51 Charts That Will Shape AI in 2026

0:00Today, we are looking at 51 charts to tell the story of artificial intelligence heading into next year. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:36And if you want to learn more about our recently released AI ROI benchmarking survey, you can find out all about that at aidbintel.com. Now, we are in the midst of end-of-year episodes, which is a combination, of course, of both looking back and looking forward. And this episode is all about the charts that sit right at that intersection. They are charts that tell us where AI is today and give us some idea of what we should be planning on heading into 2026. Given that there are 51 of these things, I am going to rip through them. So buckle up and let's talk about the 51 charts that explain AI in 2026.

1:10Quick note on the production of this. The charts were all sourced entirely by me. Part of my process for preparing this show is spending a ton of time on X slash Twitter and using those bookmarks heavily. And I have a folder where I actually keep these types of charts. So step one was just going back and looking at the charts that I thought were most reflective of the current moment and had something to say about the year we're heading into. The second part of the process was outlining a somewhat rough organization of which charts I wanted to include. From there, I turned it over to Claude, ChatGPT, and Gemini to see how they would organize it.

1:38I liked Opus 4.5 best, so we went with that with a few tweaks. And then I handed that and the charts off to GenSpark and Manus to put it all together. And while GenSpark looked much better, it made some really weird leaps in terms of how it was describing things and had some errors. So ultimately, we went with Manus, which was then exported to Google Drive for a final edit by me. Apologies for those of you who don't care about that. I just think a lot of you are also interested in the operator and production side of AI, so I like telling you how these things get put together. All right. As you can see, we've divided this into seven categories.

2:09Capabilities, infrastructure, markets, economics, vibe, coding, jobs, and politics. We kick off with capabilities. First chart comes from Open Router and is the reasoning versus non-reasoning token trends over time. You've probably seen this one a couple times now. Basically, at the beginning of 2025, reasoning models were not yet really a thing. OpenAI had announced O1 Preview back in September, and it had finally become available at the very end of December, but we were just starting to get our hands on these things. That would change dramatically over the course of the last year, and by November of 2025, reasoning tokens represented meaningfully over 50%.

2:41This has brought with it new capabilities, new use cases, and new ways of thinking about how we scale. Our next chart is the one that for much of this year held up the entire world, it felt like. This is the chart from Meter that measures the time horizon of software engineering tasks that different LLMs can complete at 50 and 80 % success rates. So the task duration here is not how long the model works for independently, it's how long in human equivalent time a task can complete. Coming into this year, Meter had shown a doubling of capability roughly every seven months, but it had started to inch up to closer to four months, and this year reified that four-month doubling time.

3:17In these charts, you can see the seven-month doubling line in green and the five-month doubling line in red, and you can see how at 50%, it hews really closely to the four-month line, and at 80%, it's mostly on the four-month line, with a few recent ones in between the four - and the seven-month line. Now, whether it's seven months or four months, the point is, capabilities have not plateaued. They continue to increase dramatically and quickly. We are also seeing major efficiency gains. This chart shows the performance efficiency of Gemini 3 Flash, which is better performing than Gemini 2.5 Pro, which was state-of-the-art just a few months ago, for around a third of the cost.

3:52Especially as we move into a world where production workloads are getting bigger and bigger and we are consuming more tokens, the fact that it's not just capabilities but also efficiency and costs that are improving is a big deal. Another measure of the efficiency gains came with 5.2's performance on the ARC AGI 1 exam. The ARC AGI benchmark folks noted that between a tweaked O3 model last year and GPT 5.2 this year, there was a 390 % efficiency gain in a single year. Now, what this all adds up to in terms of when we get AGI is kind of anyone's guess. As you can see from this chart, people are all over the place in terms of when they think we're actually going to get AGI.

4:30By the way, there's no common definition of AGI, and there are even plenty of folks out there who think that the term is getting more and more meaningless. One interesting note is that I think that if anything, people's timelines actually got moved back slightly heading into 2026 from where they were heading into 2025, despite all these capability gains. Andre Karpathy, in particular, in a big interview he did, might have single-handedly set back the timeline a couple of years. Now, as we look at these new releases, they are not just incremental. In many cases, they are solving key challenges.

4:59The charts we have here are for a long context test that basically tests how an LLM's performance degrades the more context you give it. With GPT-5-1, which is like a month old at this point, you can see the performance on this test went from around 85 % or 90 % at 8k tokens to a little under 50 % at 256k tokens, whereas with GBT52 thinking, it was at 100 % to start and stayed very close all the way to the end. This makes that context window actually usable in a way that it just wasn't before, which is extremely powerful and opens up new use cases. Still, AI capabilities, as much as they are evolving, are not evolving evenly.

5:36There are a bunch of different versions of this chart that have different size of spikes depending on how good you think AI is, but the idea here is that AI progress is not uniform. It is instead jagged, where a model can be superhuman at certain tasks and unbelievably incompetent at basic things that a kid could do. This jaggedness is a key facet of AI and is part of the challenge in implementing it well. Indeed, when it comes to what slows down AI, there are a set of three different bottlenecks. This chart comes from a recent essay by Professor Ethan Mollick, who organized it into capability bottlenecks, process bottlenecks, and verification bottlenecks.

6:09Keepability bottlenecks are the ones we think about most, the weaknesses in AI and that jagged performance. Process bottlenecks are the ones that we really started to reconcile with this year, the things that make it hard to overlay AI onto existing systems, particularly in the enterprise, and have it do what it's able to do. A third layer, which we talk about not very much, is a new category that is native and endemic to AI, which is verification bottlenecks. basically where humans become crucial to reviewing edge cases and ensuring final accuracy, which is often a whole new set of processes that humans need to be organized around.

6:41I feel like in some ways, the first profession to really reconcile with these verification bottlenecks is software engineering, which has seen such a shift over the course of this year in how AI and agentic coding supports what they do, but has created all these new challenges and shifted a lot of the work to that verification step. Regardless of the challenges, one thing that this year showed is a massive explosion of diversity in the model set. The major labs are putting out more different types of models that have different strengths and weaknesses and are optimized for different types of use cases.

7:09But Chinese labs have also exploded and become a major player when it comes to the choices that builders have access to. Next up, we move to infrastructure. Obviously, the big theme of this year, which will continue to dominate heading into next year, is the hyperscalers making just historically large capital investments into AI infrastructure in the form of data centers. This represents one of the largest coordinated technology investments in history, and something that the market has really had to reconcile and wrap its head around. The level of investment is why people are asking questions about whether the output of AI and the revenue that comes from that can possibly justify it.

7:43And yet, all of the big labs feel exactly the same, which is, as Mark Zuckerberg has articulated many times this year, it is a much greater risk to under-invest than to over-invest. Another expression that we got that tells the story of the moment in such a simple chart is the capital going into office construction versus data center construction. This was actually from a couple months ago, and I'm sure that it is actually fully flipped at this point. But starting in 2023, you see a shift where less capital is going into offices and more capital is going into data centers. And sometime in the middle to the late part of 2025, those lines actually overlapped, where we're now seeing more money spent on data center construction than on office construction.

8:21Now, one of the things that some people ask is how much all of this new infrastructure really matters. This chart shows that slower growth in compute could lead to substantial delays of possibly years in terms of certain capability milestones. Now, it's a whole separate question as to whether there are actually benefits to those delays. For example, if you ask Bernie Sanders, there absolutely are. But the point that this chart is trying to make is that there will be consequences to the speed of AI development if these labs don't have access to the compute that they're looking for. And this chart shows that as much as the labs have to service their existing customers, they are still heavily investing in the future.

8:56Now this is 2024, and I'd be interested to see an update for 2025. But in 2024, you can see OpenAI's R &D compute was 5 billion, as opposed to its inference compute, which was at about 2 billion. I'm not sure that they were able to maintain this ratio this year. Given the release of their images model, which was their most viral moment of the year, plus the release of Sora, plus just continued growth in their base usage, you have to think that servicing their existing customers started to compete with R &D a little bit more this year in ways that could be challenging heading into the future. Certainly there is scuttlebutt going around that some folks inside OpenAI aren't particularly happy about how that ratio looks right now.

9:32From there, we move into markets. The first chart to note is sort of the most obvious. The chatbot adoption is absolutely massive and faster than anything else we've ever seen before. I hardly need to spend a lot of time on this, but we now have two chatbots in ChatGPT and Gemini that are absolutely careening towards a billion active users, something that it took the previous fastest growing technologies five plus years at the very minimum to achieve. Still, if there was one chart that defined AI for markets this year, it's all of the various permutations of this circularity chart. This shows how much revenue and dealmaking flows between the major players like Microsoft OpenAI and Oracle.

10:09Now, to some, this chart is exhibit A in why AI is a house of cards. But of course, what this chart is missing is a visualization of the very significant and real revenue that is also coming in to help fuel this. Of course, the revenue isn't even close to the total scale of dealmaking right now, but it is growing faster than anything we've ever seen, and we really have barely started to scratch the surface on monetization. Still, this chart is as good a Rorschach test as we have for how you feel about the markets heading into the next year. Paired with this, we have a recent chart of OpenAI's estimated balance sheet that shows just how much external capital they're going to need to get to the point where they're actually profitable.

10:44Now, so far, it doesn't seem like they're going to have any problem accessing that capital. The most recent rumors are that they are raising tens of billions of dollars, if not up to $100 billion at an$830 billion valuation, suggesting that capital markets are very comfortable with what they're seeing from OpenAI and very willing to fund this party to keep going. Now, for those who are AI bears, one of the things that they are concerned about is the reduction in inference costs, where as models get good enough, AI products can actually run on much cheaper, simpler GPUs and computers. They worry that if that happens, all of this massive investment in complex architectures doesn't really make sense anymore.

11:20One investor I saw called this chart the most important and misunderstood chart in AI. Now, it should be noted that not everyone agrees with this and there are lots of counter-arguments, but if we're trying to understand where the market's head is at heading into next year, this is a key consideration. Now we move into some model competition. Anthropic, by any measure, had a very good year. Their market share in coding was massive, and that dragged their market share across the enterprise up as well. According to Menlo Ventures estimates, they now claim 40 % of the enterprise market ahead of OpenAI.

11:50Google also saw a lot of growth in enterprise this year as well. And as fast as OpenAI's revenue has been growing, Anthropics has been growing even faster. They went from a billion dollars annualized at the beginning of this year and will end the year at somewhere around$8 or$9 billion, it seems. OpenAI started at around$4 billion and will end the year at$13 or$14 billion, which is absolutely incredible, but still growing more slowly than Anthropic, at least for the moment. Still, if you are just taking a step back and don't have a particular horse in this race, the thing to note is just the incredible pace of revenue growth for both of these companies, which has to be bullish for their ability to actually make good on all these big deals that they're signing over the course of the next five years.

12:30Another key story of 2025 that sets up the battle for 2026 is, of course, the massive resurgence of Google. You can see this moment around the launch of GPT-5, which also happened to be the launch of the first version of Nano Banana, that Gemini really starts to take off. The release of Gemini 3 has also purportedly even increased this competition, and Gemini is absolutely surging heading into next year. Another way that you can see this expressed is in the betting markets, where the likelihood that Alphabet is the largest company by the end of next June has increased significantly. NVIDIA held a commanding percentage of that at the beginning of the year, and Alphabet is now creeping up on them.

13:05Now the other way to see the sentiment shift between OpenAI and Alphabet is in the basket of correlated stocks. Bloomberg and Morgan Stanley put together a basket of stocks that are exposed to Alphabet and a basket of stocks that are exposed to OpenAI and showed how around November they started diverging, with the Alphabet-exposed stocks continuing to rise and the OpenAI-exposed stocks taking a bit of a hit. Now, this doesn't mean anything fundamental. It just means a shift in what markets believe, but it's a pretty clear and dramatic signal of where things are. Still, if you are one who is just looking at the performance of these different models, I think one of the most powerful charts is this one, which shows that no one stays on top for long.

13:43OpenAI introduces the world's most powerful model, followed by Anthropic, who introduces the world's most powerful model, followed by Gemini, who introduces the world's most powerful model, followed by Grok, who introduces the world's most powerful model. On and on forever, infinity. Just from a performance capability standpoint, this is absolutely the chart that best shows what we are going to experience throughout 2026, I strongly believe. However, if we're talking about the competition between the labs, we do have to give China its due. This chart shows the massive increase in China as a share of open source tokens.

14:14At the beginning of the year, it was all meta and mistral, with almost nothing coming from China. By the end of the year, it was something like 80 % Chinese models. They are a factor heading into next year and will be a bigger factor throughout 2026.

14:57change inside your enterprise and as case studies, expert panels, and a lot more practical goodness that I hope will be extremely valuable for you as the listener. Search You Can With AI on Apple, Spotify, or YouTube and subscribe today. Today's episode is brought to you by my company, Superintelligent. Superintelligent is an AI planning platform. And right now, as we head into 2026, the big theme that we're seeing among the enterprises that we work with is a real determination to make 2026 a year of scaled AI deployments, not just more pilots and experiments. However, many of our partners are stuck on some AI plateau.

15:36It might be issues of governance. It might be issues of data readiness. It might be issues of process mapping. Whatever the case, we're launching a new type of assessment called Plateau Breaker that, as you probably guessed from that name, is about breaking through AI plateaus. We'll deploy voice agents to collect information and diagnose what the real bottlenecks are that are keeping you on that plateau. From there, we put together a blueprint and an action plan that helps you move right through that plateau into full-scale deployment and real ROI. If you're interested in learning more about Plateau Breaker, shoot us a note, contact at bsuper.ai with plateau in the subject line.

16:12AI changes fast. You need a partner built for the long game. Robots and pencils work side by side with organizations to turn AI ambition into real human impact. As an AWS-certified partner, they modernize infrastructure, design cloud-native systems, and apply AI to create business value. And their partnerships don't end at launch. As AI changes, robots and pencils stays by your side, so you keep pace. The difference is close partnership that builds value and compounds over time. Plus, with delivery centers across the US, Canada, Europe, and Latin America, clients get local expertise and global scale.

16:45For AI that delivers progress, not promises, visit robotsandpencils.com slash AI Daily Brief.

17:18work autonomously while providing a guide for the final 20 % of human development work required to complete the sprint. Public companies are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding pilot of choice to bring an AI-native SDLC into their org. Visit blitzy.com and press get a demo to learn how Blitzy transforms your SDLC from AI-assisted to AI-native.

17:45Next up, we move over to economics. by which we mean the impact of AI. We've talked already a little bit about how the cost of using AI has fallen precipitously. What's interesting, of course, is that Jayvon's paradox has fully locked in and the total amount that enterprises are spending on AI has gone nothing but up. In fact, price decreases are fueling usage growth as price decreases unlock new types of use cases that were uneconomical before. This has led to enterprise AI being the fastest scaling software category in history. According to Menlo's estimates, it now captures 6 % of the$300 billion global SaaS market.

18:18Now, it should be noted that I don't think that the total addressable market for AI is the$300 billion global SaaS market. I think it is multiples larger than that. But still, even in historically slow-moving and lumbering enterprise adoption, things are moving very, very quickly. What's more, despite some rumors to the contrary, companies are actually seeing measurable ROI from AI even now. In a Wharton study of something like 800 executives, around 75 % reported positive ROI from their AI investments. On our AI ROI benchmarking study, we found 82 % saw current positive ROI. Of the remainder, by the way, only 5.5 % were currently at negative ROI, and even those that were at negative ROI anticipated that becoming ROI positive by next year.

19:02In fact, 96 % of overall respondents anticipated positive ROI within the next 12 months. And across the entire sample, 37%, almost 4 in 10 were already seeing high ROI from their use cases, reporting either significant or transformational impact. In another chart from our ROI study, we also found a correlation between how diverse an organization's use of AI was and how much benefit they got. We organized impact into eight different benefit categories, and we found that when an organization had use cases with just one benefit type, their ROI was lower than organizations that had four different benefit types, who were lower than organizations who had all eight different benefit types, in a pretty significant way.

19:41Three was a measure of modest ROI, four was a measure of significant ROI, organizations that had one benefit type were just over the edge of modest at 3.13, and organizations that had eight benefit types were starting to creep on significant at 3.65. What about the idea of 2025 as the year of agents? Well, it turns out in practice, agents remain nascent. In that same Menlo study, they found about 10 times as much money being spent on assistants and co-pilots than were being spent on agents so far. In our ROI study, we found something similar. We divided use cases into three categories, assisted, automated, and agentic, and found 57 % were in the assisted category, 30 % were in the automation category, where the AI was managing a discrete workflow, and 14 % were in that agentic category of autonomous work execution.

20:25One more in the economics section that is under-discussed but I think is going to be extremely important next year, I anticipate that we're going to start seeing a much deeper integration of ads into the AI landscape. There are a variety of reasons for that, but they are not just about the business model needs of the labs, although that's a part of it. LLMs also appear to be a really good platform for ads. Check out these recent numbers from SimilarWeb. They looked at the average minutes spent on site after referral, the average page views on site after referral, and the average conversion rate of referrals if the source was ChatGPT versus Google.

20:57And in each case, ChatGPT absolutely thwumped Google. The average minutes spent on site were three times higher from referrals from ChatGPT. Average page views were 25 % higher, and the conversion rate jumped from 5 % to 7%. Basically, people who are finding sites through LLMs are more high-intent, it appears, than your average Google browser, which again makes it a great place for sponsored links and ads. Next up, let's look at a category that was extremely important to 2025, vibe coding. Our first chart is just vibe coding grew really fast. We saw multiple companies surge into the nine figures of revenue.

21:31Some like Cursor creep up on a billion dollars in ARR, and some like Cloud Code blow past that. The combination of meaningful token costs and high consumption, plus implications for other use cases in LLMs, made coding-related performance become the industry's number one priority. As we head into next year, however, especially engineering organizations are trying to figure out how to redesign themselves around AI coding. A chart that I've seen from Swix, Sean Wang, and others is this semi-async valley of death chart. It looks at agent autonomy and measures the experience or observed productivity at various autonomy levels.

22:04On the one end of the spectrum, when coding agents are extremely responsive in very fast order, they can be extremely valuable for deep work focus on the hardest problems. On the other end of the spectrum, when they have a lot of autonomy, they can be great for simpler texts that are handled in the background. The challenge is in the middle range, where as the chart puts it, it's not enough to delegate and it's not fun to wait. Now, different organizations are handling this differently, and I think SWIX even has some questions around whether this is exactly the right way to think about things.

22:30But for our purposes, this chart represents not just the semi-async valley of death, but just the broader set of questions that engineering organizations are going through heading into next year to redesign themselves around AI coding. The reason that this matters outside of software engineering is that I believe that they are the first department that will fully reorganize themselves around AI capabilities, and in so doing set a template that other departments and functions can start to follow. Another interesting chart showing the impact of vibe coding, after a long period of being flat, in 2024 and 2025, we started to see the number of apps and games released to the App Store going back up.