Cost optimization

PromptGlass gives finance and IT one view of AI spend across seat-based subscriptions (ChatGPT Enterprise, Microsoft 365 Copilot, Claude, GitHub Copilot, Gemini) and usage-based APIs (OpenAI, Anthropic, Azure OpenAI, Bedrock, Vertex). It then recommends specific actions: reclaim idle seats, remove duplicate subscriptions, and move workloads to cheaper models where the quality holds.

Where cost data comes from

Spend type	Source	Precision
API usage via gateway	Measured tokens × price list	Per request
API usage via connectors	Vendor usage and cost APIs	Per project / key / day
Seat subscriptions	License assignments from connectors + contract price you enter	Per seat / month
Internal / self-hosted models	Tokens × your internal rate (GPU cost)	Per request

Enter contract prices under Cost → Contracts so seat costs reflect your negotiated rates.

Recommendations

Idle seats

A seat is idle when its holder had no interactions in the window you choose (30, 60 or 90 days). PromptGlass lists idle seats by tool and team and the monthly amount you could reclaim. It can open a ServiceNow or Jira ticket for license removal, or notify the user first (“use it or lose it”).

Duplicate subscriptions

Finds people paying (or being paid for) more than one tool that covers the same job, such as ChatGPT Plus on an expense report plus a company Copilot seat, or Cursor plus GitHub Copilot. Expense-report data can be imported as CSV.

Model right-sizing

For gateway traffic, PromptGlass groups requests by use-case topic and compares them across models. Where a smaller model’s outcome signals (re-prompt rate, error rate, user feedback) are equivalent for a topic, it recommends a routing change and estimates the savings:

Topic: ticket-classification   Team: Support   Current: gpt-4o
Candidate: gpt-4.1-mini        Outcome delta: none detected (n=18,402)
Estimated monthly saving: shown in console, based on your current volume and price list
Apply: promptglass route suggest --topic ticket-classification --target gpt-4.1-mini

Prompt caching and waste

The gateway flags repeated identical prompts (candidates for caching), very large contexts with low reuse, and runaway agents (sudden token spikes from a single key).

Budgets and showback

Set budgets by team, cost center, app or virtual key. Showback reports allocate shared seat and API costs to cost centers monthly. Export them to your FinOps tool:

promptglass export cost --group-by cost_center,app,model --month 2026-10 --format csv

Alerts

alerts:
  - name: spend-spike
    when: { metric: cost.usd, scope: virtual_key, increase: "300%", window: 1h }
    notify: ["slack:#finops", "pagerduty:ai-platform"]