On August 1, OpenAI announced that an internal version of its Astra model solved ten open problems across mathematics and theoretical computer science. These included a construction proving the existence of non-sofic groups, a longstanding question in group theory. The model published formal proofs on GitHub. Fields Medal winner Timothy Gowers said he would recommend one of them for publication in a top journal without hesitation. The reported compute cost was roughly two thousand dollars.
That is a result on a particular research task. For a business considering AI, the next step is to identify a task it needs done and test the available system against it.
What the business surveys measure
BCG’s 2026 analysis of AI-first cost advantage reports that sixty percent of companies get minimal or no value from their AI investments. It also attributes 10% of the value in a typical implementation to the algorithm, 20% to technology and data, and the remaining 70% to managing process change.
Those proportions put a substantial part of the work with the people responsible for operations. Someone has to decide where an AI output enters a process, who checks it, and what happens when it fails. Buying access to a capable model leaves those decisions with the business.
BCG reports that AI leaders deliver three times greater cost reduction than peers, with 1.6 times higher EBIT margins and 2.7 times higher return on invested capital. The firms it calls “future-built” see five times the revenue increases and three times the cost reductions from AI compared to everyone else.
These comparisons describe groups of companies. They do not establish that a particular reorganization will produce the same return for another business, or that access to Astra would explain the difference.
Put an owner on the test
McKinsey identifies roughly 6% of enterprises as AI high performers, defined as those attributing more than 5% of EBIT directly to AI. That definition is useful because it asks for a business result. A team evaluating a new capability needs its own measure, even if the first test is much smaller: review time, completed work, corrections, or cost per accepted output.
Meta also narrowed its 2026 capital expenditure range upward to $130 billion to $145 billion, driven by AI infrastructure buildout. That spending describes investment in capacity; a company using AI still needs to establish the value of its own deployment.
If your company deployed AI tools in the last year but cannot point to measurable business outcomes, choose a workflow with a known baseline. Give one person responsibility for the trial, include the time spent checking and correcting the output, and compare the result with the existing process. That gives the next investment decision something concrete to rest on.