On July 17, 2026 OpenAI proposed measuring AI value not by tokens or seats but by the work completed per dollar spent. The company framed the approach as a scorecard called “Useful Intelligence per Dollar,” composed of four questions about output, cost, dependability and scale.
Measuring useful work
OpenAI urged organisations to start from concrete workflows and define what “done” means in the system where work happens. Examples given include resolved customer issues, code changes that pass tests, and contracts reviewed on time. The company argued that models create value when tokens translate into usable outcomes and noted that more capable models can take on longer, multi-step tasks.
Cost, dependability and model tiers
The firm said the full cost of a successful task must include model price, compute, employee time, human review, retries and rework. OpenAI warned that lowest price per token can still yield higher cost per outcome if a model requires more attempts.
As an example of its tiered approach, OpenAI described GPT‑5.6 with three variants: Sol (flagship), Terra (balanced) and Luna (fastest and most affordable). The company reported that GPT‑5.6 Sol set a new state of the art on the Artificial Analysis Coding Agent Index, used 54% fewer output tokens than another leading model in a cited comparison, and reached 72.7% on DeepSWE v1.1 versus Claude Fable 5’s 69.9% at a 36.2% lower estimated API cost.
Dependability was presented as a separate measure: results that are “ready to use,” those that “need correction,” and those that “need escalation.” OpenAI recommended clear boundaries for data access, system control and human review before granting models authority to take actions.
Value at scale and infrastructure
OpenAI linked improvements in useful work per dollar to advances in both training and inference compute, plus infrastructure gains such as better routing and purpose-built hardware. The company said those elements combine across products—ChatGPT, ChatGPT Work, Codex and the API—to translate research and compute into customer outcomes.
OpenAI provided specific claims about GPT‑5.6’s performance and its three tiers (Sol, Terra, Luna), and noted integration with ChatGPT Work and ChatGPT Enterprise as part of the proposed scorecard framework.
Original source: OpenAI News