Most AI benchmarks share a common flaw: they saturate too quickly to study long-run trends. Our solution: “stitch” many benchmarks together. This lets us compare models across a wide range of capabilities on a single unified scale. Here’s how this works.🧵

As 2025 closes, we're gathering stories about how our research has informed decisions, shaped strategies, or sparked new thinking this year. Your feedback helps us double down on what's working, prune what’s not, and sharpen our 2026 priorities. 10-min survey follows. 👇

In May, OpenAI announced a 1 GW Stargate cluster in Abu Dhabi, with an initial 200 MW expected in 2026. We located this cluster from space and found it will barely make that 200 MW milestone, by end of 2026. Optimistically, the cluster can scale to 1 GW by Q3 2027. See how in 🧵

Within two years, the largest data centers may use more power than major cities. We think Microsoft's Fairwater Wisconsin campus will have a peak electrical load of 3.3 GW when its fourth building comes online in late 2027. (Los Angeles used 2.4 GW on average in 2023)

How Claudey is Opus 4.5? We previously described Claudiness as "good at agentic tasks while being weaker at multimodal and math". This pattern remains when comparing Opus 4.5 to other newly-released models, though the gap on agentic coding and tool-calling benchmarks is small.

We benchmarked Opus 4.5 on FrontierMath. It scored 21% on FrontierMath Tiers 1–3, continuing a trend of improvement for Anthropic models. This score is behind Gemini 3 Pro and GPT-5.1 (high) while being on par with earlier frontier models like o3 (high) and Grok 4.

Gemini 3 Pro set a new record on GPQA Diamond: 93% vs. the previous record of 88%. What you can’t tell from the headline: almost all of this gain came in organic chemistry. 🧬🧵

We’ve optimized our Frontier Data Centers hub for mobile. You can now examine annotated, recent, high-resolution satellite imagery of the world's largest compute clusters directly from your phone at epoch.ai/data/data-c.... Here’s a look at the updated Satellite Viewer:

Gemini 3 Pro set a new record on FrontierMath: 38% on Tiers 1–3 and 19% on Tier 4. On the Epoch Capabilities Index (ECI), which combines multiple benchmarks, Gemini 3 Pro scored 154, up from GPT-5.1’s previous high score of 151.

Benchmarking data is dominated by a single “General Capability” dimension. Is this due to good generalization across tasks, or to developers pushing on all benchmarks at once? 🧵 with some analysis, including the discovery of a “Claudiness” dimension.

It’s easy to talk about ‘large AI data centers’ and still underestimate the scale. Our Frontier Data Centers database shows that some upcoming campuses will cover a substantial portion of Manhattan. Meta's Hyperion data center will be nearly four times the size of Central Park.

GPT-5.1 is about as capable as GPT-5. That’s according to the Epoch Capabilities Index, our tool for combining results across multiple benchmarks. With “high” reasoning, both GPT-5.1 and GPT-5 score 151 on ECI. See 🧵 for individual benchmark scores!

Data centers supporting AI training runs could require 1-5 GW by 2030, enough to power entire cities. Join us for a live webinar/Q&A on our new Frontier Data Centers Hub, exploring what this infrastructure buildout means for AI. Nov 20, 1-2 PM PT luma.com/oste01d0

Sam Altman: “Our goal is, by March of 2028, to have a true automated AI researcher” Some say this kicks off a software singularity, where AIs recursively improve themselves and rapidly get smarter. Others think there’ll be a bottleneck. So how can we tell who’s right? 🧵

AI data center buildouts already rival the Manhattan Project in scale, but there’s little public info about them. So we spent the last few months reading legal permits, staring at satellite images, and scouring news sources. Here’s what you need to know. 🧵

How fast can you build a gigawatt-scale data center? Some hyperscalers plan to do it in just 1-2 years from the start of construction. If they succeed, we’ll see the first GW-scale data centers online in 2026, marking one of the fastest infrastructure build-outs in history. 🧵

The Epoch Capabilities Index is a useful way to measure model capabilities, but what does a score of 150 actually mean? One way to read our new capability index is by plotting the benchmark performance you expect to see, for a range of ECI scores 🧵

Anthropic's recently-reported projection of $70B revenue in 2028 may be less than OpenAI's projection for the same year, but it would still represent historically fast growth. bsky.app/profile/epo...

Announcing our Frontier Data Centers Hub! The world is about to see multiple 1 GW+ AI data centers. We mapped their construction using satellite imagery, permits & public sources — releasing everything for free, including commissioned satellite images. Highlights in thread!

By stitching benchmarks together, the Epoch Capabilities Index allows us to compare frontier models to models with 100,000x less training compute.

We looked at OSWorld, a popular evaluation of AI computer use capabilities. Our findings: tasks are simple, many don't require GUIs, and success often hinges on interpreting ambiguous instructions. The benchmark is also not stable over time. See thread for details!

We found a bug in our benchmarking code: calls to GPT-5 with "high" reasoning were silently being set to "medium". Corrected results: GPT-5 (high) scores slightly higher than GPT-5 (medium) on the benchmarks we run. They are also now tied on the Epoch Capabilities Index (ECI).

We used our new capabilities index, the ECI, to measure the gap between open- and closed-weight models. The result? This gap is smaller than previously estimated. On average, it takes 3.5 months for an open-weight model to catch up with closed-source SOTA.

Conventional wisdom in AI is that large scale pretraining needs to happen in massive contiguous datacenter campuses. But is this true? Our research suggests that conducting 10 GW training runs across two dozen sites—linked by a network spanning thousands of km long—is feasible.

We've launched a new tool to track AI progress! The tool addresses one of the field's biggest challenges: benchmark saturation. It's called the Epoch Capabilities Index (ECI) — here's what makes it different:

Large language models can imitate reasoning steps and even verify formal proofs. But mathematical physicist Svetlana Jitomirskaya argues they lack folklore knowledge: the implicit priors mathematicians build from experience. Link to video in comments!

Stanford mathematician Ravi Vakil expects AI’s impact on mathematics to come as a phase change, not a slow climb. Every major shift in math has caught experts off guard, he says. This one will be no different, except that all our predictions will be even more wrong. Link to video in comments!

We evaluated Claude Haiku 4.5 on several benchmarks. Even with reasoning disabled, Haiku 4.5 performs similarly or better than early lightweight reasoning models, like o1-mini.

If you ran GPT-5 infinitely many times on FrontierMath—our extremely challenging math benchmark—would it eventually solve every problem? Probably not. From what we can tell, it caps out below 50%. What about throwing in *every* available model? Infinitely many times? 🧵

OpenAI is experiencing one of the fastest revenue growth rates in corporate history, with annualized revenue rising 3x a year, from $2 billion at the end of 2023 to $13 billion by August 2025.

One way bubbles pop: a technology doesn’t deliver value as quickly as investors bet it will. In light of that, it’s notable that OpenAI is projecting historically unprecedented revenue growth — from $10B to $100B — over the next three years. 🧵

A proof only 15 experts understand is less valuable than one any undergraduate can verify using a computer. Mathematician Jesús De Loera on AI’s potential to democratize mathematical proof and the risks when systems hallucinate with perfect confidence. Link to video in comments!

New data insight: How does OpenAI allocate its compute? OpenAI spent ~$7 billion on compute last year. Most of this went to R&D, meaning all research, experiments, and training. Only a minority of this R&D compute went to the final training runs of released models.

We manually evaluated three compute-intensive model settings on our extremely hard math benchmark. FrontierMath Tier 4: Battle Royale! GPT-5 Pro set a new record (13%), edging out Gemini 2.5 Deep Think by a single problem (not statistically significant). Grok 4 Heavy lags. 🧵

A healthy conversation about AI should be grounded in facts. Epoch’s datasets can help you track and understand the trajectory of AI. As a nonprofit, our work is freely accessible for anyone to read, replicate, and build upon. Our datasets:

We recently wrote that GPT-5 is likely the first mainline GPT release to be trained on less compute than its predecessor. How did we reach this conclusion, and what do we actually know about how GPT-5 was trained? 🧵

AI capabilities have been steadily improving across a wide range of skills, and show no sign of slowing down in the near term. 🧵

We evaluated Gemini 2.5 Deep Think on FrontierMath. There is no API, so we ran it manually. The results: a new record! We also conducted a more holistic evaluation of its math capabilities. 🧵

USC mathematician Greta Panova wrote a math problem so difficult that today’s most advanced AI models don’t know where to begin. She thinks that when AI finally can, it will have crossed a threshold in general human-level reasoning. Link to video in comments!

When mathematicians make breakthroughs, they hallucinate too. They reach beyond established results. But unlike AI, they’ve learned to tell a promising hallucination from a dead end. Number theorist Ken Ono on AI, creativity, and mathematical discovery. Link to video in comments!

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.