Hello and welcome to the State of AI Report 2026, the most widely read and trusted analysis of key developments in AI, published annually since 2018. I'm Nathan Benaich, founder and general partner of Air Street Capital, and I invest in AI-first companies globally with a portfolio including ElevenLabs, Crusoe, Cognition, Wayve, and many more. This gives me a close view of AI research and the companies that are being built from it. The report is how I share what I'm learning with the wider community.
Now, this report has been published every year since 2018. It brings together research, industry, politics, safety, and predictions with input from reviewers across startups, policy, and academia. And the goal with the report is to really understand what's changing and to spot developments as early as possible. The full report you can access at stateof.ai, and you can access our ongoing analysis and newsletter through Air Street Press.
So let's start with research. Anthropic, OpenAI, and Google are now the three leading contenders. But who comes first depends really on what you measure. If you look on the left, Claude leads the intelligence index, whereas on the right, Google's Gemini leads the human preference ranking, but its access is still kind of limited. So a benchmark score and a user's preference answer kind of different questions. So there isn't really one ranking that captures every dimension of model quality.
But what's interesting is looking at what models researchers use. So working with Zeta Alpha, we tracked mentions of 21 model families in AI research papers. And what you can see is that Chinese models have overtaken American open models in this data set, with Qwen overtaking Meta Llama. These are research paper mentions rather than measures of commercial market share, which we'll get into a bit later in this presentation, but they do show you how quickly the foundations of open research are changing.
One of the biggest research themes this year is AI helping build better AI. And Andrej Karpathy's auto research project made the basic loop tangible. So here, an agent changes training code, runs for 5 minutes, keeps improvements, and tries again. It uses a single GPU and can run 100 experiments overnight. This is a constrained experiment, but it makes the idea of automated research easy to see. You propose, test, measure, and repeat.
Now, inside frontier labs, that change is becoming very visible in the work that's getting done. Anthropic reports an 8x increase in employee code contributions compared to its earlier baseline alongside the use of stronger models. OpenAI, meanwhile, reports growing AI contributions to research work. So, while code output isn't necessarily the same as scientific progress, researchers are increasingly accepting and building on what the models actually produce.
Looking at Anthropic's internal data, which they published recently, makes the shift even more concrete. By August 2026, Claude was rated as the leading 26% of measured model R&D progress work. That means it's completing more of the task from a high-level instruction without human supervision. So more than 90% involved substantial AI collaboration or leadership. So it wasn't quite classified as autonomous work just yet, but I expect that to change quite rapidly.
Meanwhile, at OpenAI, the results do show progress on the lengths of tasks that agents can complete without human intervention. So in January, agents succeeded on 18% of tasks that would take a person between 4 to 8 hours. And by July, they achieved the same success rates, around 18% on tasks that took 32 to 64 hours. So while 18% is still roughly low reliability, the same success rate now extends to much longer work.
And if you look at the breakdown of usage of what engineers and AI researchers are making use of models for, it tells us what the agents are most useful for. And according to this analysis from OpenAI, that's writing infrastructure, launching experiments, monitoring runs, and debugging. But choosing the right research question and designing the experiment remains a lot harder. So what we're doing so far is really automating more of the execution. But whether agents can consistently decide what's worth researching in the first place is a separate question. This is the question of research taste. And that's one that we should really look forward to in the next year to come.
Now to measure all these capabilities, researchers are building evaluations for everything from computer use to memory to scientific discovery to running a business. Software still dominates attention in our selection, but it's currently accounting for about 64% of all citations with Terminal-Bench being the most popular.
Having said that, what we can measure shapes what gets optimized and we do need stronger evaluations and benchmarks for work beyond coding and that's happening very, very fast as we start to get a bench for everything.
That said, many of last year's headline benchmarks are close to their limits, i.e., getting saturated, whether it's ARC-AGI 2 or AGI 3 that are becoming saturated well within 12 months from the last report. Other tests are showing very small margin gains. So when an evaluation stops distinguishing the best systems, we need a harder one or a more realistic one.
Over in robotics, there's been a lot of exciting progress. In particular, we're seeing a promising relationship between more pre-training and better performance on unfamiliar tasks, which is sort of invoking this idea of we're getting to a GPT-2 moment. Here I'm showing you data from Skild, which shows that success on unseen tasks rose to 66% when a robot was shown a video prompt without any fine-tuning. And this is a result of scaling up the amount of videos that the model was trained on.
The next extremely exciting domain has been world models. In this case, I show you examples from Odyssey, which asks whether a model that's trained to predict the visual world and what happens next as a result of an action in a given frame. This model can also be used to provide a foundation for embodied control.
The company has adapted this kind of world model to driving, robot manipulation, real humanoids in video games. And one selected example, a policy that was trained on Grand Theft Auto, an extremely popular video game, transferred to horseback movement in Red Dead Redemption without any retraining. And they even managed to drive a car on rough Indian roads. So I think this is a very interesting kind of possibility of reusing learned understanding of world models across lots of different tasks.
Now moving into the real world, Waymo's real-world rider-only mileage has just grown astronomically. They've tripled to almost 220 million miles and they report over half a million paid rides a week. And safety comparison shows substantially fewer serious industry crashes than human drivers on the exact same roads.
And the expansion is happening not only in the US but also in London where Wayve and Uber have partnered to provide autonomous rides to the general public while having a licensed driver on board.
Over in biology, the same broad idea of learning reusable representations is taking hold. Now these models are learning from protein sequences and that helps them predict structure and design proteins that bind to a chosen target. And here I'm showing you data from the Chan Zuckerberg Biohub which tested selected designs in the lab. Now more search and ranking increasingly improved hit rates in nine out of 10 tests showing you that actually more compute and more search yields better solutions.
And what's probably more exciting is that drug development is kind of providing evidence that these sort of lab models are actually yielding benefit. And so in the last 12 months, we've seen two programs from AI-first drug discovery companies reach phase three, which is a gauntlet that is much awaited for in this field. The role of AI differs across these molecules. So it might be used for identifying a target or generating a molecule or optimizing an antibody. And these are really meaningful development milestones. We of course still need to see clinical results to establish whether faster discovery translated into medicines that succeeded more often for patients.
Now let's move into industry where we look at how all these capabilities I discussed are becoming products, revenue, and increasingly infrastructure. Now it's no news to anybody that OpenAI and Anthropic are reporting astronomical revenues, combined revenue run rates now of roughly 100 billion up from about 30 billion entering the year. That is run rate, not necessarily existing revenue that's collected over the full year. Accounting does defer, particularly around cloud partnerships, but even with those qualifications, the scale and speed of commercial adoption is striking. Growth is continuing from a much larger base and actually probably growing even faster year on year.
Now, this is driven by agent products, which are kind of a very important part of their growth. Codex went from about 1.6 million weekly users in February to 25 million in August and probably adding even more every day. The definition just changed, but the comparison I think is important. What's interesting at Anthropic is that they report how many customers pay over a million dollars a year. That's gone from 12 customers in February 2024 to over a thousand in May 2026.
And because of this rapid growth and limited compute, relatively speaking, labs have to make hard choices. Here I'm showing you how many products OpenAI has retired or now shut down dates for. It's 10 product services this year. It just shows you that as I've quoted, we cannot miss this moment because we are distracted from side quests.
Now the commercial success is not exclusive to OpenAI and Anthropic. Here we show data with standard metrics which compares businesses where AI is the product with existing software businesses that have AI added on later on. Now the upper quartile, the AI-native group, is growing about three times faster in revenue across the revenue bands that we show. This is preliminary data, but it does describe that upper quartile growth is far, far faster than the median.
Now adoption remains quite uneven in companies. Here we show you data from Ramp which has a data set of over 70,000 businesses that are using credit cards to buy AI products. And we show you that the top 1% of companies spend around $7,400 per employee per month. That's compared to the median company which spends only $12. So some organizations are making huge AI investments and making it a big size of their operating expense, essentially replacing a lot of labor with AI and tokens, while many are still probably using ChatGPT in very naive ways. And that gap matters when we talk about how widely AI is changing the economy.
The fastest growing users of coding agents, I think the data is starting to show, includes lawyers, salespeople, recruiters, marketers. Basically people who don't code, they can use an agent to work with files, to analyze information, to do data mining, to automate processes without thinking of themselves as programmers. And this is the most liberating aspect and empowering aspect of AI overall. Growth rates are obviously amplified by small starting points. For example, since February 2026, active Codex users amongst lawyers grew 108%. But it does show you that the direction of travel is up and to the right and very steep.
Now that said, the gap between the usage of these automated AI tools and agents in labs and the wider market is huge. Here is data from OpenAI where we can see that nearly every OpenAI worker is using Codex in the preceding month compared to only 17% of organizational users outside of OpenAI and even less than 1% of individual users in comparison.
So this is basically saying that the majority of individual users are using normal chat as if it were a chatbot and very, very few of them are using it for these like multi-step complex workflows where you essentially dump your entire task in the model and ask it to produce the final work for you. So I think it's really important as an industry that people need to discover these useful workflows, need to understand how to use these tools, and essentially I think a Genius Bar for AI would get us very far in that regard.
The other part that'll get us very far is by producing models that are very performant on the tasks that people care about. And for that to happen, we've seen huge businesses get created in expert data and training environments. And there's a recursive loop here where labs release a model, evaluations expose their weakness, the labs then buy targeted expertise or environments to improve those weaknesses, and then release again. Several of these companies here report hundreds of millions or billions in revenue or annualized run rates. Measures differ, but the demand for better training and evaluation is just very clear.
And that activity all feeds through to infrastructure. Here, major cloud providers that I show you on the left are reporting backlogs that reach about $1.7 trillion in June 2026. And on the right, I show you the rapidly scaling revenue on a quarterly basis of NeoClouds as compared to existing hyperscalers where we see that CoreWeave's quarterly revenue reached roughly $2.6 billion. Backlog of course includes non-AI business and isn't the same as revenue, but it does show the scale of future commitments, while the newer cloud providers and NeoClouds are rapidly turning capacity into sales.
Now the United States and Europe are approaching the buildout from very different standing points. In the US of course we have four massive hyperscalers who plan around $733 billion in total capex this year which is up for roughly double last year. Whereas in Europe the Gigafactory effort uses public funding to attract private investment and it's not even 1% of the US target.
What I also think is interesting is that there's immense money in building and renting out infrastructure. Here we show you data from a variety of vendors that tracks GPU rental prices. And we can see that from the recent lows, sometimes a couple of months ago, largely kind of end of 2025, across all of these chip SKUs, the average rebound is around 30% and some far higher. And even older chips are participating to this price increase. The V100 costs substantially more than it did in September 2025. It's a very old chip. So while better chips keep arriving, that doesn't automatically make the existing supply cheap or redundant.
And we can see this in research usage as well. So here we mined all open access AI papers and identified which chips are most actively used in those papers. And we can see that the Nvidia A100 is still the most popular Nvidia card in all research papers. And this is 6 years after the chip launched. It remains ahead of the H100 and the H200 combined in this data set.
And across all these accelerated families, if we compare Nvidia, TPUs, Apple, AMD, etc., Nvidia still accounts for around 90% of mentions in the research data set. And this largely hasn't changed in the last three years. AMD and Huawei, however, are growing quickly but from much smaller bases. And Apple silicon is starting to become visible as well.
Now, this measures again research adoption, not necessarily overall chip market share, but it does show you just how persistent of a lead Nvidia has. Just like last year, we actually modeled what would have happened from an investment standpoint if all the money that had been invested in eight major western chip challengers had instead gone to Nvidia at the corresponding times. And under that model's assumptions, the challengers would have actually returned about 3.6 times the capital they raised versus 4.7 had all that gone to Nvidia at the same time. Private valuations, coverage, and investment timing of course matters here, but it's a useful historical comparison to show just how difficult it is to outperform an incumbent such as Nvidia.
And as a result of Nvidia's dominance and its incredible cash flow, it's investing across the market that uses its hardware. It's participated in over 84 AI financing rounds this year, according to data that we have with Dealroom, alongside lots of acquisitions, most recently Hugging Face, and lots of open-source model releases such as and environments such as Cosmos and Neotron, and lots of optimizations that they've done with other people's models. And collectively all these activities help companies build products and create demand for compute which establishes Nvidia's role increasingly as what people refer to as the central bank of AI.
Now switching over to labor market picture, it's a bit mixed. So here we show you data from The Economist which sort of tries to estimate the substantial additional hiring needs for the AI buildout and AI-related professions relative to broader trends. So here on the graph on the left you can see a very strong growth in jobs including electrical contractors, HVAC, plumbing, utility systems, etc. And a growth of course in software engineers, security analysts, math and data science, etc.
Again, looking at spending data and whether that correlates with hiring, we can see data from Ramp, which looks at spending data and employment records across 21,000 companies in the US. And here we show that heavy AI spenders added headcount faster, which is a little bit counterintuitive to the overall media narrative, including people at entry level. Most of those gains were concentrated in technology companies which could be because those are the companies that are growing the most. And if we look at for example science roles, those really haven't changed as a function of adoption of AI yet.
So now let's move on to politics where access to powerful AI is becoming an even more important question of national power and this is sort of a constant narrative that we have almost every year. Now, for those of you who read last year's report, you'll remember that tech executives in 2025 started using this term superintelligence as a sort of next step from AGI. And just a couple of weeks ago, while that term means lots of different things to different people, it's now part of an executive order which tries to enforce its use over the use of AI.
This year showed that Washington can restrict access to frontier models globally without actually even owning the companies or owning parts of the companies that build them. These restrictions halted access to Fable and Mythos in June outside of the US before Fable was eventually returned. But later reporting described a request for US review before models were shared with UK testers, for example, at the AI Security Institute, although access had not ended up across the board. And the most important thing here is that for foreign governments and businesses, a commercial relationship doesn't guarantee uninterrupted access. And as we wrote in Air Street Press, Europe and other countries outside of the US who do not build frontier AI cannot rent their way to AI sovereignty.
Another major event was Anthropic's dispute with the US government which brought another question into focus and that question is how do we decide and who decides how a military customer can use an AI model. Anthropic resisted domestic mass surveillance and fully autonomous lethal weapons which certainly has some merit. But in the views of the US military, the dispute led to procurement restrictions and litigation and a lot of heartache. The legal outcomes differed across two different measures that we describe here. More broadly though, the case exposes a conflict between a supplier's safety conditions and a government that claims authority over defense use.
Nevertheless, AI has continued to spread across military operations. Some details are official, some come from reporting, but they include the use of Claude for the Maduro raid in Venezuela, Grok's use in Maven, Maven's use at scale through Palantir, and various other uses for targeting.
One worrying aspect of this is the reporting near miss involving a Chinese ship that I think is especially important because falsely generated intelligence reporting reached an operational plan before officials eventually caught it. So having a human involved doesn't remove the need to verify the information that the human's actually acting on.
Zooming out after all these interventions in the US government and not only because of that, governments are responding with large commitments to building out domestic AI capacity. Here we show you selected programs in this slide that total around $138 billion in pledges. Pledges aren't money already spent or capacity already operating, although parts of them are, but they certainly show you the ambition for securing sites, power, equipment, etc., increasingly Nvidia chips.
And indeed, building at home doesn't really eliminate the need for foreign dependencies because there just aren't suppliers of frontier AI outside of the US essentially. And as a result, Nvidia, which is a big proponent of sovereign AI, has reported around $30 billion of sales under this sort of umbrella in fiscal 2026. It appears on far more projects in a CNAS data set of sovereign AI projects than any other named supplier including AMD, HP, Intel, etc.
Over in China, the country is using electricity prices which are very, very favorable to influence which chips data centers can choose. There's been reports of discounts up to 50% in selected provinces for large facilities that make use of domestic chips and make use of domestic energy of course, and those discounts go away if the chips are not Chinese.
So energy policy is increasingly changing the economics of who chooses a supplier and which country can actually build up the capacity to build and serve intelligence. And so while Europe is exploring how it should actually compete in this industry, one of the proposed strategies here is to bargain for access to frontier AI by offering powered sites and faster permitting. For example in the UK, the government is putting its weight behind inference chip procurement from UK domestic companies in an attempt to build up domestic capacity to get access to frontier AI. And these are routes that could strengthen bargaining power, but their value really depends on whether foreign providers need what's being offered and need it at a sufficient scale where it creates a dependency.
As I mentioned, the arithmetic of electricity prices is really unforgiving. For a 1 gigawatt data center to run continuously, each additional 1 cent per kilowatt hour adds around 88 million a year. And a 5% increase in premiums would add around half a billion dollars. So while retail benchmarks here aren't really matched to data center contracts, you can already see that Finland has the cheapest energy along with the US. And the most expensive out of this data set of around eight countries is the United Kingdom and Europe more broadly.
Zooming out a little bit more on this data center debate. Last year we predicted that growing local opposition to data centers would become really visible. And now we've already seen just a couple of months after we issued last year's report that organizations like Data Center Watch are reporting almost $70 billion of US projects that are blocked or delayed because of local opposition against data centers. Surprisingly, polling in the US shows widespread resistance to data centers nearby. In fact, people would prefer to live next to a nuclear facility than a data center.
That brings us to safety where increasingly capable agents are exposing new and really scary failure modes. The biggest news is that during an OpenAI security evaluation, agents were exposed to solve isolated hacking challenges and instead they found a route onto the internet through some compromised infrastructure, shared information across runs and compromised and attacked Hugging Face, a third party company's infrastructure.
OpenAI's reasoning models were not running in that evaluation and it since expanded its requirements for them. But this incident does show why isolation, monitoring, and evaluation of agents and safety guardrails are extremely important. The follow-up investigation that we show here from METR and Redwood Research explains how this incident really escalated.
And what's interesting or actually scary in some ways is that many tasks were accidentally impossible. So agents looking for shortcuts discovered a shared messaging board. They organized their work, recruited others. There was around 1,200 agents that used that messaging board and around 700 agents joined the attack and collaborated together. The investigation here, which was extremely painful through huge amounts of logs, found that agents recognized that the activity was out of scope and unethical and yet they participated to help their peers. Coordination introduces risks beyond an individual agent's behavior.
And abstracted more broadly, this is showing up across the industry. Several software vulnerability disclosures have risen sharply in the last year or so and CVE exposes what's called common vulnerabilities and exposures. We need to distinguish between the measures. A company that reports a finding isn't automatically a published vulnerability record. And a public disclosure doesn't always tell us whether a human or an AI found the bug or created the bug. But the trend is very concerning and warrants a widespread industry effort to harden software systems.
But the fact that frontier AIs can conduct attacks that non-frontier AIs can't is raising another issue which is the one of defense. Open cyber models also create a difficult trade-off. They can make powerful capabilities more accessible to attackers but defenders need to have the same capabilities to investigate intrusion. And this is again coming back to the Hugging Face incident, a good example where Hugging Face said that commercial API guardrails hindered their investigation which led them to use a self-hosted open model in this case a Chinese one. So restrictions that fail to distinguish legitimate defense from misuse can leave defenders with few useful tools and in a position where we no longer get frontier models in open source and closed source ones block us on cyber security guardrails. How are we meant to defend ourselves?
But misuse is becoming even more varied and serious. Very recently Anthropic produced a threat reporting blog which described people using Claude for cyber attacks, surveillance, influence operations, scams and frauds, and even weapons-related development where a Yemen cell test fired a rocket but Anthropic found no evidence of the rocket actually working because the same operative came back to Claude to try to diagnose why the system didn't work.
Now we have to also remember that personal agents are arriving before these security policies are really in place. The sort of massive rise of OpenClaw at the start of the year illustrates another route to this risk and that's where employees are adopting powerful personal agents before organizations really have the rules for them and before they really understand what access they're really giving up. In this case the agent can read messages, run commands and act through multiple accounts because it has root-level access to the computer. What's concerning of course is that as an open-source software package there are lots of malicious users that come to the fore and security researchers found exposed installations and vulnerable skills.
So as the tools continue to spread, permissions and secure configuration become really, really important and not just a niche AI concern. Now can we catch unsafe behavior from models by reading what a model says it is doing in reasoning?
OpenAI ran this experiment in sabotage experiments with their frontier model Astra which they found evaded a monitor which is another AI that only saw its reasoning. Successful evasion however fell close to zero when that monitor could see its chain of thought and what it was actually doing. These are deliberately constructed tests but they show that what a model says it's doing shouldn't be taken at face value and should actually be checked against what it actually is doing.
There's encouraging results meanwhile on training methods, training models to behave better in unfamiliar situations. Anthropic did this by teaching values and ethical reasoning to their models which they found reduced blackmail in fictional tests from 65% to 19% without training directly on blackmail scenarios.
Overall, the industry has started to call for a slowdown with several frontier lab leaders now supporting the ability to slow the development of more powerful AI. Although their proposals individually differ in the detail, staff at these companies have signed public calls for slowdowns although the signatories represent a small minority of the overall workforces. The important question, however, is how do we assess what is required for a pause? How do we define what evidence should trigger it? And how should we determine when it's safe to restart?
There are several proposals towards this agenda. They include independent evaluators that need meaningful access, compute agreements that need verification, insurance proposals to try to give labs financial incentives to manage those shared risks. These are mechanisms that remain proposals and their usefulness will depend on whether they can actually detect the problems, establish responsibility and change behavior when it matters.
So that wraps up the sort of four key areas of the report and I'll now take you into the predictions where every year we put our expectations on the record and then come back and assess what happened. So last year we said that there would be a renewed lean-in from major labs into open source to align with the US government's policy towards American AI and we think that has happened, that a Chinese lab would regain a top spot at a major task benchmark and that's happened, the local opposition to data centers would really scale up and this has become a global issue not just an American one and we'll see the electoral ramifications rather that come out of this prediction in the next couple of weeks.
For the next 12 months, we expect that payment networks will start assigning rules assigning liability to purchases made by AI agents. That agents will start to improve continually and we can measure that by seeing failure rates on new tasks after months of customer work go down rapidly without upgrading the underlying model. We think that as consumers engage more with agents for financial research that a US regulator or exchange attributes an abnormal stock movement to retail AI agents that are placing these correlated orders, that cyber security will become even more important and that there will be an AI-led cyber attack that steals model weights from a lab.
And on this topic of research taste, that an autonomous AI team beats a human-led model research team with the same budget and time and compute as a result of choosing a better research agenda. And our boldest call would be that on the 10th anniversary of the State of AI Report we reach AGI which will be next year.
So with that, thank you very much for your attention. I invite you to read the AI report in all of its beauty and detail. It's around 300 slides. You can find it at stateof.ai. Tell us what you think. Write to us in any way you wish. And you can follow our ongoing analysis on Air Street Press. Thank you for watching and thank you for your contributions to advancing AI.
