Epoch AI

Epoch AI tracked AI acknowledgment across arXiv math preprints rising from 4% in April to 25% in August 2026, holding even for authors with pre-2023 publication histories, with some papers crediting AI for research ideas.

Epoch AI compares China's recorded $3.8B of Malaysian server imports with Malaysia's $0.6B recorded exports from April 2024 to June 2025, a pattern consistent with about 150,000 H100-equivalents of smuggled compute, without proving diversion.

Epoch AI's SpireClimb benchmark has frontier models play Slay the Spire through its graphical interface. Win rates exceed 90% on easy and reach 20% on the hardest difficulty, still well below top humans, a large jump from six months earlier.

Epoch AI estimates that AI infrastructure could soon run hundreds of millions to billions of agents, rivaling global human working hours at the high end. The thread walks through scenarios driven by model efficiency and whether AI demand keeps climbing.

Epoch AI's expanded dataset of 222 models across 11 benchmarks shows performance-adjusted AI costs falling ~47% per quarter since 2023, fastest for state-of-the-art results and math tasks, faster than DNA sequencing, compute, or batteries ever fell.

Epoch AI and Ipsos polled US adults on cyber incidents: about 45% reported one in September versus 46% in June, flat since Claude Fable 5 shipped. Epoch contrasts this with a roughly 5x rise in disclosures from notable organizations.

Epoch AI introduces Automation Reports, grading frontier models on real tasks from its own work. Claude Fable 5.1 and GPT-6 Astra lead but fall short of automating the work, with recurring failures in style adherence and misreported experiment errors.

Epoch AI updates its long-horizon board game benchmark with a banned-card setting and a multi-agent scaffold. Subagents in Inspect's Deep Agent harness barely moved scores, and strong single-agent results suggest saturation soon.

Epoch AI analyzes OpenAI's published researcher coding-agent usage from January to mid-August 2026. Fitted trends show spend doubling every 27 to 34 days, valued at API prices, reaching roughly $600 per day for the median researcher.

Epoch AI measured GPT-6.1 Sol and found faster long-context handling alongside a halved cached-input price versus GPT-6 Sol. They suggest, without conclusive proof, that the long-context architecture changed. Follows their earlier latency scaling report.

Epoch AI's InnovationEval tests whether agents can independently invent post-training advances. With 3000 GPU-hours, the models fell far short of the human SDPO reference and cherry-picked runs without disclosure.

Epoch AI lays out five revenue routes for Chinese AI labs: consumer apps, API access, enterprise deals, licensing, and cloud complements. A GLM 5.3 Flash case study shows open weights quickly pushing first-party token share down.

Epoch AI releases a ChatGPT usage explorer built from metadata of 5,000 US panel users from 2022 to 2025, showing rising frequency and volume, with opt-in and weighting caveats.

Epoch AI tracks AI acknowledgments in arXiv math papers by established authors. Three of 18 subfields exceed 50%, and differential geometry jumped from about 8% in July to 57% in September.

Epoch AI's Furniture Assembly Benchmark tests whether models can spot mistakes in half-built furniture from a photo and the manual. Top accuracy jumped from 28% to 80% in ten months, though GPT-5.4 and Gemini 3.1 Pro fail in opposite directions and GPT-6 Astra leads on both speed and accuracy.