The Challenge
Starburst Data, a data platform vendor with over $100 million in annual recurring revenue, faced a common enterprise issue: unpredictable AI costs. The company relies heavily on Anthropic's Claude for AI development, with AI tools generating about a third of its code. This has significantly increased development speed, cutting the time from idea to production by up to 60%. However, CEO Justin Borgman found that the finance team lacked visibility into AI spending.
"We don't have good predictive modeling of our spend today," Borgman said. While token costs weren't out of control, they were rising faster than anticipated due to increased consumption. The core problem is that AI work costs become clear only after the model completes its tasks.
This issue isn't unique to Starburst; it's a structural challenge. Enterprises are billed for input tokens (what you send to the model), output tokens (what it returns), and reasoning tokens (the model's internal processing). As Pega's COO and CFO Ken Stillwell noted, this third category operates as a "black box." Anthropic's documentation confirms that customers are billed for a model's full internal reasoning, with usage varying by task.
The bill essentially writes itself, a unique phenomenon in commercial software procurement.
Navigating Constraints
Starburst operates under typical enterprise constraints: the need for cost predictability, accountability to finance leadership, and scaling AI usage across teams. With Gartner forecasting global AI spending to reach $2.6 trillion this year, CFOs are taking direct control of AI expenditures.
The company also faces a technical reality: frontier models from OpenAI and Anthropic offer measurable productivity gains, but their token-based pricing complicates cost management. Bills are reviewed after the month ends, leading to reactive adjustments.
Borgman's team needed to balance three pressures:
- Maintaining AI-driven development speed
- Keeping token costs within acceptable budget limits
- Avoiding the overhead of constant manual optimization
They also had to deal with the end of promotional pricing. As Carmen Li, CEO of Silicon Data, pointed out, transitioning from subsidized to standard pricing often results in higher bills, not due to increased per-token prices, but because discounts end.
The Approach Taken
Starburst adopted a two-track strategy.
First, they implemented reactive cost controls. The company reviews monthly bills and makes adjustments to optimize consumption, such as identifying redundant tests or restructuring high-token workflows. While not predictive, this creates a feedback loop.
Second, they explored model alternatives. Borgman's team evaluated Nvidia's open weight Nemotron model, among others, due to Starburst's partnership with the chipmaker. Hosting open weight models on premises could significantly reduce token costs while maintaining data control.
Borgman also emphasized the need for Anthropic to lower costs faster to capture the market. Otherwise, frontier vendors must specialize in specific use cases where they add value beyond just the model itself.
This approach reflects a broader trend. Silicon Data's benchmarking index rose from March to June as enterprises used more premium frontier models. Since June, the index declined as enterprises optimized workloads and shifted tasks to smaller or open weight models.
Results and Metrics
Starburst achieved significant development efficiency gains: AI tools now write about a third of the company's code, reducing time from idea to production by up to 60%. These metrics justify the AI investment from a capability standpoint.
On the cost side, the company stabilized spending growth through monthly reviews, though predictive modeling remains elusive. Token costs continue to rise, but at a manageable rate through reactive optimization.
The investigation into open weight alternatives revealed potential savings of 5x to 10x for routine coding tasks, according to Max Christoff, CTO at Everlaw. While quality might be slightly lower, this is less critical for tasks not requiring frontier-level capability.
Lessons Learned
Borgman identified the core gap: the need for better internal tools to predict high token consumption before work begins. Starburst requires predictive cost modeling, not just post-hoc analysis.
The company would also push for vendor pricing transparency earlier. As Shrihari Sridhar from Texas A&M University noted, "Bills shouldn't write themselves." Borgman's team would have benefited from clearer reasoning token visibility from the start.
Another adjustment: earlier exploration of hybrid model strategies. Instead of defaulting to frontier models for all AI development, Starburst could have segmented workloads by task complexity from the outset, routing routine work to lower-cost alternatives while reserving frontier models for more complex tasks.
Takeaways for Your Team
Build cost visibility before scaling usage. Starburst's reactive approach is inefficient. Before expanding AI tool deployment, track token consumption by task type, team, and model. You need leading indicators, not trailing bills.
Segment workloads by model capability requirements. Not every task needs frontier-level reasoning. Identify workflows that can run on smaller or open weight models. The economics are compelling: potential 5x-10x cost reduction for suitable use cases. Host open weight models in your environment if data security requires it.
Prepare for CFO scrutiny. As Everlaw's Christoff showed, you can tie AI spending to clear ROI, $10,000 in tokens replacing a year of engineering time for a $1 million revenue feature. Show the math before finance leadership demands it. Track both productivity gains and token costs by project.
Demand pricing transparency from vendors. Reasoning tokens shouldn't be a black box. Your procurement terms should require vendors to provide consumption forecasting tools or detailed usage breakdowns to build your own models. If a vendor can't help you predict costs, they're asking you to sign a blank check.
Consider consumption caps as a governance control. Quadient's CIO Nina Tatsiy stays on Anthropic's token-capped Team plan instead of moving to unlimited enterprise tiers. Teams learn efficiency quickly when they hit limits. Under chargeback models, departments must justify AI consumption against either cost savings or new revenue.
The token pricing model isn't going away, but you don't have to accept opaque bills. Build the instrumentation, segment the workloads, and hold vendors accountable for cost predictability. Your CFO will thank you.
