Databricks Cuts AI Coding Costs 70%: Model Flexibility, Open Source, and the Efficiency Frontier

Databricks reports a 70% reduction in AI coding spend while maintaining developer velocity, using a combination of aggressive model switching, internal benchmarking, and infrastructure they've open-sourced. The key insight: most coding doesn't need frontier intelligence, so the real target is the efficiency frontier — best price for a given quality bar. Here's what worked.

Key Cost Levers

• Moving to open-source and lower-cost models — The biggest single lever. Databricks built an internal benchmark that found GLM models offered competitive price/performance, leading to a company-wide rollout. Stripe similarly tested Opus 4.7 but declined to deploy it because it cost more without improving quality. Databricks saw the same with Opus 5.0 vs 4.8.
• Harness and model flexibility — To adopt new models quickly, you need tooling that lets you switch. Databricks open-sourced Omnigent (an end-user meta-harness) and Unity AI Gateway to route traffic across models. They also let developers use familiar harnesses (Claude Code, Codex, Cursor) but direct them to cost-efficient models via the gateway.

The Efficiency Frontier

Frontier labs optimize for peak intelligence, but day-to-day coding doesn't require math proofs or novel security exploits. The efficiency frontier — models that give you the best intelligence per dollar — is advancing much faster. Public benchmarks fail to capture real-world coding performance, so Databricks and peers build internal evals that mirror their own dev workloads.

What This Means for Your Team

If you're managing AI coding costs, the playbook is:

• Build or adopt internal benchmarks tailored to your codebase to evaluate new models as they ship.
• Be ready to switch models quickly — don't lock into a single vendor.
• Use a gateway to dynamically route requests to the cheapest model that meets quality bars.
• Monitor cost regressions when models update; sometimes the old model is still the better economic choice.

Databricks claims these techniques can keep aggregate costs in a fixed envelope per user, even as usage grows. For more details and the full tech stack, read their post.