Databricks has achieved a 70% reduction in AI coding costs by implementing a suite of cost management techniques, according to a blog post by company leaders including Patrick Wendell and Akshat Bhatia.
The company faced a common challenge: while agentic coding measurably improved velocity metrics and delivered order-of-magnitude gains in output for some teams, costs grew exponentially. Left unchecked, this spending curve threatened to undermine the efficiency gains AI provided.
Based on experience at Databricks and conversations with other digital-native companies including Stripe, Coinbase, Uber, and Ramp, the company identified several proven cost management approaches.
The single largest cost lever, according to the post, is moving coding spend to more efficient models as they are released. Databricks distinguishes between the “intelligence frontier”—the highest-performing models—and the “efficiency frontier,” which represents models with the best price-to-intelligence ratio for typical software engineering work. The efficiency frontier advances far faster than the intelligence frontier, with new models released almost weekly that offer better cost-per-quality than predecessors.
To identify which models actually outperform incumbents, companies need automated evaluations representative of their internal development mix, since public benchmarks poorly predict real-world coding performance. Databricks built such benchmarks and found GLM models competitive on price-performance, leading to internal rollout. However, Stripe found that Opus 4.7 did not meaningfully improve quality over Opus 4.6 while increasing cost, so they declined to offer it internally.
To preserve flexibility when switching models, companies increasingly adopt a “meta-harness”—a tool that surfaces a common user experience while dispatching requests to underlying harnesses and models. This reduces developer switching costs compared to asking users to manually switch between tools like Claude Code, Codex, or Cursor.
Automatic model and tool selection based on task requirements offers additional efficiency gains over letting users choose models themselves.
Counter to intuition, hard spending budgets that cut off access when thresholds are hit proved ineffective across all companies surveyed. Such limits debilitate productivity and discourage power users who achieve monumental efficiency gains. Instead, companies adopted progressive approaches emphasizing visibility and increasing friction as spend rises. Databricks provides near-instantaneous feedback on spend through developer dashboards with tips for cost reduction using cheaper models.
Databricks has open-sourced key infrastructure components including Omnigent (an end-user meta-harness) and Unity AI Gateway to help companies implement these techniques.
Key facts
- Databricks reduced AI coding costs by 70% through systematic efficiency improvements
- The efficiency frontier—models with best price-to-quality ratio—advances faster than peak intelligence models
- Databricks built automated benchmarks to evaluate new models rather than relying on public benchmarks
- The company uses a meta-harness approach to allow model switching without requiring developers to change tools
- Hard spending caps proved ineffective; companies instead provide cost visibility and progressive friction as spending increases