AI cracks JEE Advanced 2026 with ease

JEE Advanced is the filter after JEE Main. A few lakh students sit Main. A few thousand walk into an IIT. The 2026 paper still looks like the exam families know: long diagrams, multi correct lists, numerical ranges, and matching columns that punish a single sloppy option.

Goodmarks split the official English papers into 102 single question PDFs and sent each one to four models. The prompt carried the question images and the question type. It never carried the official key. Scoring happened after the model replied, using the IIT solutions booklets and the paper’s own marking rules.

The result is blunt. The best models are no longer grazing the cutoff. They are finishing within a handful of marks of a perfect script.

What “with ease” actually means

A human who scores in the mid 300s on JEE Advanced is not having a normal day. That band is the air that All India Rank 1 lives in most years. Claude Fable 5.1 is 2 marks short of a clean sweep. GPT 6 Astra is 4 marks short. GPT 5.5 is 9 marks short and still at 97.5 percent.

Paper 1 was almost a formality. 3 of the 4 models scored a perfect 180 out of 180 on Paper 1: Claude Fable 5.1, GPT 6 Astra, and GPT 5.5. Physics was a clean sheet for both OpenAI models. Chemistry was a clean sheet for Claude Fable 5.1 and GPT 6 Astra. After the question 4 rescore, both Claude models also sit on a full 120 in Mathematics.

Claude Sonnet 5 is the only model in this set that still looks mortal. It scored 341, lost most of those marks in Chemistry, and returned two questions as errors. Even then, 94.7 percent on the official 2026 paper is a score most coaching classrooms would print on a banner.

A scoring fix on question 4

Paper 2 Mathematics question 4 is a single correct definite integral worth 3 marks. The official letter is B, which is 1/3. All four models returned letter A. That looked like a shared miss until we checked the image we had sent them.

Our split of the official PDF hid the denominators under A and B behind a white box. The models saw 1 over a blank for both plain fractions. C and D (the log forms) stayed readable. Mapping a correct 1/3 onto a letter was then a guess. That is our fault, not a calculus miss.

So we rescored the item on the written value. If a model wrote 1/3, we restored the full plus 3. That is a 4 mark swing from the raw minus 1. If a model left only the letter A and no value, the raw mark stands.

The printed question, now with the full option table, is:

The integral itself

Put u = 3 to the x. Then du = (3 to the x)(ln 3) dx, so dx = du / (u ln 3). The limits change from x = 0 to x = 2 into u = 1 to u = 9. The integral becomes

Split the fraction: 1 over u(u+3) equals (1/3) times (1/u minus 1/(u+3)). The antiderivative is (1/(3 ln 3)) ln(u/(u+3)). From 1 to 9 that is

The value is 1/3. On a complete option table that is letter B.

Who wrote 1/3

Claude Fable 5.1 rewrote the integrand, reduced the log term, and arrived at 1/3. It even noted that the printed choices were not clear, then attached that value to letter A. Claude Sonnet 5 used the swap that replaces x with 2 minus x, showed that f(x) plus f(2 minus x) equals 1/3, and also arrived at 1/3 before locking A. Both get plus 3.

GPT 6 Astra and GPT 5.5 left no written work on this item. They returned only letter A. We cannot show that they held the value 1/3, so we did not add marks. GPT 6 Astra’s only remaining miss is this letter. GPT 5.5 still has this letter plus a wrong multi correct on Paper 2 Chemistry question 9.

Claude Fable 5.1 now has no fully wrong item: 101 correct and one partial on Paper 2 Physics question 5 (A and D instead of A, C, and D). Claude Sonnet 5 still drops the rest of its marks in Chemistry, including two questions that returned no parseable answer.

Why this is not memorisation

GPT 5.5 shipped on April 23, 2026. JEE Advanced 2026 was held on May 17, 2026, 24 days later. The 2026 paper did not exist when that model went public. The other three models sat the same printed pages with the same blind prompt.

Each request was one question. Diagrams stayed as printed. Numerical answers counted only if they landed inside the official range. Multi correct items used the paper’s partial marking, so a missing option is not silently forgiven.

What this means if you are sitting the paper

Models can now finish JEE Advanced. That does not sit the exam for you. Partial marking still bites. Chemistry still separates a 358 from a 341. A clipped option table can also turn a correct value into the wrong letter, which is why we publish the image defect and the rescore.

If you are preparing, the useful lesson is narrower. The paper is readable. The marks live in careful options, not in a secret chapter no one taught. Practice the same shapes the models saw: multi correct lists, numerical ranges, and matching columns, with instant checking and a written solution after every attempt.

The full GPT 5.5 question log still shows the raw letter A on this item, because that run left no written 1/3. See the GPT 5.5 JEE Advanced 2026 report.

For more details check the JEE Benchmarks page. It has the full 2025 and 2026 leaderboard, paper by paper marks, and every miss.