Source: Databricks, Introducing the Databricks AI Index: The state of enterprise AI tokenomics (2026-11-16). Shared on LinkedIn by Databricks CEO Ali Ghodsi.
What it is
A public, weekly-updated index built from two sources: aggregated, anonymized traffic from about 10,000 companies routing models through Databricks' gateway (including traffic to external providers), and an internal coding-agent benchmark built from real Databricks engineering workflows (PRs, reviews, investigations) with private-codebase tasks and withheld tests to limit contamination. It reports token and spend share by model and provider, per-user spending, and success rate and cost per completed task for model-plus-harness combinations (Claude Code and Codex, models from nine labs).
Headline findings (as reported)
Launch cadence. Since May 2026 a new flagship model has shipped about weekly, every 3.7 days counting regional labs.
Share moves fast with quality. Claude Opus 5.5 went from 7.2% to 37.9% of proprietary-model tokens in six weeks, lifting Anthropic from 57.6% to 68.8%, after a period where OpenAI's GPT-6 Sol and Astra had gained share.
Multi-provider is now normal. Enterprises using a single provider fell from about 64% in July to 28% by November 2.
Open-weight models quadrupled. From 7.3% of enterprise tokens in July to 30.9% in November. Three open-weight models are in the coding benchmark's top 10; DeepSeek V4.1 and GLM 5.3 tie for third at 93.1%.
Token price is not task cost. GPT-6 Luna and Claude Opus 5.5 both hit 89.1% success, at about $0.05 vs $1.62 per task (roughly 32x). Gemini 3.8 Flash has a low token price but averaged 140 turns per task and cost $11.97 per task, among the most expensive.
The top model is rarely the best value. Claude Fable 5.1 led at 96.0% for $7.47 per task; Muse Wonder scored 94.9% for $1.53. Fable 5.1's proprietary-token share fell from 5.5% to 0.6% in six weeks.
Spend is concentrated. Median user spend was $2.49 vs $699.74 at the 99th percentile (about 281x). Median weekly spend more than doubled over six weeks while the tail stayed flat.
Internal practice. Databricks replays its own traces against the cost-quality frontier before upgrading: Opus 5.5 gave better quality at 29% lower cost per session than its predecessor and became the default, while an earlier Opus 4.8 to Opus 5 upgrade was judged not worth the cost.
What it means for engineering teams
Measure cost per completed task, not per token. Turn count, output length and retries dominate; a cheap model that loops is expensive.
Evaluate model and harness together. The same model ranks differently under different harnesses, and five models from four labs sit on the frontier.
Build for switching. With weekly launches and share swinging 30 points in weeks, a gateway or abstraction that lets you swap models per task is cheaper than re-platforming.
Budget for the tail. A few heavy agent users drive most spend; per-user visibility matters more than average cost.
Caveats
Vendor data from the vendor's own gateway customers, who skew toward multi-provider routing by construction.
The coding benchmark is built on Databricks' own codebase and workflows; results may not transfer.
The CEO's LinkedIn summary phrases provider counts differently (75% with two or more, 50% with three) from the blog's single-provider figure.
The post promotes the vendor's gateway and routing products.
Review checklist
Do you track success rate and cost per completed task for each model and harness you use?
Can you replay your own traces against a new model before switching the default?
Is model choice configurable per task type, with a cheaper default for routine work?
Do you have per-user and per-agent spend visibility, with alerts on the tail?
Related entries
R113 (cutting coding-agent cost from session history), R111 (the open-weights gap), and the inference-provider entries under index subtopic 2.4.