AI for Science and Mathematics 2026: From Benchmarks to Real Discoveries

Two numbers from the same research group sum up where AI in science stands at the end of 2026. On FrontierMath Tier 4, Epoch AI's set of research-level maths problems written to resist AI, the best model went from 0% in January 2025 to 100% in September 2026. On FrontierMath: Erdős, 68 genuinely open problems that must be resolved with a machine-checked proof, the best system scores 4.4%. Between those two numbers lies the real question for every R&D team: which parts of research AI already does better than people, and which parts it has only just started on.

Research-level maths benchmarks fell in under two years

FrontierMath was launched by Epoch AI in late 2024 with several hundred unpublished problems written by mathematicians, from advanced undergraduate level (Tiers 1–3) to problems that take a specialist days (Tier 4). The answers are checked automatically, so the scores are hard to game. When the benchmark appeared, the best model solved about 2% of it. In June 2026 Epoch re-released an error-corrected version ("v2") and re-ran the models, which gives one clean series for the chart below.

The pace is the story. GPT-5 (August 2025) solved 22% of Tier 4; GPT-5.2 Pro (December 2025) 46%; GPT-5.5 Pro (April 2026) 78%; Claude Fable 5 (June 2026) 90%; GPT-6 Astra (September 2026) 98%; and GPT-6.1 Sol solved every problem. On Tiers 1–3 the best score is 94%. For comparison, the original Tier 4 set peaked at 48% in May 2026, reached by Google DeepMind's AI co-mathematician system rather than a single model.

The same holds outside maths. On GPQA Diamond, the PhD-level biology, chemistry and physics quiz, the best score is 96.4% (Gemini 4 Argon), and eleven labs are above 90%. On Humanity's Last Exam, the best score in Epoch's data is 54.8% (GPT-6 Astra). These tests are effectively solved or close to it, which is why our benchmark comparison now treats them as an entry ticket rather than a differentiator.

The labs: OpenAI and Anthropic lead, Mistral is a generation behind

• OpenAI leads FrontierMath (93.7% on Tiers 1–3, 100% on Tier 4) and HLE. It also produced the most visible results on open problems (below).
• Anthropic's Claude Opus 5.5 is close behind (91.2% and 95.0%), with Claude Fable 5 briefly holding the Tier 4 record in June 2026. Anthropic's science push is product-led: Claude for Life Sciences (October 2025), the standalone Claude Science product (June 2026) and a 10,000-seat programme for researchers.
• Google DeepMind leads GPQA Diamond (96.4%) and the CritPt physics benchmark (36.0%), and its specialised AI co-mathematician held the record on the original Tier 4.
• Mistral AI's Mistral Large 4 (October 2026) is competitive on GPQA (90.2%) but scores 58.6% on Tiers 1–3, 26.8% on Tier 4 and 10.6% on CritPt. That is roughly where OpenAI was in the second half of 2025: a strong option where data must stay in the EU, not the tool for frontier research maths.

The staircase: tests vs. open research

The chart shows why "AI solves research maths" and "AI has not solved research maths" are both true. Problems with a known answer that a test author can check are now largely within reach. CritPt, built from unpublished physics research problems, tops out at 36%. And on FrontierMath: Erdős, where Thomas Bloom (who runs erdosproblems.com) picked 68 open problems and each system had a $300 budget per problem to produce a Lean proof, GPT-6.1 Sol resolved 3 (4.4%), GPT-6 Astra, Claude Opus 5.5 and Claude Sonnet 5.5 2 each (2.9%), Gemini 4 Argon and GPT-6 Sol 1, and Claude Fable 5.1, Claude Opus 5 and GPT-5.6 Sol none.

Real discoveries: what actually happened in 2026

Mathematicians add two caveats. Erdős problems vary enormously in difficulty, and many AI wins are problems that standard techniques could crack once someone looked closely; several results are not yet peer-reviewed. And the proofs are piling up faster than people can read them. Still, the direction is clear: in 2024 and the first eight months of 2025, 111 Erdős problems moved from open to solved, and in 2026 the AI-assisted results arrived month after month.

Beyond maths: protein design and lab analysis

In Anthropic's experiment, Claude ran the design loop itself (choosing tools, generating and filtering candidates) and a wet lab tested the results. 22–35% of designs bound their target, against the 10–15% typical of current campaigns, and against at least six targets Claude produced binders with nanomolar affinity. On the RBX1 target, 40% of its designs bound, versus 3.7% of entries in a public competition, and its best design beat the winner. It also failed: none of 90 designs against MBP was confirmed.

The same post shows the less glamorous but more transferable use. Claude Opus 5 analysed NMR and LC-MS data for a compound in 23 and 19 minutes, reporting 96.4% purity against the lab's 96.33%. Manual analysis takes half an hour to an hour per sample, and the lab's finished report arrived four days after the first spectrum. That is the pattern most companies can copy today, and it is the same one we see in pharma and drug discovery: AI compresses analysis and iteration, people keep the judgment and the lab.

What this means for R&D teams in Croatia and the DACH region

1. Use AI for the "known answer" layer now. Derivations, literature synthesis, statistics, spectra and code for simulations are where models are at or above expert level. Pair them with reasoning models at a high effort setting and require the model to show its working.
2. Make verification part of the workflow. The strongest results in 2026 came with a checker: Lean proofs, wet-lab assays, Epoch's automatic grading. Treat any open-problem claim as a hypothesis until a person or a tool has checked it.
3. Plan for long, autonomous runs. The protein campaigns ran for 24–48 hours without a human in the loop, which matches the growth in agent time horizons. Budget compute and define stop criteria up front.
4. Pick the model per task, not per vendor. OpenAI leads formal maths, Google leads physics and GPQA, Anthropic has the strongest published results in agentic lab work, and Mistral Large 4 is the EU-hosted option for less demanding analysis.

Conclusion

In 2026 AI moved from passing science exams to producing results that experts check and accept: an 80-year-old conjecture disproved, Erdős problems closed every month, protein binders that beat a public competition. The benchmarks with known answers are saturated. Open research is still at single-digit percentages when the proof has to be machine-checked. For companies that is good news: the reliable part is large enough to change how R&D teams work, and the unreliable part is clearly marked.