GPT-5.6 Sol shifts AI ROI from token price to cost per finished task

The useful number is not the token price

Enterprise AI spending is starting to be judged by completed work, not by the cheapest model line item. In a July 17, 2026 post, OpenAI framed the buyer question as whether AI spend produces more useful work than it costs. The company calls the target “Useful Intelligence per Dollar,” but the practical metric is simpler: what does one successful task actually cost after retries, review, latency, and human correction are included?

The sharpest numbers sit around GPT-5.6. OpenAI says GPT-5.6 Sol with max reasoning set a new high on the Artificial Analysis Coding Agent Index while using 54% fewer output tokens than another leading model. On DeepSWE v1.1, a benchmark for long-horizon engineering tasks, it says GPT-5.6 Sol reached 72.7%, ahead of Claude Fable 5 at 69.9%, with 36.2% lower estimated API cost.

That framing matters because AI adoption is moving from drafting and search into workflows where the output has to survive contact with production systems. For support teams, the unit is a resolved customer issue. For engineering teams, it is a code change that passes tests. For legal teams, it is a contract reviewed accurately and on time. Tokens only create business value when they become work that a person or system can use.

OpenAI positions the GPT-5.6 family as a tiered buying decision. Sol is the flagship model for stronger reasoning, Terra balances performance and cost, and Luna is the faster and more affordable option. The point is not that one model should win every task. It is that customers should evaluate the full economics of a workflow: a cheaper model can become expensive if it fails repeatedly, while a more capable model can lower the cost per successful outcome by finishing in fewer attempts.

This is partly product marketing, but it also marks a more concrete phase of the AI ROI debate. If vendors want enterprise budgets to grow, they will need to show successful tasks, correction rates, escalation rates, and review time, not just context windows and token prices. Watch whether independent evaluations and customer case studies reproduce OpenAI’s cost-per-outcome claims in messy internal workflows.