Databricks launched Smart Routing in Beta inside Unity AI Gateway on August 13, 2026. The feature automatically assigns each coding task to the lowest-cost model that can handle it. Developers using ucode or Omnigent get that routing without choosing a model themselves.
The cost difference is meaningful. On internal coding workloads, Smart Routing cut per-task cost 35% versus a single leading model. On public coding benchmarks, it matched Anthropic's Opus 5 on quality at 56% lower cost per task. Databricks expects savings to grow as it learns from real use cases and design partners.
Why Default Model Selection Gets Expensive
The problem is structural. Databricks counted 33 new AI models released in 2026 alone, spanning a wide range of capability and price. When developers face that many options, many set the most powerful model at the highest effort level and leave it there. Most everyday coding tasks, flipping a flag, a single-file edit, a well-scoped bug fix, do not need frontier-model performance.
Matching tasks to the right model rather than defaulting to the most expensive is one of the biggest cost drivers Smart Routing targets.
What the Feature Actually Requires
Smart Routing is a Beta feature with real prerequisites. The workspace must be in a Unity AI Gateway supported region (currently AWS). An account admin must enable the Smart Routing preview; propagation takes one to two minutes after enablement.
Smart Routing selects only among system.ai-prefixed model services. Custom model endpoints are not currently supported. Each user also needs EXECUTE access to every candidate model the router might select, if any model is inaccessible, the request returns an error rather than falling back silently.
Outside Omnigent, Smart Routing is available only through the ucode CLI. Native Claude Code and Codex sessions run directly, not through ucode, and do not support it. Routing across harnesses rather than just models requires Omnigent version v0.8.0 or later.
For teams using Omnigent, Databricks' meta-harness for coding agents, routing goes further. Omnigent selects both the model and the coding harness for each task, including every sub-agent a session spawns. Databricks says that design can produce savings beyond the reported benchmark figures.
Unity AI Gateway is generally available, providing spend caps, per-user rate limits, and observability. Smart Routing operates inside those controls. Customers including Rivian, Asana, and Edmunds already route production AI traffic through the gateway. Zepto's technical lead reported handling over 100 billion tokens per month with no availability issues.
The Practical Decision
If your team runs coding agents on Databricks today and pays per token, Smart Routing is worth a direct evaluation. Before enabling, confirm your workspace is in a supported region and that users have EXECUTE access to all candidate system.ai models, or the router will error on inaccessible options.
If your team is not yet on Databricks infrastructure, the savings figures still carry context. Matching tasks to model tiers rather than defaulting to frontier can reduce costs substantially. The useful question is whether your current platform offers equivalent routing controls, and whether that changes how you weigh platform options.
Beta status means availability, supported models, and the API are subject to change. Treat the current benchmark numbers as directional. Run your own workload against it before committing to contract or architecture changes based on projected savings.
