A developer reviews an AI coding workflow dashboard showing bot checks, performance metrics, and project progress.
The cheapest AI tokens didn’t buy the cheapest work. In a Databricks coding-agent benchmark, Google’s Gemini 3.8 Flash had a low token price but averaged 140 turns and $11.97 per task, putting it among the most expensive models tested. OpenAI’s GPT-6 Luna and Anthropic’s Claude Opus 5.5, meanwhile, both posted an 89.1% task success rate, but their average costs were $0.05 and $1.62 per task, respectively. That’s a roughly 32-fold difference on the same benchmark. The rate card tells you what a token costs. It doesn’t tell you how many attempts it will take to get usable work.

A developer reviews an AI coding workflow dashboard showing bot checks, performance metrics, and project progress. Compare the whole workflow, not just the model​

The Databricks AI Index combines anonymized usage data from more than 10,000 companies using Unity Gateway with evaluations based on Databricks engineers’ own coding workflows, including pull requests and issue investigations. The initial dataset covers July 20 through November 7, 2026. Hidden tests determine whether the coding tasks succeed, and cost estimates use publicly listed provider prices rather than each customer’s negotiated rates. The results describe specific models paired with the software that runs them, not what every buyer should expect to pay.

The usage data also shows how quickly buyers can move. About 64% of enterprises in the Unity Gateway sample relied on a single provider in July. By November 2, that share had fallen to 28%. That’s evidence of diversification among customers of a gateway built to connect to multiple providers, not a census of enterprise AI. And the benchmark gives those customers a reason to keep their options open: Claude Fable 5.1 scored 96.0% at $7.47 per task, while Muse Wonder scored 94.9% at $1.53. Whether that extra 1.1 percentage points is worth nearly five times the spend depends on what a failure costs.

A useful internal test is to replay a representative batch of work against each candidate model and agent configuration. Decide up front what counts as success—a patch that passes tests and review, for example—then divide the total cost of all runs, including failures and retries, by the number of accepted results. Track human review time alongside that figure. A model that looks cheap on the invoice isn’t necessarily cheaper if engineers spend the savings cleaning up its output. Databricks used workload replays itself and found that one proposed model upgrade wasn’t worth the added cost for its use cases.

Find out who—or what—is driving the bill​

Databricks reported median user spending of $2.49, versus $699.74 at the 99th percentile, a roughly 281-fold gap. The $2.49 figure is a median, not a daily average. Databricks also reported that median weekly spending more than doubled over six weeks while spending at the 90th and 99th percentiles stayed relatively flat. The gap is a reason to investigate high spend, not automatically throttle it: an agent doing useful work all day and one retrying the same failed step can both top a consumption report.

The broader budget pressure is real, even if individual bills vary. Gartner’s July forecast put worldwide end-user spending on AI models and platforms at about $64 billion in 2026, up 63.4% from 2025. That forecast covers models and platforms, not all AI spending. In Microsoft’s Inside Track guidance, corporate vice president Tal Saraf made the complementary point: knowing who uses the most tokens and which models they choose should help IT teams steer employees toward the right tools, not simply shut heavy users down.

What Microsoft administrators can measure now​

Microsoft Digital recommends tracking usage by application, workflow, agent and user, then using those measurements to set budgets, rate limits and fallback behavior. In a Microsoft environment, the tools answer different questions: Microsoft 365 Cost Management can show Copilot Credit consumption by user, group, service or agent and apply spending policies and alerts. Copilot Studio and Power Platform controls can monitor consumption by environment and agent and set monthly agent limits. For Azure workloads, Microsoft points to Azure Cost Management for spend and Foundry observability for near-real-time token and per-agent cost views.

A practical starting point is to pick one metered workflow, give it an owner, and agree with finance on what a usable result looks like. Review the most expensive runs before setting a blanket cap: narrow an overbroad prompt, stop repeated retries, or test a lower-cost model on routine work while keeping an escalation path for harder cases. Budget for a peak month as well as a typical one. Keep the token dashboard—it helps explain where the money went—but judge the workflow by what that money delivered.

 

References

  1. Tokens only tell part of the enterprise AI spend story - CIO Dive CIO Dive 2026-11-20T15:30:24+00:00
  2. Introducing the Databricks AI Index: The state of enterprise AI tokenomics | Databricks Blog databricks.com
  3. Gartner Forecasts Worldwide AI Platforms and Models Market to Grow 63% in 2026 gartner.com