Chief Financial Officers are no longer satisfied with the vague promise of productivity gains. For the past few years, the software industry has measured success through adoption metrics: how many seats were purchased, how many users are active, and whether the license was renewed. But as generative AI moves from the experimental phase to the core of the enterprise stack, the conversation in the boardroom has shifted. The question is no longer whether a team is using AI, but exactly how much business value is being extracted from every dollar spent on compute. This tension has created a gap between the technical cost of running a model and the actual economic value of the output.
The Three Tiers of Useful Intelligence
OpenAI is addressing this valuation gap with the release of the GPT-5.6 model family and a new guiding metric called Useful Intelligence per Dollar. This framework moves the goalpost from simple token consumption to the quantification of completed tasks. Instead of tracking how many tokens a model generates, OpenAI suggests that enterprises track the volume of work accomplished, such as the number of customer issues fully resolved, the amount of code successfully deployed to production, or the number of contracts reviewed without requiring human correction. As models evolve to handle longer contexts and multi-step reasoning, the value is found at the precise moment a token is converted into a usable business result.
To implement this economic strategy, GPT-5.6 is deployed across three distinct tiers designed to match specific workload requirements. The flagship model, Sol, is engineered for maximum reasoning capability and is positioned for high-complexity problem solving where accuracy is non-negotiable. Terra serves as the balanced mid-tier, optimizing the trade-off between performance and cost for general analytical tasks. Luna is the high-velocity, low-cost option, designed for massive volumes of simple, repetitive workflows. By distributing tasks across these three tiers, organizations can align their compute spend with the actual complexity of the work, ensuring that expensive reasoning is reserved for tasks where a single, correct answer is more valuable than a series of cheap, iterative attempts.
The Hidden Math of Total Workflow Cost
The critical insight provided by the GPT-5.6 release is that token price is a deceptive metric. When a company calculates the cost of an AI workflow, the formula is not simply token price multiplied by usage. The true cost is the sum of the API spend plus the time spent on human review, the cost of multiple retries, and the overhead of rework. This creates a paradox where the cheapest model can become the most expensive option. A low-cost model like Luna may have a negligible token price, but if it requires three retries and twenty minutes of human auditing to fix a hallucination, the total workflow cost skyrockets.
In contrast, a frontier model like GPT-5.6 Sol is designed to minimize this hidden overhead. By increasing the probability of a first-pass success, Sol reduces the need for human intervention and repeated API calls. The technical data supports this efficiency. In the Artificial Analysis Coding Agent Index, GPT-5.6 Sol achieved state-of-the-art performance while reducing output token usage by 54% compared to other leading models. By optimizing the path to the correct answer, the model generates less noise and reaches the solution faster.
This efficiency is further evidenced in long-term engineering tasks. On the DeepSWE v1.1 benchmark, which tests the ability to solve complex software engineering issues, GPT-5.6 Sol recorded a 72.7% success rate. This outperforms Claude Fable 5, which recorded a 69.9% success rate. More importantly, the estimated API cost for GPT-5.6 Sol was 36.2% lower than that of Claude Fable 5. The result is a double-win for the enterprise: higher intelligence leads to a higher success rate, which in turn lowers the total cost of ownership by eliminating the cycle of failure and correction.
To operationalize this, OpenAI encourages firms to define a strict state of Done. For a customer support team, Done is not a generated response, but a closed ticket. For an engineering team, Done is code that passes all automated tests. For a legal team, Done is a contract approved within the deadline. By measuring the success rate against these concrete definitions, companies can finally calculate a real ROI.
To support this enterprise-grade scaling, OpenAI has introduced three reliability metrics: the success rate, the escalation rate, and the rework rate. The escalation rate measures how often the AI fails and must hand the task back to a human, while the rework rate tracks how often a human must edit the AI's output. As these rates drop, the cost of completing a unit of work falls, regardless of the token price. This logic is integrated into ChatGPT Work, which builds upon the security and compliance foundations of ChatGPT Enterprise to allow deeper integration into internal corporate contexts and workflows. When the rate of completed tasks grows faster than the total cost of the AI pipeline, the investment reaches a point of positive ROI.
The era of paying for AI by the word is ending, replaced by a regime where the only metric that matters is the cost of a finished job.
