What these things can do

Jonas Moss · Data Science and Analytics

16 October 2026

Competitive programming: the world finals

AtCoder World Tour Finals: the AI’s score as a multiple of the best human’s score. Two divisions, each with its own scoring.

Best human 0.9× Psyho, a human, wins: “Humanity has prevailed (for now!)” about 8× 50 billion vs 6.3 billion points 1.9× 8,300 vs 4,300 points Heuristic July 2025 Heuristic July 2026 Algorithm July 2026

ICPC World Finals 2025: AI solves 12 of 12 problems, best human team 11.

Epoch AI chart: best score so far on maths benchmarks by model release date, 2020 to 2026. FrontierMath Tiers 1 to 3 (teal) rises from about 0% in early 2024 to about 90% in 2026; FrontierMath Tier 4 (orange) rises from 0% in early 2025 to nearly 100% in 2026.

FrontierMath: unpublished problems written by mathematicians to stump AI. Tier 4 (orange) is research level.
These tests are close to full, so the labs now also test on open problems.

Chart: Epoch AI, CC BY

A released model on open problems

OpenAI, September: a curated set of open maths problems, ones nobody had solved.

OpenAI chart: pass rate on a curated set of open maths problems against test-time compute on a log scale. The internal model rises from about 0.26 to about 0.47; GPT-6 Astra rises from about 0.08 to about 0.17.

GPT-6 Astra, released 3 September: up to about 1 in 6

Internal model, unreleased: nearly 1 in 2

The same Astra on FrontierMath Tier 4, the research-level test on the previous slide: 97.6%

Claude users: on Epoch’s capability index, Claude Fable 5.1 scores within two points of GPT-6 Astra (164.5 against 166.3).

OpenAI, 8 September 2026 · Capability index: Epoch AI, 20 September 2026 · Tier 4: Epoch AI, Astra at medium reasoning, run data 20 September 2026

6 October: the same internal model

OpenAI posed about 4,000 open problems to it, as a test: problems nobody had solved.

4,000 open problems tried

372 results

719 papers

3 hours of compute per result

42% checked by computer

By one count, 90 of the 500 best-known open problems in mathematics.

OpenAI: “Some of the unformalized results could have issues.” Three papers were withdrawn in the first two days.

The biggest day in history?

It isn’t an overstatement to say that this is probably the biggest day of scientific advancement in history.

Joshua Gans, economist, Rotman School of Management, Toronto

Big, big, big, big props for quasiriemann and no Siegel zeroes […] It’s obviously the most significant moment in mathematical history.

Levent Alpöge, mathematician at Anthropic, on OpenAI’s results

The New York Times front page that day had no mathematics.

A Fields medallist’s reaction

Not a single one of the major open problems I have publicly mentioned throughout my career […] was left untouched by the announcement. Everything has been claimed to be proved.

I expected to see a few of them in the list. But not all of them. Not all at once.

Hugo Duminil-Copin, Fields Medal 2022

“The shock is immense. I am paralysed. Tomorrow, we will find a way forward.”

For teaching: animations, written as code

Animated explainer: saving 1,000 dollars a year at 7 percent for 40 years. The bars grow year by year; the interest part overtakes the savings part. After 40 years the total is about 200,000 dollars: 40,000 saved and 160,000 interest.

Make a 15-second animated SVG that explains [compound interest] to first-year students. Three scenes, one caption each.

Made by Claude Code for this slide: one SVG file, no video software. Also in the Claude app: Design animations, Motion.