AI coding tools have significantly improved development velocity at companies like Databricks, with some teams seeing order-of-magnitude gains in output. However, the rapid adoption of these tools has led to exponentially growing costs that threaten to offset efficiency gains if left unmanaged. To address this, early large-scale adopters including Databricks, Stripe, Coinbase, Uber, and Ramp have converged on a set of approaches aimed at achieving a ‘dual mandate’: providing broad access to AI tooling with minimal friction while keeping aggregate costs within a fixed envelope per user. According to the source, the single greatest lever for cost reduction is shifting to more efficient models as they become available. This relies on the concept of the ‘efficiency frontier’—the set of models offering the best price for a given level of intelligence relevant to typical software engineering tasks. The efficiency frontier is advancing faster than the intelligence frontier, with new models released weekly that offer better intelligence-per-unit-price. To identify which models truly improve cost-efficiency, companies have built internal automated evaluations that better reflect their real-world development needs than public benchmarks. For example, Databricks used such an evaluation to adopt GLM models after observing competitive price/performance. In contrast, Stripe found that Opus 4.7 did not meaningfully improve quality over Opus 4.6 despite higher costs, so they did not deploy it internally. Capturing savings from model shifts requires flexibility in tooling. Some companies ask developers to switch between harnesses (like Claude Code, Codex, or Cursor) when migrating to lower-cost models, though this can create high switching costs and de facto lock-in. A more effective approach is using a meta-harness that provides a consistent user experience while routing requests to underlying harnesses. Databricks uses its open-source Omnigent as a default meta-harness, enabling model and harness independence. Beyond model selection, companies are implementing visibility and progressive friction strategies instead of hard budgets. Every company surveyed provides near-instantaneous feedback on user spend, often with tips to reduce costs by using cheaper models. Hard token budgets are avoided because cutting off access harms productivity and may penalize high-output users whose AI use drives significant gains. Instead, progressive friction increases as spend rises, encouraging efficiency without blocking access. Additional cost-saving techniques include reducing context bloat and leveraging prompt caching, which minimize unnecessary data sent to LLMs during inference.
Key facts
- AI coding tools have improved velocity metrics and driven order-of-magnitude gains in output at Databricks
- The single greatest cost lever is shifting to more efficient models along the advancing efficiency frontier
- Databricks open-sourced Omnigent (meta-harness) and Unity AI Gateway to support cost management
- Stripe found Opus 4.7 did not meaningfully improve quality over Opus 4.6 despite increased cost
- Hard token budgets are ineffective as they hinder productivity and may penalize high-value users
- Companies use developer dashboards to provide near-instantaneous spend visibility and cost-reduction tips