Databricks AI Index: Open-Weight Models Jump to 30.9% of Tokens as Enterprises Go Multi-Model

Databricks' new AI Index, based on Unity Gateway data from more than 10,000 companies, shows open-weight models rising from 7.3% to 30.9% of tokens on Databricks-hosted models since July and single-provider companies falling from 64% to 28%. Its coding benchmark finds cost per task varies up to 32x at equal quality.

Christopher Ort
By Christopher Ort · 9 min read
Share
A network engineer seen from behind patches colored fiber cables between server racks in a dim data center aisle.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. The key findings
  5. Background: why Databricks can see this
  6. How the coding benchmark works
  7. Pros and cons
  8. What X is saying
  9. Critical analysis
  10. What's still open
  11. About the analysis
  12. i10x Perspective

For most of 2026, the enterprise AI story was simple: pick a frontier lab, sign the contract, build on it. Databricks now has data saying that era is ending fast. Its new AI Index, built on traffic from more than 10,000 companies that route models through its Unity Gateway, shows open-weight models jumping from 7.3% to 30.9% of tokens on Databricks-hosted models in under four months, and the share of companies relying on a single model provider collapsing from about 64% to 28%. The thread announcing it trended on X on Monday evening. Here is what the numbers say, and where the fine print matters.

i10X Quick Take

Quick Take

  1. What: Databricks launched the Databricks AI Index on November 16, a public, weekly updated tracker of which models enterprises use, what they spend and how coding agents perform on real tasks.
  2. Headline numbers: Open-weight models rose from 7.3% to 30.9% of tokens on Databricks-hosted models since July. Companies using only one provider fell from about 64% to 28% by November 2.
  3. Cost surprise: In Databricks’ coding benchmark, GPT-6 Luna and Claude Opus 5.5 both solved 89.1% of tasks, but Luna cost $0.05 per task versus $1.62 for Opus 5.5.
  4. Who is affected: CIOs and platform teams deciding whether to standardize on one lab, and developers choosing models and coding agents.
  5. The catch: The data only covers Databricks customers, spend is estimated at list prices, and Databricks sells the routing product that benefits from a multi-model world.

In plain English

Databricks, a company that helps businesses run their data and AI, looked at how more than 10,000 of its customers actually use AI models. It found that companies are no longer sticking with one AI provider, and that free-to-download “open” models now handle almost a third of the work on its own hosted models, up from under a tenth in July. It also found that the most expensive model is rarely the best deal, because a cheaper model often gets the same job done. The numbers come from one company’s customers, so they show a strong trend, not the whole market.

What happened

On November 16, Databricks published a blog post titled “Introducing the Databricks AI Index: The state of enterprise AI tokenomics” and opened a public index site. The post, credited to “The Databricks AI Product and Engineering Team,” says the company is making its internal evaluation suite “permanently public, with weekly updates as we evaluate new proprietary and open-weight model releases.”

Databricks co-founder and VP of engineering Patrick Wendell launched it with a six-part thread on X that had passed 100,000 views by Tuesday morning:

The Index combines two data sources. The first is aggregated, anonymized usage and spending data from 10,000+ companies on Unity Gateway, including traffic routed to outside providers such as Anthropic and OpenAI. The second is a coding-agent benchmark built from real workflows of Databricks engineers, such as creating and reviewing pull requests, run against tasks from Databricks’ private codebase with withheld tests. The current dataset covers September 21 through November 7, 2026.

The key findings

  • Model churn is fast: Since May 2026, Databricks counts a new flagship proprietary or open-weight model roughly every week, or every 3.7 days when regional labs are included.
  • Anthropic bounced back with Opus 5.5: OpenAI’s share of enterprise token traffic “grew meaningfully,” driven by its GPT-6 series Sol and Astra models. Then Claude Opus 5.5 went from 7.2% to 37.9% of proprietary-model tokens in six weeks, lifting Anthropic’s overall share from 57.6% to 68.8%.
  • Multi-provider is the norm: About 64% of Unity Gateway companies used one provider in July. By November 2, that was 28%.
  • Open-weight is surging: Open-weight models went from 7.3% to 30.9% of tokens on Databricks-hosted models. In the week of November 2, Zhipu’s models held 60.4% of open-weight tokens (down from 92.3% in July), DeepSeek 28.2% (from 0.0%) and Moonshot 8.8% (from 1.4%). The single biggest models were GLM 5.3 Flash (33.3%), DeepSeek V4.1 Flash (28.0%) and GLM 5.3 (21.6%).
  • Spend is concentrated: The median user spent $2.49 on AI in a week, while the 99th-percentile user spent $699.74, roughly 281 times more. Median weekly spend more than doubled from $1.10 over six weeks, while the top percentiles stayed flat.

Selected coding-benchmark results as reported in the Databricks blog:

ModelTask success rateCost per task
Claude Fable 5.196.0%$7.47
Muse Wonder94.9%$1.53
DeepSeek V4.193.1%not stated
GLM 5.393.1%not stated
GPT-6 Luna89.1%$0.05
Claude Opus 5.589.1%$1.62
Gemini 3.8 Flashnot stated$11.97 (140 turns per task)

Background: why Databricks can see this

Unity Gateway is Databricks’ central access and governance layer for AI models. Companies use it to reach Databricks-hosted models and external providers through one place, with usage dashboards, budgets and a Smart Routing feature that picks a model per task. That vantage point is what makes the Index possible, and it is also why the numbers have to be read with care (more on that below).

The coding benchmark builds on one Databricks published in July. Databricks says it uses the same approach internally: it rolled out Opus 5.5 as the default model for Claude Code after finding “improved quality at 29% lower cost per session than its predecessor,” but decided not to upgrade from Opus 4.8 to Opus 5 because replaying its own traces showed “the upgrade wasn’t worth the additional cost for our use cases.”

The shift fits a pattern we have tracked in our Enterprise & Work coverage, including the move toward total cost of ownership and model routing and the Opus 5.5 surge inside coding tools.

How the coding benchmark works

  • Real tasks: Workflows from Databricks engineers captured through Unity Gateway, from pull requests to issue investigations.
  • Hidden tests: Results are graded against withheld tests, not an LLM judge, which reduces the risk of models having seen the answers.
  • Model plus harness: Models from nine AI labs are run inside Claude Code and Codex, because the same model can behave very differently in different agent harnesses.
  • Two scores: Task success rate and cost per completed task. Five models from four labs (OpenAI, DeepSeek, Meta and Anthropic) sit on the cost-quality frontier.

The most striking finding is about cost. Gemini 3.8 Flash has a low price per token, yet it averaged 140 turns per task and ended up one of the most expensive models at $11.97 per task. Claude Fable 5.1 tops the benchmark at 96.0% but costs $7.47 per task, while Muse Wonder reaches 94.9% for $1.53. Databricks notes that Fable 5.1’s share of proprietary tokens fell from 5.5% to 0.6% in the six weeks after its launch. Three open-weight models made the top 10, with DeepSeek V4.1 and GLM 5.3 tied for third at 93.1%.

Pros and cons

For users

  • Pro: More competition between labs and open models tends to push prices down and quality up for the AI tools employees use at work.
  • Con: When companies switch models every few weeks, the assistant you rely on can change behavior without notice.

For developers

  • Pro: A public, contamination-resistant benchmark that reports cost per task is far more useful for picking a coding setup than price-per-token tables.
  • Con: The tasks come from Databricks’ own codebase, so results may not transfer to very different stacks.

For companies

  • Pro: The data supports a multi-provider strategy and suggests real savings from routing routine work to cheaper or open-weight models.
  • Con: Running several providers adds governance, security review and evaluation work, and the savings depend on measuring your own workloads.

What X is saying

Wendell’s explanation for why the most capable models barely show up in the data got almost as much attention as the launch itself:

In the same thread he wrote that open-weight models are used “mostly to reduce costs on high volume workloads (large batch workloads, coding, etc).” Zhipu was quick to claim its piece of the result:

Others focused on cost. Account @liro90106788 picked out DeepSeek on the frontier and asked of Claude Fable 5.1’s lead: “+2.9 pts for 7.3x the cost?” Automation builder @xyzhani drew a practical lesson: “don’t marry one model. Route each task to the one that fits it.” Data engineering manager @SulArifL pushed back on the idea that switching models solves much: “Swapping a model is one config change. Messy tables, fuzzy metric definitions and broken permissions follow you to every model.”

We did not find a Grok summary or reply on the Index at the time of writing.

Critical analysis

The direction of the data is convincing. Several signals point the same way: falling single-provider use, rising open-weight share, and a benchmark showing that cheaper models often match expensive ones. Publishing cost per completed task, rather than price per token, is a genuinely useful correction to how models are usually compared.

But four caveats matter. First, the 30.9% figure is narrower than many posts suggest. Databricks’ blog calls it a share of “enterprise token traffic,” while the chart on the Index site defines it as “tokens to Databricks-hosted models on Unity Gateway.” That appears to leave out the traffic Unity Gateway forwards to external providers such as Anthropic and OpenAI, so it is not 30.9% of all enterprise AI usage. Second, the multi-provider numbers vary by source: the blog’s 28% single-provider figure implies 72% with two or more, Wendell wrote “75% have two or more,” and Databricks’ Yuchen Jin wrote that “71% of active enterprises have used 2+ model providers.” Different time windows probably explain it, but Databricks has not reconciled the figures publicly.

Third, spend is estimated at public list prices, and the site states it is “not what Databricks bills.” Negotiated discounts, which large enterprises routinely get from frontier labs, are not reflected. Fourth, Databricks is not a neutral observer. It sells Unity Gateway and Smart Routing, and a world where companies juggle many models is exactly the world where a routing layer is valuable. That does not make the data wrong, but it is worth keeping in mind.

What’s still open

  • Whether Databricks will publish total market share across hosted and external traffic, not only hosted-model shares.
  • How the numbers look once spend is weighted by actual contract prices rather than list prices.
  • Whether the open-weight surge holds as capacity constraints at open-weight providers ease or tighten.
  • How the next frontier releases from OpenAI and Google shift the proprietary split in the weekly updates.

About the analysis

Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated image. Read our AI Use Policy. Spotted an error? Request a correction.

This independent i10x analysis is based on the Databricks blog announcement, the public Databricks AI Index site and its chart definitions, and public posts on X by Databricks staff, Zhipu staff and developers. All figures are as published by Databricks at the time of writing; the Index is updated weekly, so later values may differ.

Sources (plain text, no links):

  • Databricks Blog: Introducing the Databricks AI Index: The state of enterprise AI tokenomics (November 16, 2026)
  • Databricks AI Index (public index site, data updated November 2, 2026)
  • Patrick Wendell, Databricks co-founder, launch thread and replies on X (November 16, 2026)
  • Yuchen Jin, Databricks, post on X (November 16, 2026)
  • Zixuan Li, Zhipu (Z.ai), post on X (November 17, 2026)

i10x Perspective

The most important number in the Databricks AI Index is not 30.9%. It is $0.05 versus $1.62 for the same 89.1% success rate. For a year, enterprise AI buying has been driven by leaderboards and brand loyalty. This data says the winners in enterprise budgets will be the teams that measure cost per finished task on their own work and route accordingly. That is good news for open-weight labs and for buyers. It is less comfortable for frontier labs whose top models are, in Wendell’s words, competing with “good enough.” Just remember who collected the data, and that a routing vendor will always find that routing matters.

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Timeline

What happened before

  1. Release

    Claude Haiku 5.5

    1M context · $0.1 / $0.5 per 1M tokens

    support.claude.com (source, opens in a new tab)
  2. Release

    Claude Sonnet 5.5

    1M context · $2 / $10 per 1M tokens

    support.claude.com (source, opens in a new tab)
  3. Release

    Claude Opus 5.5

    1M context · $4 / $20 per 1M tokens

    First model of the Claude 5.5 family; performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.

    anthropic.com (source, opens in a new tab)
  4. Partnership

    Accenture partnership on independent frontier-AI evaluation (≥$1B each over five years)

    With Accenture · Research

    anthropic.com (source, opens in a new tab)
  5. Release

    Claude Fable 5.1 & Mythos 5.1

    1M context · $10 / $50 per 1M tokens

    Anthropic's most advanced models for coding and knowledge work. Price/context shown for Claude Fable 5.1 (models overview).

    support.claude.com (source, opens in a new tab)
Full Anthropic timeline ›

Editor's picks

  1. Genesys Launches Standalone AI Agents and Arena, With Pricing Based on How Hard Each Task Is
  2. Seedance 2.5 4K Reaches Developers: 10-Bit Video up to 30 Seconds With 480p Draft Mode
  3. Perplexity Computer Adds Multiagent Mode, Where Up to 16 Agents on Rival Models Check Each Other
  4. Gemini Can Now Make Calls, Pay Bills and Shop, Free in Six Countries With Gemini 4 Hydrogen
  5. Hinton on Firing Line: AI Is Making ‘Alien Beings’ and Self-Regulation Is ‘Absurd’