How many "digital workers" could OpenAI deploy? We did a back-of-the-envelope estimate: restricting to tasks GPT-5 can do, OpenAI has the compute to run ~7 million digital workers. As AI automates more tasks, such a digital workforce could have big implications. 🧵

We've updated the look of our benchmarking hub! The new dashboard makes it easier to compare trends in performance across multiple benchmarks.

Sora 2 can solve questions from LLM benchmarks, despite being a video model. We tested Sora 2 on a small subset of GPQA questions, and it scored 55%, compared to GPT-5’s score of 72%.

What are current economic models missing about AGI? How would we know if we were approaching explosive growth?

Announcing our new AI Companies Data Hub! We collected key data on frontier AI companies, including revenue run rates, funding, staff, usage rates, and compute spend. This free resource will help you understand the trajectory and economics of AI. Highlights in thread!

Why did OpenAI train GPT-5 with less compute than GPT-4.5? Due to the higher returns to post-training, they scaled post-training as much as possible on a smaller model And since post-training started from a much lower base, this meant a decrease in total training FLOP 🧵

AI companies often grade their own homework. So are they cheating on the test? We wanted to see how companies’ test results compare to our own findings on the GPQA diamond benchmark. So what did we find? No signs of cheating! Here’s our methodology: 🧵

The future of long context inference isn’t just reading long documents. Inference with much longer contexts could unlock continual learning and massive RL scaling. But this will require overcoming high costs and slower research iteration cycles. 🧵

When unemployment spiked during COVID, a several trillion dollars stimulus package passed in the US almost overnight. In @YafahEdelman modal story of how AI unfolds, by 2035 we see a similarly dramatic response to AI-driven unemployment and de-skilling.

What will AI look like by 2030 if current trends hold? Our new report zooms in on two things: (1) whether scaling continues (compute, data, power, capital), and (2) the capabilities this enables—especially for scientific R&D.

What does it take to train cutting-edge AI? We looked into the numbers for Grok 4, which required the largest training run to date. Here's what we found: 🧵

Should AI regulations be based on training compute? As training pipelines become more complex, they could undermine compute-based AI policies. In a new piece with Google DeepMind’s AI Policy Perspectives team, we explain why. 🧵

AI progress has been driven by enormous compute scaling, but this is likely to slow down within the next few years. The reasons: investor uncertainty, the heavy costs of overinvestment, and increasing lead times. 🧵

New podcast episode: How far can current AI trends continue? Jaime Sevilla & Yafah Edelman on where current AI trends carry us. They disagree on mechanisms and outcomes, but agree on this: fast diffusion now, broad cognitive automation by ~2035, and extreme uncertainty after. youtu.be/Ab6HfmmCFQs

In less than a year LLMs have climbed most of the high school math contest ladder. Every tier of problem difficulty has either been saturated or is well on its way—except for the very highest tier.

Has LLM progress slowed? Initial reactions to GPT-5 were mixed: to many, it did not seem as dramatic an advance as GPT-4. Benchmarks may help clarify the picture: GPT-5 is both an incremental release following many other OpenAI advances, and a major leap from GPT-4.

We interviewed some of the mathematicians who contributed the hardest questions to FrontierMath on when they expect AI to solve their problem, what a correct solution would reveal about AI’s abilities, and the future of math in the age of AI. www.youtube.com/watch?v=nko...

We’ve renamed the ‘AI Supercomputers’ dataset to ‘GPU Clusters’ to better reflect that it covers all types of GPU clusters, including both general-purpose clusters and the supercomputers used to train the largest AI models. 🧵

What's the best model you can run on a single consumer GPU? We've updated last week's Data Insight with 3 additional benchmarks to judge capabilities. The result? Across all four benchmarks, small open models lag frontier performance by less than a year. 🧵 x.com/EpochAIRese...

Frontier AI performance typically reaches consumer hardware in just 9 months. With a single gaming GPU, you can run open-weight models matching the benchmark performance of the absolute frontier from less than a year ago. 🧵

We're expanding the Epoch AI Benchmarking Hub with five new external benchmarks: TerminalBench, METR Time Horizons, DeepResearchBench, GSO, and WebDevArena! These benchmarks test AI's ability to perform complex tasks through coding or tool use. 🧵

Let’s take a look into GPT-5’s record-setting performance on FrontierMath. How did it perform on the holdout vs. non-holdout set, how did it do across tiers, and what new Tier 4 problems did it solve? 🧵

We’ve independently evaluated the GPT-5 model family on our benchmarking suite. Here is what we’ve learned 🧵

The power required to train frontier AI models has been growing exponentially over time. What happens if trends continue? In a new white paper written in collaboration with @eprinews.bsky.social‬ , we analyze this question and forecast multi-gigawatt individual training runs by 2030!

The past 5 years have seen big successes in language, image and video generation, but relatively limited success in robotic manipulation. Why don’t we have laundry robots in every house? One thing seems clear: training compute is not the blocker. 🧵

OpenAI has historically scaled up training compute by around 100x with each new generation of its GPT. However, GPT-5 appears to be an exception to this trend.🧵

Multiple AI systems won gold medals at the 2025 International Mathematical Olympiad (IMO). Exciting as that sounds, @GregHBurnham argues that it represents little progress: an unlucky draw of problems made the event relatively uninformative. Is that cope? Judge for yourself. 🧵

GPT-5 sets a new record on FrontierMath! On our scaffold, GPT-5 with high reasoning effort scores 24.8% (±2.5%) and 8.3% (±4.0%) in tiers 1-3 and 4, respectively.

We have independently evaluated the new Claude Opus 4.1. We see 63% on SWE-bench verified and 7% on FrontierMath Tier 1-3 — a minor improvement over Claude Opus 4.

Is China about to leap ahead of the West in AI compute? According to Veronika Blablová and Robi Rahman,, the answer is no. Although Chinese hardware has been catching up, China still faces critical bottlenecks in chip manufacturing and their software ecosystem. 🧵

xAI commissioned us to analyze Grok 4’s math capabilities. Our findings: + It’s good at involved computations, improving at proofs (from a low base), and useful for literature search. - It favors low-level grinds and leans on background knowledge. Read on for examples!

Major AI developments happen fast, and often come with incomplete or misleading information. We want to help: you can now schedule a paid call directly with Epoch whenever you need quick access to our expertise on AI.

Should you start your training run early, so you can train for longer, or wait for the next generation of chips and algorithms? Our latest estimate suggests that it’s not effective to train for more than ~9 months. On current trends, frontier labs will hit that limit by 2027. 🧵

How fast has society been adopting AI? Back in 2022, ChatGPT arguably became the fastest-growing consumer app ever, hitting 100M users in just 2 months. But the field of AI has transformed since then, and it’s time to take a new look at the numbers. 🧵

We have graded the results of OpenAI's evaluation on FrontierMath Tier 1–3 questions, and found a 27% (± 3%) performance. ChatGPT agent is a new model fine-tuned for agentic tasks, equipped with text/GUI browser tools and native terminal access. 🧵

New time horizons doubling time estimates from METR, using data on GPQA Diamond, SWE-bench Verified and Mock AIME from the Epoch AI benchmarking hub. x.com/METR_Evals/s...

LLM context windows have grown, but can models really use all this content? We find signs of recent, rapid progress in their ability to do so. Read on to learn more!

The bottlenecks to >10% GDP growth are weaker than expected, and existing $500B investments in Stargate may be tiny relative to optimal AI investment In this week’s Gradient Update, Andrei Potlogea and Anson Ho explain how their work on the economics of AI brought them to this view.

How does token generation speed trade off against cost in LLM inference? We introduce a framework to answer this question and use it to examine the token economics of known LLM architectures.

Do the biorisk evaluations of AI labs actually measure the risk of LLM-driven bioweapons development? To answer this question, we analyzed the biorisk safety evaluations (or lack thereof) of 8 major labs. Here’s what we found:

Accuracy takes time: LLMs with higher benchmark scores take longer to run. Anyone who has compared models knows this intuitively, but what exactly is the trade-off? Here’s what we found!

Our AI Supercomputers database is now live! Explore specs for 700+ training & inference clusters—compute, chip types, power draw, ownership, location, and more.

How do reasoning models solve hard math problems? We asked 14 mathematicians to review o3-mini-high’s raw, unsummarized reasoning traces on 29 FrontierMath problems. Here’s what they found:

SWE-bench Verified is one of the main benchmarks to assess AI coding skills. But what does it actually measure? We found that it's one of the best tests of AI coding, but limited by its focus on simple bug fixes in familiar repositories. Here’s a summary of our article 🧵

How do models perform on the holdout set of FrontierMath? To test this, we evaluated six models on both the holdout and non-holdout sets of FrontierMath. We find that performance is broadly consistent across these two subsets.

Three years and 100+ projects in, our mission remains the same: give everyone clear, trustworthy insight into where AI is headed. Our new post breaks down the principles that shape every research decision we make—why we pick certain projects, and why we pass on others. epoch.ai/blog/what-is...

We're expanding the Epoch AI Benchmarking Hub with four more external benchmarks: VPCT, Fiction-liveBench, GeoBench, and SimpleBench! These benchmarks test visual physics understanding, Geoguessr ability, long-context comprehension, and reasoning and logic skills. 🧵

Anthropic’s progress on SWE-bench Verified really stands out. Naive extrapolation suggests they’ll have it mostly solved in another year.

Today’s largest AI models are trained on over 10^26 FLOP. How many will surpass this threshold in the coming years? We project 30 notable models by 2027, rising to 200 by 2030. This rapid growth could pose challenges for threshold-based AI regulation.

The speed of computations on GPUs depends directly on the numeric format: less precision means more calculations on the same hardware. We analyzed the numerical format used to train 272 models from 2008 to 2025. Here’s what we found. 🧵