Open-weight models hit 31% of enterprise AI traffic in four months

Databricks Unity Gateway data shows open-weight models jumped from 7% to 31% of enterprise token traffic between July and November. Chinese labs (GLM, DeepSeek, Moonshot) dominate the category. Single-provider enterprises dropped from 64% to 28% in one quarter.

Open-weight models hit 31% of enterprise AI traffic in four months

Open-Weight Went From 7% to 31% in Four Months

Open-weight models claimed 30.9% of enterprise token traffic by November, up from 7.3% in July, according to Databricks AI Index data from more than 10,000 organisations routing AI through Unity Gateway.

The shift is driven by cost optimisation on high-volume workloads: classification, extraction, summarisation. Enterprises are not replacing Claude on complex problems. They are moving bulk work to models that cost a fraction of frontier pricing.

Chinese labs own the category. GLM (Zhipu), DeepSeek and Moonshot (Kimi) account for the majority of open-weight tokens. GLM 5.2 alone represents roughly 40% of open-weight traffic, one month after mid-July availability.

Single-Provider Strategies Collapsed

Enterprises relying on one model provider fell from 64% to 28% between July and November. Average unique models per customer rose from 4.2 to 7.1 since January.

Enterprises now treat models like cloud regions: routed by workload, swapped when something better ships. A new flagship model launched roughly every 3.7 days since May 2026.

For sales ops and revenue tooling: "built on GPT" is not a moat. Buyers want model flexibility, including open-weight options running inside their perimeter.

Anthropic Owns Proprietary, Opus 5.5 Took Over in Six Weeks

Anthropic's share of proprietary tokens rose from 57.6% to 68.8%. Claude Opus 5.5 went from 7.2% to 37.9% of proprietary traffic in six weeks.

The flip side: Claude Fable 5.1, the top performer on Databricks' coding benchmark, dropped from 5.5% to 0.6% of tokens over the same period. Good enough at lower cost wins.

Cost Per Task Matters More Than Cost Per Token

Databricks' coding benchmark measured real cost per completed task:

  • Claude Fable 5.1: 96.0% success, $7.47 per task
  • Muse Wonder: 94.9% success, $1.53 per task (5x cheaper for 1.1 points less)
  • GPT-6 Luna: 89.1% success, $0.05 per task
  • Claude Opus 5.5: 89.1% success, $1.62 per task (32x more than Luna)
  • Gemini 3.8 Flash: $11.97 per task despite low token price (140 turns per task)

Open-weight DeepSeek V4.1 and GLM 5.3 both hit 93.1%, ahead of Opus 5.5 on this benchmark.

What This Means for Sales Ops AI Stacks

If your product makes a model call on every ticket, email or record, a third of enterprise traffic now runs on models costing a fraction of frontier pricing. Competitors will rebuild COGS on open-weight.

Price per million tokens is the wrong metric. A cheap model that takes 140 turns costs more than an expensive one that finishes in 10. Measure cost per completed task, not cost per token.

Multi-model strategies are now table stakes. Enterprises that run Opus for strategy calls and GLM for ticket classification are optimising margin while maintaining quality where it matters.


Context: Databricks (founded 2013) reported more than US$1.5 billion revenue run rate in 2023, with third-party sources citing US$7 billion run rate and US$190 billion valuation in 2026. The company appointed Adam Beavis as VP and Country Manager for ANZ in 2023, doubling regional headcount to more than 150 employees. Databricks hosts open-weight models and has commercial incentive toward those workloads, so this data likely overstates open-weight adoption relative to the broader enterprise market. The underlying traffic includes Fortune 500 and Global 2000 companies.