PhD-level science questions: best AI rose from 42% to 96%; PhD experts score 65–70%
On 198 hard multiple-choice science questions, independent tests put the best AI at 96.4% in 2026. PhD experts get 69.7%; guessing 25%.
PhD-level science questions (GPQA Diamond)
Best independently measured score, by model release date. Multiple choice, 4 options.
What this measures: 198 hard biology, chemistry and physics questions written by PhDs. Experts in the field get about 70% right; smart non-experts with Google get 34%; guessing gets 25%.
- Best AI score (Epoch AI runs)
- Reference level
Saturated: Top scores (about 96%) are near the test's ceiling, where errors in the answer key cap what's achievable, so differences between today's best models are mostly noise. Now watch: Humanity's Last Exam.
Independent runs by Epoch AI, plotted at each model's release date (many runs happened later). The expert baseline comes from a study OpenAI ran with recruited PhDs; the original paper reports 65% on the full question set. Questions have been public since 2023, so training-data contamination can't be ruled out, and scores above ~95% may be near the benchmark's error floor.
Show data table and sources
| Date | Label | Value |
|---|---|---|
| 14 Mar 2023 | GPT-4 (0314) | 36% |
| 6 Nov 2023 | GPT-4 Turbo | 42% |
| 29 Feb 2024 | Claude 3 Opus | 47% |
| 13 May 2024 | GPT-4o | 49% |
| 20 Jun 2024 | Claude 3.5 Sonnet | 54% |
| 12 Sept 2024 | o1-mini (high) | 62% |
| 17 Dec 2024 | o1 (high) | 77% |
| 31 Jan 2025 | o3-mini (high) | 77% |
| 24 Feb 2025 | Claude 3.7 Sonnet (thinking) | 79% |
| 25 Mar 2025 | Gemini 2.5 Pro | 84% |
| 5 Jun 2025 | Gemini 2.5 Pro (06-05) | 85% |
| 9 Jul 2025 | Grok 4 | 87% |
| 13 Nov 2025 | GPT-5.1 (high) | 88% |
| 18 Nov 2025 | Gemini 3 Pro | 93% |
| 19 Feb 2026 | Gemini 3.1 Pro (high) | 94% |
| 5 Mar 2026 | GPT-5.4 Pro | 95% |
| 13 Aug 2026 | Gemini 3.7 Flash (high) | 95% |
| 2 Sept 2026 | Gemini 3.8 Flash (high) | 95% |
| 3 Sept 2026 | GPT-6 Astra (max) | 96% |
| 30 Sept 2026 | Gemini 4 Argon (high) | 96% |
| PhD experts in the field (69.7%) | 70% | |
| Skilled non-experts with web access (34%) | 34% |
- GPQA Diamond (Epoch AI Benchmarking Hub) (Epoch AI)
- Epoch AI benchmark runs (CSV, CC BY 4.0) (Epoch AI)
- GPQA: A Graduate-Level Google-Proof Q&A Benchmark (Rein et al.)
GPQA, released in November 2023, is a set of PhD-level science questions written to be hard to look up. The paper calls them “Google-proof”. Each has four answer options, so guessing scores 25%. In the original paper, PhD experts in the relevant field got 65% right, skilled non-experts with unrestricted web access got 34%, and the best GPT-4 setup got 39%.
The standard version today is GPQA Diamond, a 198-question subset. Epoch AI, an independent research group that runs the tests itself, compares models with a 69.7% baseline from PhD experts recruited by OpenAI.
In Epoch AI’s runs, by model release date, the best score rose from 42.4% (GPT-4 Turbo, November 2023) to 62.4% (o1-mini, September 2024), 87.0% (Grok 4, July 2025), 92.6% (Gemini 3 Pro, November 2025) and 96.4% (Gemini 4 Argon, September 2026). Labs’ own scores follow the same path: Anthropic listed Claude Mythos Preview at 94.6% in April 2026.
What it means
On written multiple-choice science questions, the best AI now beats PhD experts in their own fields; in 2023 it scored well below them. That is not the same as doing science: choosing among four answers is far from designing and running research. The test is also close to its ceiling, so it no longer separates the top models well.
Caveats
- It’s multiple choice, so it measures picking the right answer, not open-ended research. OpenAI itself said beating the expert baseline doesn’t make a model more capable than a PhD in every respect.
- The 69.7% expert figure comes from a study run by OpenAI, reported via Epoch. The paper’s 65% applies to the full question set, not only Diamond.
- The questions have been public since 2023, so models may have seen them in training.
- Scores above about 95% may be close to the benchmark’s own error rate; no verified count of mislabelled Diamond questions was found.
- Epoch plots scores by model release date; many runs happened later (for example, Gemini 3.1 Pro was run in August 2026).