OpenAI's AI Scorecard

CFOs are asking a fundamental question: how to maximize value from AI investments. Traditional software metrics like user adoption no longer suffice for artificial intelligence. A new approach is needed, one that measures the actual work AI accomplishes. This is the core of OpenAI's proposed AI work accomplished scorecard.

The ultimate metric, dubbed 'Useful Intelligence per Dollar', answers four critical questions: Is AI performing valuable work? What is the true cost of a successful AI task? Can AI outputs be relied upon? Does AI's value increase with usage?

1. How much useful work gets done?

This first pillar focuses on tangible outcomes. Did AI help resolve customer issues, ship code, or review contracts? The value of AI lies not in tokens, but in transforming those tokens into actionable work. For instance, AI can automate preparatory tasks for finance teams, freeing them for higher-level analysis.

2. What does a successful task actually cost?

Calculating the cost of an AI task requires looking beyond per-token pricing. It includes compute, employee time, human review, and retries. A cheaper model might incur higher total costs if it requires more iterations or corrections. OpenAI's OpenAI GPT-5.6 family, with tiers like Sol, Terra, and Luna, aims to offer optimized cost-performance options.

Frontier models can provide better value by delivering accurate results in a single pass, reducing overall expenses.

3. How often does AI get the work right?

Dependability is crucial for deeper AI integration. As AI moves from drafting to taking action, its accuracy and consistency become paramount. Tracking outcomes like 'ready to use', 'needs correction', or 'needs escalation' provides a clearer picture than raw model accuracy.

Defining clear boundaries for AI access and actions is essential for safety, security, and trust.

4. Does each AI dollar buy more work as usage grows?

The final measure examines AI's economic scalability. As usage increases, the cost per successful task should ideally decrease, or the value generated should outpace costs. This efficiency is driven by improvements in compute, algorithms, and model architecture.

This virtuous cycle, where infrastructure improvements fuel research, leading to better models and products, ultimately benefits customers through enhanced capabilities and reduced costs. This is the essence of Useful Intelligence per Dollar in practice.