What Was the GPT-4 SAT Score? Plus the GRE and LSAT Percentiles
GPT-4 scored 710 and 700 on the two SAT sections, 163 on the LSAT and 169 on GRE Verbal. OpenAI's report ranks each among test-takers, and those ranks are not IQ points.

What you need to know
- The GPT-4 SAT score is 1410 by our sum of the two sections in OpenAI's 2023 report: 710 in Evidence-Based Reading and Writing (about the 93rd percentile) and 700 in Math (about the 89th). The report gives no percentile for the total.
- GPT-4 scored 169 on GRE Verbal (about the 99th percentile), 163 on GRE Quantitative (about the 80th, 157 and the 62nd without image input), 4 out of 6 on the GRE essay (about the 54th) and 163 on the LSAT (about the 88th).
- Each percentile ranks GPT-4 among people who chose to take that exam. Two of the five percentile sources the appendix cites are test-prep blog posts, and none is about the LSAT.
- On our IQ scale those six percentiles translate to about 101.5 up to 135, a 33-point spread. That is a translation of ranks, not an IQ: GPT-4 has none.
The GPT-4 SAT score is 1410: 710 on Evidence-Based Reading and Writing and 700 on Math, on the 400 to 1600 scale. GPT-4 also scored 163 on the LSAT and 169 on GRE Verbal. These are the figures in OpenAI's 2023 technical report, and each comes with a percentile among people who took that exam. A percentile among test-takers is not an IQ.
Below: the numbers exam by exam, what the percentiles measure, what they look like on the IQ scale (our arithmetic), and where GPT-4 did worst.
What is the GPT-4 SAT score?
GPT-4 scored 710 out of 800 in Evidence-Based Reading and Writing, about the 93rd percentile, and 700 out of 800 in Math, about the 89th. The report gives a percentile for each section, not for the total. We added the two sections to get 1410, and no percentile for that total is printed.
OpenAI also ran a version with no image input. Its reading and writing score stayed at 710, and its math score was 690, still about the 89th percentile. The reading and writing section has no images, so the two runs match there.
What was the ChatGPT SAT score?
The ChatGPT SAT score is not in the report as such, because it does not test the ChatGPT app as a product. Its comparison model, GPT-3.5, was run on a standard ChatGPT snapshot, and it scored 670 in reading and writing (about the 87th percentile) and 590 in math (about the 70th). That is 1260 in total by our sum, so GPT-4 gained 150 points.
What is the GPT-4 GRE score?
The GPT-4 GRE score is 169 out of 170 on Verbal, about the 99th percentile, and 163 out of 170 on Quantitative, about the 80th. On the essay it scored 4 out of 6, about the 54th percentile. Verbal and Quantitative are scored from 130 to 170.
Without image input the Quantitative score was 157 (about the 62nd percentile) and the Verbal score was 165 (about the 96th). The report does not explain each change. The GRE essay was graded by one or two outside contractors, using prompts from a prep book.
What is the GPT-4 LSAT score?
The GPT-4 LSAT score is 163, about the 88th percentile, and 161 without image input, about the 83rd. GPT-3.5 scored 149, about the 40th. The LSAT is scored from 120 to 180.
The report's check for questions that also appear in its training data flagged 39% of the LSAT questions. Extrapolated from the questions that were not flagged, GPT-4's score would be 169, about the 97th percentile. The report publishes the lower of the two, 163, in Table 1.
That 88th percentile also looks high against LSAC's current table. As quoted in our LSAT percentile guide, a 160 is above 71.06% of scores and a 165 above 85.17%. A 163 falls between those two. The report names no source for its LSAT percentile, so we cannot say which table it used.
Percentile of whom? What the exam percentiles mean
Each percentile ranks GPT-4's score among people who took that exam. People choose to take the SAT, the GRE or the LSAT, and each exam draws a different crowd. A percentile among LSAT takers and one among SAT takers are ranks in two groups, so they cannot be averaged into one number.
The sources are a second limit. For the SAT, GRE and other exams the appendix cites five sources for percentiles. Two of the five are blog posts on test-prep sites, one about SAT percentiles and one about GRE percentiles. None of the five is about the LSAT or the bar exam. Even the SAT scaled scores came from a chart republished on a test-prep site.
One more rule shapes the table. Where a share of an exam's questions had been seen in training, OpenAI ran a second version without them and reported the lower score. That is cautious, but it is the developer's own check. We read the report and did not rerun any exam.
How were the exams run?
OpenAI used official past exams or practice exams from 2022-23 study material it bought, according to the appendix. For each multiple-choice section it tuned its method on one exam, then ran the held-out exam once for the final score. The essays were run once, with no tuning. The multiple-choice parts were run on a snapshot of the model from 1 March 2023, and the essays on a non-final snapshot from 23 February 2023.
So the table is one run per exam, by the developer, on a model that was still changing. Each multiple-choice prompt also gave GPT-4 a few worked examples with gold-standard answers. Those are fair ways to test a model, but they are one more reason to treat a percentile as a rough guide.
What would those percentiles be on the IQ scale?
A percentile of 99 matches an IQ of about 135 on a scale with a mean of 100 and a standard deviation (SD, the typical spread of scores) of 15. A percentile of 54 matches about 101.5. Here are all six, translated on that scale. You can check any pair with our IQ percentile calculator.
- GRE Verbal, about the 99th percentile: IQ-scale equivalent about 135
- SAT reading and writing, about the 93rd: about 122
- SAT Math, about the 89th: about 118
- LSAT, about the 88th: about 118
- GRE Quantitative, about the 80th: about 113
- GRE essay, about the 54th: about 101.5
Read these as a translation of a rank, not as a score. They would equal an IQ only if the exam takers were a random sample of adults, and they are not. An IQ also needs a test built and normed to measure general ability, which these exams are not. GPT-4 has no IQ.
On our scale GPT-4's six exam percentiles run from about 101.5 to 135, a 33-point spread, so no single number describes it.
For human scores, see how SAT and GRE results compare with IQ in our guides to SAT scores and IQ and GRE and LSAT scores and IQ. They show why those conversions are rough.
Where did GPT-4 do worst, and has anything changed?
Within these three exams, the GRE essay was lowest at about the 54th percentile. Across all of Table 1, the weakest rows were the AMC 10 maths contest at 30 out of 150 (6th to 12th percentile, extrapolated) and Codeforces programming at a rating of 392 (below the 5th). Next came AP Environmental Science at 2 out of 5 (8th to 22nd) and AP English Literature at 2 (14th to 44th).
The table also lists the bar exam, where GPT-4's 298 out of 400 was marked as about the 90th percentile. Other work disputes that figure, which our GPT-4 bar exam article explains, so we keep it to one row here.
Has it changed? OpenAI published this report in March 2023, and we read the 4 March 2024 version of the same paper. Newer models have been released since. We did not check their exam results for this article, so every number here is GPT-4's. For the wider picture, see why AI benchmark scores disagree, what MMLU measures and how smart AI is.
Where would your own score land?
These exams rank test-takers. To see a rank on the standard IQ scale, take the IQ Metrics test.
Find your IQ score now! →To place a human score among adults, take the IQ Metrics test or read our guide on whether AI is smarter than humans.
Common questions
What was the GPT-4 SAT score?
710 out of 800 in Evidence-Based Reading and Writing (about the 93rd percentile) and 700 out of 800 in Math (about the 89th), so 1410 on the 400 to 1600 scale by our sum. The report gives a percentile per section, not for the total. GPT-3.5 scored 670 and 590.
What did GPT-4 score on the GRE and the LSAT?
GRE Verbal 169 out of 170 (about the 99th percentile), GRE Quantitative 163 (about the 80th) and the GRE essay 4 out of 6 (about the 54th). On the LSAT it scored 163, about the 88th percentile, or 161 and the 83rd without image input.
Does GPT-4 have an IQ?
No. The percentiles in OpenAI's report rank GPT-4 among people who took the SAT, GRE or LSAT, not among all adults. Translated to the IQ scale they run from about 101.5 to 135, so no single IQ fits them.
Is the ChatGPT SAT score the same as GPT-4's?
No. The report tested GPT-4 and GPT-3.5, and ran GPT-3.5 on a standard ChatGPT snapshot. That model scored 670 and 590 on the two SAT sections, 1260 in total by our sum, against GPT-4's 1410.
Sources for this story
- OpenAI. GPT-4 Technical Report (2023), arXiv 2303.08774, version 6 of 4 March 2024: Table 1, Appendix A and Table 9; read 16 November 2026 — OpenAI (arXiv)
- LSAC percentile table for the 2023-24, 2024-25 and 2025-26 testing years, as quoted in our LSAT percentile guide — LSAC
- IQ Metrics, Was the GPT-4 Bar Exam Score Really in the 90th Percentile? (published 14 November 2026) — IQ Metrics
- IQ Metrics computation, 16 November 2026: 710 + 700 = 1410, 670 + 590 = 1260, and percentile-to-IQ translation on a normal curve with mean 100 and SD 15 — IQ Metrics
Corrections: spotted an error? Email corrections@iqmetrics.org and we will update this story and note the change here.
Related stories
All news →
Was the GPT-4 Bar Exam Score Really in the 90th Percentile?
OpenAI reported that GPT-4 scored 298 of 400 on a simulated bar exam, about the 90th percentile. A 2024 re-analysis put the same score near the 62nd percentile of first-time takers and the 45th among licensed lawyers. The score held; the comparison group changed.

How Smart Is AI? Five Ways It Is Measured, and What Each One Says
Five yardsticks give five answers: up to 17.4 hours of expert work, 96.3% on graduate science questions, under 1% on new games in March 2026 and 73% in a chat test. None is an IQ.

What Is an Artificial Intelligence Test? Three Meanings, Compared
An artificial intelligence test can be a Turing test, an AI benchmark or a skills test for people. Here is who is tested, what comes out and what, if anything, links each one to IQ.
Read the research.
Then find your own number.
Our IIF-certified assessment reports your score with its scale, percentile and confidence range — and a breakdown of the cognitive domains behind it.
Start IQ Test →
