Are you measuring AI investment by useful work per dollar? OpenAI's 5-point playbook

Hey everyone, it's Shiichan! Today I found a thought-provoking piece about how to think about AI investment!

OpenAI News openai.com

What was announced?

OpenAI's News published a piece on how enterprises should manage AI investment in the agentic era. OpenAI's core stance is to make AI more affordable, capable, and accessible, but this piece argues that, from an enterprise's perspective, you shouldn't just look at token price. Instead, focus on useful work per dollar: how many tasks got done, how much time was saved, how much better decisions became, and how workflows scale.

Why it matters

The backdrop here is that model costs keep falling fast. Token prices dropped 97% from GPT-4 to GPT-5.4, and GPT-5.6 uses 54% fewer output tokens while taking 57% less time per task.

With costs falling this fast, just picking whichever model has the cheapest token price can actually leave value on the table. That's exactly why the piece argues enterprises need to update how they think about investment.

What changes

The piece lays out 5 practical strategies for leaders managing AI investment:

• Visibility into usage and spend: see clearly who's using which product and model, how much capacity they're consuming, and what work that usage supports. The Admin Console tracks adoption and credit usage by user, product, and model, plus trends over time and emerging patterns.
• Outcome-based model efficiency: the cheapest model per token isn't always the lowest actual cost. A pricier model may need less trial-and-error and less review. Measure with evals that reflect real tasks, define a "good enough" quality bar upfront, and track cost per approved outcome.
• Set up governance before you scale: treat governance as the operational layer that determines what AI work can scale. Decide in advance what context ChatGPT can use, what tools it can access, what actions are allowed, and who approves high-stakes steps. This matters even more as capabilities like plugins, connectors, and Computer Use come into play.
• Fund workflows that compound: think in a portfolio: everyday productivity (broad access), function-specific workflows (repeatable improvements), and strategic investments (a few high-priority bets that lean on your company's specific context), and fund in phases: exploration, validation, then production.
• Match capacity to proven demand: once a workflow's value is proven, match your products, capacity, and support model to actual demand. ChatGPT Work has ready-made features for chat, coding, and agentic workflows, plus connectors, plugins, and Computer Use, while large strategic rollouts can get support from programs like OpenAI Frontier and Deployment Company.

Wrap-up

• OpenAI argues enterprises should measure AI investment in the agentic era by useful work per dollar, not just token price
• Token prices fell 97% from GPT-4 to GPT-5.4, and GPT-5.6 cut output tokens by 54% and task time by 57%
• The 5 strategies are: visibility, outcome-based efficiency, governance set up early, portfolio-based funding, and demand-matched capacity
• The piece recommends measuring impact through quality and reliability metrics and business value like time saved, faster cycle times, and protected revenue

This one is for managers and decision-makers responsible for measuring AI investment impact and governance at their company!