A Guth Labs publication ↗
Back to News

Databricks opens an AI Index that compares coding-agent cost with performance

Databricks opens an AI Index that compares coding-agent cost with performance

By guthlabs.newsroom

Databricks launched a public AI Index on November 16, combining anonymized Unity Gateway usage with coding-agent evaluations on its own engineering tasks. The company says the resource draws on more than 10,000 companies and will receive weekly updates. Launch announcement

One headline finding is that open-weight models' token share rose from 7.3% in July to 30.9% in November. But the dashboard defines that comparison as traffic to Databricks-hosted models on Unity Gateway. It is not a measurement of all enterprise AI traffic, and its scope differs from charts that also include externally routed requests. Index definitions

The company also reports equal 89.1% task success rates for GPT-6 Luna and Claude Opus 5.5 in its coding evaluation, but average costs of $0.05 and $1.62 per task respectively. Those are workload-specific model-and-harness results, not universal rankings. Tests use Databricks' private codebase, with hidden checks and spending estimated from providers' published prices. Methodology and results

The launch date should not be mistaken for the data cutoff. The dashboard labels usage charts November 2 and its coding frontier November 10. Its multi-provider metric counts every provider an active company has used since July 20, rather than requiring simultaneous use during one week. Dashboard

Guth News analysis: This is useful evidence for evaluating the complete agent system, not just a token price. A company's own tasks, permissions and retry behavior can change the economics. The index provides a comparison starting point, while its population and methodology limit how widely the results can be generalized.

Sources and citations

Revision history

Published .