In the Enterprise, Loyalty to AI Models Lasts Only a Few Weeks

In the Enterprise, Loyalty to AI Models Lasts Only a Few Weeks
3 min read

An AI model can gain favor with enterprises in a matter of weeks, while another loses almost all its place. This is shown by the Databricks AI Index, a public dashboard updated weekly. It relies on traffic from Unity Gateway, the gateway through which Databricks customers access AI models, and on an in-house code agent benchmark. These figures describe the usage of Databricks customers, not the entire market.

A new model can establish itself in a few weeks

Since May 2026, a new flagship model has been released roughly every week, according to Databricks’ count. By mid-year, most customers relied primarily on a handful of Anthropic models. OpenAI then gained ground with GPT-6 Sol and Astra, before the release of Claude Opus 5.5 in late September gave the advantage back to Anthropic.

In six weeks, Claude Opus 5.5 went from 7.2% to 38% of the tokens consumed on proprietary models, making it the top model on the gateway. Over the same period, Claude Fable 5.1, despite leading the publisher’s code benchmark, fell from 5.5% to 0.6%. ‘Model quality continues to strongly influence vendor market share,’ observes Databricks. The token, a unit of text processed by a model and the basis for its billing, serves here as a measure of usage.

Databricks says it applies the same logic internally, evaluating each new model in the days following its release. The company thus decided not to upgrade from Opus 4.8 to Opus 5, due to insufficient performance gain for its uses relative to the extra cost. However, it made Opus 5.5 the default model for Claude Code after measuring a cost per session 29% lower than its predecessor.

Have enterprises therefore multiplied their suppliers? According to Databricks, about 64% of its customers used only one model provider in July, compared to 28% as of November 2. This indicator is cumulative: it counts all providers used by a company since July 20. The July point therefore covers one week of usage, the November point over three months. It measures the share of companies that have tried multiple providers, without indicating if they use them in parallel.

The most adopted model is not the most economical

The index’s benchmark submits models from nine labs to tasks drawn from the work of Databricks’ own developers, then measures their success rate and average cost per task. At a similar level, cost differences are significant. GPT-6 Luna and Claude Opus 5.5 both succeed in 89% of tasks, but Luna costs an average of $0.05 per task, compared to $1.62 for Opus 5.5, or about 32 times less.

Gemini 3.8 Flash shows a low token price, but it requires an average of 140 exchange turns per task, for a bill of $11.97, one of the highest in the benchmark. Claude Fable 5.1 leads with a 96% success rate, for $7.47 per task. Muse Wonder, from Meta, achieves 95% for $1.53, which, according to the index, is nearly five times cheaper.

Models do not require the same volume of work to succeed. ‘Lower token prices do not guarantee lower costs’ when a model has to do more to complete a task, summarizes the publisher. Databricks, which sells an automatic routing tool between models, advises reserving the most expensive ones for complex tasks.

Open weight models are gaining ground

The index also notes the rise of open weight models, whose weights are published and can be hosted by anyone. Among models served directly by Databricks, their share of tokens rose from 7.3% in July to 31% in early November. This scope excludes traffic routed to Anthropic or OpenAI. The increase was concentrated between July and late September, and the share has plateaued since.

The open weight providers detailed by the index, Zhipu, DeepSeek and Moonshot AI, are all Chinese. Three open weight models appear in the top 10 of the code benchmark, including DeepSeek V4.1 Flash and GLM 5.3, tied for third place with a 93% success rate. On the European side, Mistral AI published the weights of Mistral Large 4 on Hugging Face on October 29, but this model is not among the families tracked by the index.

The index is published by Databricks, without a third-party institute. It aggregates anonymized usage from over 10,000 customer companies, from digital startups to Fortune 500 groups. Usage data covers the period from July 20 to November 2, 2026, and the code benchmark is as of November 10.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x