AI news

Similar AI results can cost 32 times more, Databricks finds

In brief

According to Databricks, 64% of companies using Unity Gateway had used only one model provider in July 2026, compared with 28% on 2 November 2026. The share of tokens sent to open-weight models rose from 7.3% to 30.9% in that environment. GPT-6 Luna and Claude Opus 5.5 each scored 89.1% success, at an average cost of 0.05 and 1.62 dollars per programming task.

Source : Le Monde Informatique - « IA en entreprise : Databricks révèle l’essor des modèles ouverts et du multi-fournisseur »

What it changes for an SME

These figures cover Databricks Unity Gateway customers and internal coding tests; they do not describe a typical SME with 10 to 250 staff. They do show that, in this sample, relying on a single vendor is less common, even counting one-off trials, and that two models with the same success rate can differ about 32 times in cost per task. Spending is highly concentrated: Databricks reports a weekly median of 2.49 dollars per user against 699.74 dollars at the 99th percentile. For you, the practical change is not a forced tool switch, but closer tracking of who spends, comparison of cost per completed task, and not treating token price as the same thing as real cost.

What you can do

  • List every AI provider you already pay for and the monthly amount of each bill.
  • Identify the few people whose AI use is far above the rest of the team.
  • On one recurring task, compare cost per result across at least two models before you lock in a single vendor.
Intro call · 20 min · free

Let's talk about the time you could save.

In 20 minutes, we find where AI could save you time or money. Free, no commitment.

[email protected] Belgium · Europe (remote)
Book a call Call