On November 16, Databricks launched the AI Index, a public tracker of enterprise AI model adoption, spending, and coding-agent performance. The data comes from more than 10,000 companies using Unity Gateway, plus evaluations on Databricks' own real coding tasks, updated weekly. The question it answers is a practical one: with a new model arriving almost every week, which one should a company switch to, and what is worth paying more for?
Enterprises are switching models faster than expected
At mid-year, most companies still concentrated on a handful of Anthropic models; by November, the picture had fragmented, with OpenAI winning back share on the back of its GPT-6 series Sol and Astra models. But quality still moves share: within six weeks of launch, Claude Opus 5.5 jumped from 7.2% to 37.9% of proprietary-model tokens, lifting Anthropic's overall share from 57.6% to 68.8%. Open-weight models are rising in parallel: they accounted for just 7.3% of enterprise token traffic in July, but 30.9% by November 2. Three of the top ten in Databricks' coding benchmark are open-weight models, with DeepSeek V4.1 and GLM 5.3 both at a 93.1% pass rate, tied for third.
A cheap model can be the most expensive to run
Token prices alone mislead. GPT-6 Luna and Claude Opus 5.5 both achieve an 89.1% task success rate, yet completing one task costs $0.05 on average with the former and $1.62 with the latter — a roughly 32x gap. Gemini 3.8 Flash has a low token price but averages 140 turns per task, ending up at $11.97 per task, among the most expensive in the benchmark. The way you call a model can also swing cost by dozens of times: an OpenRouter test once found the same task burning 42x more tokens over MCP than over CLI. Another pair of numbers shows the trade-off: Claude Fable 5.1 tops the benchmark at 96.0% success and $7.47 per task, while Muse Wonder scores 94.9% at just $1.53. Paying nearly five times more for 1.1 extra points was not worth it to most buyers — Fable 5.1's share fell from 5.5% to 0.6% in six weeks.
Spending is just as concentrated: the median user spends $2.49 a week on AI, while the 99th-percentile user spends $699.74 — a 281x gap — and the median more than doubled from $1.10 in six weeks. Databricks' advice is to route by task: use expensive models for complex, high-stakes work and cheaper ones for routine jobs, and above all test against your own real workloads instead of trusting a leaderboard. The index data covers September 21 to November 7, and its conclusions only reflect the models and workloads evaluated — results may differ in another setting.