consumption reaches $2.49 per user, compared to $699.74 for users at the 99th percentile, a gap of about 281 times.
This concentration reflects the diversity of uses, between employees who occasionally consult a conversational assistant and developers who run coding agents daily. For companies, these differences reinforce the importance of budget tracking, usage governance, and resource optimization.
With this AI Index, Databricks says it aims to help organizations compare models based on their performance, costs, and business needs. Unity Gateway notably offers tools for supervision, budget control, and intelligent routing of queries to the most suitable models. The results, however, remain limited to the scope of companies using the platform and the coding tasks evaluated, and do not constitute a representative measure of the entire global artificial intelligence market.
