Databricks Says Unity Gateway Cut Its AI Coding Costs by Up to 90%
Databricks says it cut internal AI coding costs by as much as 90% using Unity Gateway, an internal system that routes tasks, enforces spend controls, and shifts work to cheaper models.
Quick answer
What did Databricks announce about cutting its AI coding tool costs?
Databricks says it cut its internal AI coding costs by as much as 90% using Unity Gateway, a system that routes tasks to the cheapest capable model, applies escalating spend controls instead of hard caps, and gives engineers visibility into their own usage, while reserving costlier models like Astra for complex problems.
Key takeaways
- Databricks says its internal AI coding costs grew exponentially after a broad rollout because the tools are priced by consumption rather than by seat, but token spend per engineer is now roughly flat even as usage and value have risen.
- Databricks replaced hard spending caps with Unity Gateway, an internal system that applies warnings, required approvals, and automatic downshifts to cheaper models before cutting access, which the company says avoids disrupting work mid-task.
- Databricks says four levers, cheaper model defaults, task-based routing, spend visibility, and reduced token overhead, have cut its unit AI costs by as much as 90%, with switching to cheaper models alone accounting for 50% or more of the savings.
- Using a benchmark built from real pull requests on its own production codebase, Databricks says the open-source model GLM 5.2 matched Opus 4.8 on quality at about a third of the cost per task.
- Databricks rolled out a model called Astra to roughly 3,500 engineers and says it is the first company to publicly share such a rollout; spend rose about 60% before the company restricted Astra to harder problems through its own budget in Unity Gateway.
Rising costs, then a flattening
In a blog post published October 22, 2026, Databricks said its internal spending on AI coding tools grew exponentially after it rolled the tools out broadly to engineers, because the tools are priced by consumption rather than by seat. Patrick Wendell, a Databricks co-founder who oversees the company's AI products, said costs began threatening to erase the productivity gains the tools were meant to produce. Databricks says its token spend per engineer is now roughly flat even as total usage and the value generated have continued to rise.
From hard caps to a gateway
Wendell said Databricks' first attempt to control costs, a hard spending cap, caused problems because it cut off work mid-task and tended to affect its most active users, some of whom moved to personal accounts the company could no longer monitor. Databricks says it replaced hard caps with an internal system called Unity Gateway that applies escalating steps before cutting anyone off: first a warning, then a required approval, then an automatic downshift to a cheaper model. Databricks says cutting someone off is treated as a last resort. The company also said it sends engineers regular breakdowns of their own AI spending, which it says accounts for roughly a 10% reduction on its own.
Four levers and their impact
Databricks described four methods it uses to lower unit costs: shifting default usage to newer, cheaper models as they are released; routing individual tasks to the least expensive model capable of handling them; giving developers visibility into their own spending; and reducing token overhead. Wendell said moving traffic to better, cheaper models as they appear has been the single biggest factor, cutting costs by 50% or more on its own. Combined with the other three levers, Databricks says unit costs have fallen by as much as 90% in some cases. The company's Smart Router, part of Unity Gateway, is said to cut average task cost by more than 30% while keeping output quality close to that of the most expensive model.
Internal benchmarking
Databricks said it built its own benchmark from real pull requests in its multi-million-line codebase, grading results by running tests rather than using another AI model as a judge, because public benchmarks do not reflect how tools perform on a large enterprise codebase. Using this benchmark, Databricks said the open-source model GLM 5.2 matched Opus 4.8 on quality while costing about a third less per task, and that a model priced 1.7 times cheaper per token actually cost more per task because it used more tokens. The company also said the harness, or coding tool wrapped around a model, affected cost about as much as the model choice, producing more than a 2x cost difference for the same model, and that tuning harness and caching settings alone saved close to 50%.
Astra rollout
Databricks said it rolled out a model called Astra to about 3,500 engineers and says it is the first company to share such a rollout publicly. The company said Astra performed better on complex tasks such as system design but increased spend by about 60% without a clear benefit on routine tasks, prompting Databricks to give Astra its own separate budget within Unity Gateway restricted to harder problems.
All coding-tool traffic at Databricks, including Claude Code, Codex, Cursor, and the company's own Omnigent harness, is routed through Unity Gateway, which Databricks says began as an internal fix for its own AI spending and became the basis for the gateway.
Source: Databricks Blog, "The AI Bill is coming," published October 22, 2026.
Frequently asked questions
- What is Unity Gateway?
- Unity Gateway is Databricks' internal system for routing all AI coding tool traffic, handling access, model routing, spending budgets, and tracing. Databricks says it began as an internal fix for its own rising AI bill.
- How much did Databricks cut its AI coding costs by?
- Databricks says combining cheaper model defaults, task routing, spend visibility, and reduced token overhead cut its unit AI costs by as much as 90% in some cases.
- What happened when Databricks tried hard spending caps?
- Databricks said hard caps disrupted work mid-task and pushed some of its most active users toward personal accounts outside company visibility, so it replaced caps with escalating warnings and automatic model downshifts.
- What did Databricks find when it rolled out the Astra model company-wide?
- Databricks rolled out Astra to about 3,500 engineers and says spend rose roughly 60%, with clear gains on complex tasks like system design but no obvious benefit on routine work, prompting it to give Astra a separate, restricted budget.
Sources
- The AI Bill is coming | Databricks Blog — Databricks