The September 2025 SAT: the exam nobody could describe the same way

Genuinely unclear, and unusually so. September produced the mildest difficulty reaction of any recent sitting: a fair number of students described the math as straightforward apart from two or three items, while others found the same section a struggle. Neither group was wrong.

What separated them was not ability but which form they were given. Students who worked through what they remembered found they were describing different questions in the same slots, and in at least one case two different words as the answer to what sounded like the same vocabulary question. That is not faulty memory, and it is not two difficulty levels of one module. It is a much deeper pool of questions than most students assume exists.

The other constant was quieter. Almost nobody reported running out of time. September was a test of whether you recognised what was being asked, not whether you could work fast.

Students know the SAT is adaptive, and most can tell you that the second module comes in an easier and a harder version. What September made visible is the layer underneath that: each of those versions exists in several forms, and which one you get is not information about you at all.

The consequence is worth stating plainly, because it runs against the strongest instinct students have after a test. Comparing questions with other people cannot tell you how you did. If someone describes a question you never saw, the possibilities are that you were routed differently, or that you sat a different form of the same module, and neither is a verdict on your performance. The comparison feels like evidence because it produces a specific, checkable fact. It is not evidence about anything you care about.

There is a second, more practical consequence. Any advice that promises the exam contains a fixed number of a particular question type is describing one form, not the test. Build a plan around skills, not around counts.

Statistics and data analysis came up more than any other math topic in September, and the items students found difficult had almost no calculation in them. The recurring shape was a study description followed by a question about what the result licenses you to conclude: which group the finding applies to, whether a margin of error is meaningful, whether a sample supports a claim about a wider population.

These reward reading rather than computing. The trap in nearly all of them is scope. A result drawn from a random sample of one school’s students supports a conclusion about that school’s students, and about nothing else. Choices that widen the population, or narrow it to only the people sampled, are wrong for the same structural reason, and they appear in almost every version of this question.

This is the highest-yield thing to learn in the whole Problem Solving and Data Analysis domain, because the reasoning transfers unchanged from one form to the next while the numbers do not.

Beyond the sampling questions, September’s math ran through familiar ground: exponential models, a percent relationship between two quantities, a graph feature to identify, a linear model built from a description. The hard items clustered in advanced math rather than geometry, which is unusual for a recent sitting and worth noting if you are comparing dates.

What made the difficult ones difficult was translation. Several asked for a relationship rather than a value, which is a shape that has become steadily more common and that a calculator cannot help you with until you have already done the thinking.

Paired passages featured heavily, and they are the item type where the gap between a strong reader and a trained one is widest. The question is never what each text says. It is what relationship holds between them: whether the second text qualifies the first, contradicts a specific claim in it, or answers a question the first left open.

The vocabulary students reported was mostly about fine distinctions between near-synonyms, including one pair close enough in meaning that students were still arguing about which the test wanted. When two options genuinely both fit the general sense, the question is testing which one the sentence structure requires, not which one you would use in conversation.

Original questions written around what this exam tested, one for each argument above. Nothing here is a real test question.

The lesson of September is about how to read every other test-date account, including this one. When students disagree about whether an exam was hard, the most likely explanation is that they met different questions, and the disagreement tells you nothing about the exam. What does transfer between forms is the set of skills and the shapes those skills arrive in.

That is why we write about question types rather than question lists. The list changes every time. The types have barely moved in two years.