Nate: why I keep opening GPT-5.6 Sol even though Fable 5 is the smartest model

By Nate from Nate's Substack · July 13, 2026.

Nate, who benchmarks AI models, admits something that feels uncomfortable given that role: last night he opened ChatGPT 5.6 Sol even though he considers Fable 5 to be the smarter model. According to his own benchmark suite, Fable ended up ahead of Sol in the overall average, proving superior at reading between the lines, making sense of incomplete intentions, and finding what the user meant before they managed to express it clearly.

Even so, Sol is the model he opens when he sits down to do real work. Nate explains that this has to do with his way of working: he talks through problems in long and sometimes quite technical prompts, and usually already knows many of the nuances he wants to cover, even if it takes him a while to get them all out. Once the first result arrives, he corrects the work, inspecting what changed and trying to carry that lesson into the next run.

Sol fits that pattern: it takes the full instruction seriously, persists at the task, and works well within the skills and tools Nate has built around it. Fable, by contrast, helps him more when the important part of the work is still halfway between his intention and his words.

For Nate, both capabilities are real and valid; which one is needed depends on the person, not just the task. He believes this divergence is the future: the better these models get, the further apart the reasons for choosing one or the other grow, and people who work differently can look at the same lineup of models and make different decisions for equally good reasons.

The previous week he had defended the idea of 'starting with the task' when choosing a model. Now he qualifies that this is only half the equation, because it leaves out the person doing the work—something that, according to him, most of those who recommend how to choose a model overlook.

With that idea as the driver, Nate introduces 'Model Fit,' a tool/guide of his own along with the benchmark suite behind it, suggesting that each reader's answer will probably not match his own. The email indicates that paid subscribers have access to the full analysis, the guide, and membership in his Slack community, though the body of the message does not detail further contents of that paid portion.