Using the right model and the right prompt is only part of the enterprise AI challenge, it's also critical to optimize the prompt.
It's a challenge that Databricks has been working to solve with its Agent Bricks technology, which was launched back in June. That technology has steadily improved and today Databricks revealed new techniques it is using to further improve prompt optimization. New research released today from the company shows how its GEPA (Generative Evolutionary Prompt Adaptation) technique improves prompt optimization by an order of magnitude. Databricks claims the enhancement to Agent Bricks can now enable enterprises to make models up to 90x cheaper to operate.
The breakthrough in prompt optimization arrives alongside Databricks' $100 million partnership with OpenAI. This deal makes GPT-5 natively available to Databricks' enterprise customers, following similar deals that Databricks had previously announced with Anthropic and Google. To be clear, neither Databricks nor OpenAI is paying $100 million to the other; rather, the figure represents an expectation by Databricks of the potential revenue that the partnership will bring.
While the integration and partnership with OpenAI is noteworthy, the real story is how Databricks' research and advanced prompt optimization techniques prove enterprises don't need premium prices for premium AI performance.
"Prompt optimization is not really taking an existing query and just optimizing its execution, It's actually changing the query itself," Hanlin Tang, Databricks' Chief Technology Officer of Neural Networks at Databricks, told VentureBeat. "It's like, what is the best way to ask the LLM the question to get the high-quality answer that I want?"
GEPA rewrites the optimization playbook
The breakthrough technique is GEPA (Generative Evolutionary Prompt Adaptation), developed by researchers from Databricks and the University of California, Berkeley. Unlike traditional fine-tuning that adjusts model weights, GEPA optimizes the questions enterprises ask AI systems.
The approach mirrors human communication patterns.
"In the LLM world, there are different ways to ask the LLM the same question, right? Just like there's a different way to ask you a question on a quiz," Tang said. "There's like 10 different ways to ask a question about a particular fact."
GEPA uses an approach called natural language reflection, where the AI critiques its own outputs and iteratively improves them. This feedback loop automatically discovers optimal prompting strategies for specific enterprise tasks.
Results across finance, legal, commerce and healthcare domains show GEPA-optimized models consistently outperformed baselines by 4-7 percentage points.
Rewriting enterprise AI economics
The cost transformation stuns at enterprise scale. At 100,000 requests, Databricks' optimized open-source model delivers superior quality at 1/90th the serving cost of Claude Opus 4.1.
"If you're able to prompt optimize to improve the quality of a model on your task, you can also use it to sort of bring up a smaller model to the quality that you care about, so you can actually save cost as well," Tang noted.
The advantage compounds with volume. For workloads processing 10 million requests, one-time optimization costs become negligible compared to serving costs.
GEPA also outperforms supervised fine-tuning—the current gold standard for model customization—while reducing serving costs by 20%. The technique saves engineering resources, too.
"It also saves engineers and data scientists a lot of time, because usually they spend a lot of time prompting, writing the right prompt and question for the model," Tang explained. "In this case, the system can figure out automatically what the best way to query the model is."
OpenAI integration eliminates complexity
While GEPA optimization can enhance the performance of any model, the technique becomes even more powerful when enterprises can easily access and experiment with multiple high-quality models. This is where the OpenAI partnership creates a force multiplier effect.
"The most important component for us and for OpenAI is it now makes the OpenAI models natively available on Databricks," Tang said. "Any Databricks customer can query these OpenAI GPT-5 models without an external vendor relationship, without an API key."
The integration goes beyond simple API access. Enterprises can call GPT-5 directly in SQL commands. "They can easily call a SQL command and just sort of quote GPT-5 in the command to ask it to translate a row in the table, or something like that," Tang explained.
This native integration eliminates the vendor management overhead that previously complicated the deployment of premium models. "That's just part of your Databricks plan. You don't need to create an API key somewhere else," Tang confirmed.
The partnership reinforces Databricks' multi-model strategy alongside existing Anthropic and Google Gemini integrations.
"We're all about having Databricks being a multi-model platform," Tang emphasized.
Enterprise action plan
Technical decision-makers should prioritize three immediate steps:
Build robust evaluation first: Tang's core advice: "One of the superpowers of Agent Bricks is that it builds custom evaluations on your specific task. I would recommend enterprises build agents first – you don't want to be flying blind without knowing the quality of your agents."
Question fine-tuning defaults: With prompt optimization matching or exceeding supervised fine-tuning results at lower costs, enterprises should evaluate both approaches rather than defaulting to traditional fine-tuning.
Rethink model procurement: Post-deployment optimization capability changes the buying decision. A more affordable model that optimizes for frontier performance may deliver better value than premium upfront pricing.
For enterprises looking to lead in AI deployment, the message is clear: the cost barrier to frontier AI performance has collapsed. Early adopters who invest in optimization capabilities now will build increasingly insurmountable competitive advantages as their AI systems continuously improve.
