The AI Daily Brief: Artificial Intelligence News and Analysis - Grok 4.6 Shows How Fast Your AI Options Are Expanding

Starting point is 00:00:00 A year ago, if you were talking about frontier models, pretty much you were referring to a model from one of either OpenAI, Anthropic, or Google. By a couple of months ago, you were probably referring to a model just from either OpenAI or Anthropic. Now, however, things have changed. Over the past couple of months, any conversation about model performance has to include a recognition of Chinese open weight models that are pushing the frontier of both efficiency and cost, and as of this week, space is.

Starting point is 00:00:30 SpaceX AI's GROC is back in the conversation. The just-release GROC 4.6 is putting up benchmark numbers that put it in the category of a GAPT 5.6 or a Fable 5, and doing so at a fraction of the cost. Although, of course, as we know, AI in the benchmarks tends to be very different than AI in the real world. After some initial testing, while users are not ready to declare GROC 46 a Fable or GAPE class model yet, they are ready to argue fairly definitively that GROC and SpaceX AI are back in the race. The AI Daily Brief is a daily podcast and video about the most important news and

Starting point is 00:01:04 discussions in AI. All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Blitzy, Hyperagent, and Harbor. To get an ad-free version of the show, go to Patreon.com slash AI Daily Brief, or you can subscribe on Apple Podcasts. To learn more about sponsoring the show, send us a note at sponsors at AIdailydief.aI. And one other thing you should check out on AIDailybrief. As you know, we've recently updated the website, so now each ever episode has a full companion edition that includes all the key numbers, all the key quotes, all the key themes, each organized into different shareable cards that make it easy for you to find exactly the part that you want to share with someone else. We have now added an archive

Starting point is 00:01:48 as well to hopefully make it easier to find previous episodes about a particular theme. It's organized on both an episode and a card basis, and we'll be continuing to try to improve it as time goes on. Now, at that out of the way, let's get to the headlines which are all about big money and into the change in the model landscape that's the subject of our main episode. Welcome back to the AI Daily Brief Headlines edition, all the daily AI news you need in around five minutes, and the theme of today is big money. Cognition is seeking another funding round on the back of booming coding agent demand. Bloomberg reports that Cognition is in early talks with investors for new funding at a valuation of $40 billion. Cognition closed their last round just three months ago,

Starting point is 00:02:28 raising a billion dollars at a $26 billion valuation. For those doing the quick math, that means that the company's valuation would be up almost 50% in a quarter. and the revenue figures seem to back it up. Sources familiar with the fundraising efforts said Cognition has doubled their revenue run rate to a billion dollars since they were last seeking funding. One source said that Cognition is seeking a billion dollars in this round, giving themselves a substantial increase in resources to address the current agent boom. The numbers also imply that the premium attached to coding agents is growing among venture investors.

Starting point is 00:02:57 Cursor is one of the closest comps, and their last fundraising round in March saw them seeking a $50 billion valuation on $2 billion in annualized revenue. That round, of course, ended with SpaceX acquiring the company in a $60 billion all-stock deal. And honestly, if Cognition has the ability to price their round at $40 billion, the SpaceX deal could start to look like a bargain. Many think that the path that Cursor took with SpaceX feels inevitable for Cognition as well, writes Richard Wu, I wouldn't be surprised if within the next six to 12 months we see one of the hyperscalers preempt Cognition and offered to acquire them for $60 to $100 billion in stock.

Starting point is 00:03:30 Given the success with SpaceX acquiring Cursor, the boards of these companies will put pressure on them to make a move. Jeff Wu says, Google should buy Cognition for $200 billion and make Scott Wu CEO. Sond deep from Cognition responded, we aren't selling. Also, 200 billion, the stock would move three times that and after hours alone. Next up, we have Lovable, who announced their $400 million series C round at a $13.3 billion valuation. What's interesting is that you can clearly see how lovable is evolving just in the way that they describe themselves in their fundraising announcement. In short, Lovable feels to me to be inching farther away from Claude Code and closer towards something like Shopify. They write, Lovable is building the software creation platform that gives

Starting point is 00:04:13 those closest to a problem the power to solve it, a generational opportunity that spans billions of people all over the world. For most people, turning an idea into software once required so much capital, technical fluency, and time that many ideas never came to life. Lovable's first chapter was about changing that. Since our series B in December 2025, we've been building features people need to reach customers, manage day-to-day operations, and run software securely. For many builders, the product they create with Lovable is becoming the business itself. User survey data shows us that nearly 8 and 10 are building a business or side project they hope to monetize, and more than one-third of those are already earning revenue.

Starting point is 00:04:49 In CEO, Antoine Oseco's post, he absolutely emphasizes the same idea, saying that Lovable will create, quote, the most intuitive platform to build and run a business. If you are looking for a place to see the intersection of where what we'll be a way we're was once called vibe coding meets the actual transformation of small and digital businesses, look no further than lovable. Now, moving into public markets, business is booming for the neoclouds as AI demand continues to rise. This week saw Corweave and Nebius report earnings, both vastly outstripping analyst expectations. On Tuesday night, Corrieve reported that revenue had doubled over the past year to reach $2.6 billion for the quarter. At the same time,

Starting point is 00:05:27 cash burn also doubled now running at $5.7 billion per quarter. Still, the big story for investors, was a line out the door for compute. Corwee reported a $104 billion backlog in demand. In the footnotes, they added that the backlog had grown by $25 billion since they closed their books at the end of June. The story was the same for Nebius, who reported on Wednesday. They recorded 454% revenue growth over the past year to reach $582 million. Their cash burn is also escalating rapidly. But like Corweave, Nebius has endless demand, with CEO Arcade Velos telling investors, demand for what we are building continues to be enormous. We could sell today our entire 2027 capacity if we wanted.

Starting point is 00:06:05 Supply is in fact so tight that Nebius is seeing huge profits on their available capacity. Earnings per share beat analysts forecast by 83%. And Volos told investors that their auctions for Blackwell compute, which began in Q2, cleared at 15% above their previous record price for Hopper compute. Markets rewarded both stocks with CoreWeave up 19% since reporting and Nebius gaining a staggering 34%. Analysts believe that neoclods are some of the best indicators of marginal demand for AI, as they service the overflow from the hypers and even during a quarter when token austerity came into vogue,

Starting point is 00:06:36 demand is showing no signs of slowing. Meanwhile, the infrastructure boom also is coming to China as Tencent has tripled their CAPEX. Tencent reported that they spent $7.8 billion on AI infrastructure in the past quarter, boosting their training and inference fleet. Now, of course, that spending is still relatively modest compared to the U.S. hyperscalers, where meta had the slowest CAPX in their group and spent $31.9 billion in Q2. Still, there's a pretty clear attitude shift as the Chinese tech giants commit to scaling up their

Starting point is 00:07:03 data center construction. During an earnings call on Wednesday, chief strategy officer James Mitchell said, we're allocating a very substantial portion of new compute to our own models and applications. The company's revenue is growing at 11%, but free cash flow has dipped into the negative, with incremental earnings going toward infrastructure. Tencent President Martin Lau said that Tencent could monetize their compute by selling to outside customers if they wanted to, but for now they're prioritizing their own needs. Basically, just like model training, it seems like China's AI buildout and the narratives around it are three to six months behind the U.S. as well. It is uncanny how closely this is following the

Starting point is 00:07:37 narratives from the U.S. in Q1. Hyper-scalers flip to negative free cash flow, folks like Zuckerberg appeasing the market by telling them that he could sell his compute but he doesn't want to. I'm not sure I think that U.S. market participants have fully accounted for a Chinese Kappex boom and what it does for the larger global investment environment. Meanwhile, Samsung is seeing incredible efficiency gains from their use of AI in chip design. According to reports from a Korean outlet, the first three months of integrating Claude Code into the software stack have been an outstanding success. Development personnel have been able to cut down the time to complete complex tasks like

Starting point is 00:08:08 system-on-chip verification from three months to two days. In one example, a second-year engineer was able to complete a month-long task in a single day. Now, of course, this report doesn't claim that Claude Code produced efficiency gains throughout the entire chip design process, but it does seem like an interesting example of the jagged frontier of AI adoption in the enterprise. Claude was able to make highly customized jobs more efficient and able to help a junior employee contribute way beyond their expertise. Lastly today, some reported updates coming to the Trump administration's model testing framework. Last Tuesday, leading frontier labs were briefed on that framework, although the rest of us

Starting point is 00:08:42 didn't get to learn all the details. It was reported that the policy would cover only state-of-the-art models, although we didn't know how exactly that was defined. What we did hear with a fair degree of confidence was that the policy wouldn't cover open models. Open source advocates were relieved at that decision, but there was also a contingent of China Hawks who believed that this would leave a gap. On Wednesday, Wired reported that the administration has changed their mind. An official said that the White House is expected to expand the policy to cover open models in the coming months. The policy they added is aimed at ensuring that as soon as open models reach the same capabilities as Mythos or GPT5.6, they're added to the safety testing framework. White House officials said the administration had hoped the

Starting point is 00:09:18 policy would be one and done, but the exponential development of model capabilities had forced them to iterate. When it comes to the inclusion of open models, the thinking is that leaving them out of the framework could actually create a two-tiered system that would be negative for those open models. Specifically, officials are concerned that the framework could be viewed as a stamp of approval, leaving enterprises hesitant to use open models if they don't receive the same testing. The concern, then, is that leaving open models out could actually disincentivize U.S. labs from developing those open models. Adding some evidence to the idea that the government is pro-U.S. open models, Treasury Secretary Scott Besson actually retweeted Mark Zuckerberg this week,

Starting point is 00:09:53 saying, we welcome meta's release of Muse Glimmer, another win for American innovation. Sustaining U.S. leadership in AI means advancing both open and closed weight models, ensuring the future is built on trusted foundations. Overall, it's still pretty clear that there's a lot of consternation around the administration policy. President Trump himself is reportedly insisting on keeping the framework voluntary as he believes formal regulation will help China catch up, but by the same token, the safety-focused faction of the administration also isn't satisfied. and are reportedly still pushing for a more formal arrangement. Who the heck knows how that's all going to turn out?

Starting point is 00:10:23 But still, this is a perfect segue to a broader discussion of the state of models. So for now, that's going to do it for today's headlines. Next up, the main episode. Hello, everyone. One big change around AI is we've shifted our thinking from how we rank our pages to how do we become the source that AI trusts enough to answer with. At KPMG, they're seeing this firsthand. AI-generated results now surface answers directly, often without a single click.

Starting point is 00:10:51 That's why they are increasingly focused on generative engine optimization or GEO, structuring content so AI systems can retrieve it, understand it, and cite it as trusted authority. This is not just an SEO evolution, but a visibility mandate. And indeed, the GEO mandate from KPMG is simple. If AI is shaping decisions, your expertise needs to show up inside the answer. Read all about it at KPMG.com slash us slash geo. Again, that is KPMG.com slash US slash geo. Blitzy's deep code-based understanding unlocks the thing every roadmap owner cares about,

Starting point is 00:11:25 shipping new features. Here's the truth about building inside a massive enterprise codebase. Writing code was never the bottleneck. Context is. Which system does this touch? Which contracts can't break? Which standards apply? Blitzy already knows because it reverse-engineered your entire codebase into a dynamic knowledge graph before feature work began. With that complete picture, Blitzy builds features end-to-end. Architecture, APIs, UI, and tests all validated against your existing systems. One Blitzie customer built. an AI native application from scratch with 100% autonomous completion, saving over 2,700 engineering hours, features that respect your codebase instead of fighting it. Stop letting your backlog grow faster

Starting point is 00:12:01 than your team. Accelerate your roadmap at blitzie.com. That's BLYt, tzky.com. This episode of the AI Daily Brief is brought to you by HyperAgent, where you run fleets of agents your team can manage together. New users get $1,000 in inference. Forget local agents and chat workflows waiting on your laptop to be prompted. Hyperagent deploys all. always-on agents in the cloud, doing real work across the tools your team already uses. Marketing's agent turns competitor moves into landing pages. Sales as agent enriches leads, drafts emails, and updates the CRM. Ops agent chases the paperwork and tracks the budget. Every agent has access to shared context and follows your rules about scope and approvals.

Starting point is 00:12:38 It's time you add agents that feel like teammates. Hire yours at Hyperagent built by the team at Airtable. Claim your $1,000 in inference at hyperagent.com slash AI Daily Brief. Every episode, I talk about the competition between OpenAI, Anthropic, SpaceX, AI, Google, and Meta. And if you've been listening for a while, you might have a favorite. Maybe you think Open AI and Anthropic can stay ahead, or perhaps meta's open source strategy can win out. Whatever your view, every AI lab creates a different investment opportunity. Harbor Capital Advisors AI Lab Ecosystem ETF Suite lets you invest in the ecosystem behind the AI lab you believe in.

Starting point is 00:13:13 Search Harbor AI Lab ecosystem ETFs wherever you invest or follow at Harbor Capital on X to learn more. Visit Harbor Capital.com for a prospectus containing investment objectives, risks, fees, expenses, and other important information. Read and consider it carefully before investing. Risks include principal loss and artificial intelligence-related risks.

Starting point is 00:13:29 Harbor UTS are distributed by Foreside Fund Services LLC. Harbor is not affiliated with AI Daily Brief and the funds are not affiliated with sponsored by or endorsed by any AI lab. This is a paid advertisement and not personalized investment advice. investing involves risk, including possible loss of principle. Welcome back to the AI Daily Brief. The big news that we are covering today is the release of GROC 4.6,

Starting point is 00:13:52 which is getting some pretty good reviews out of the gate. But what's interesting to me is not just the model itself, but what it says about the state of the AI race and how that's changing. Now, the version of the AI race story that I am concerned with mostly here, of course, is the one that has to do not just with the achievement of some ill-defined far-flung goal like AGI or ASI, but the practical impacts on where different labs are for what we get to do with AI at home and in our companies. I think CNBC's Dear Drabossa summed up the vibes when she tweeted yesterday, what a difference a year makes.

Starting point is 00:14:23 A year ago, Frontier basically meant the big three U.S. closed labs, open AI, Anthropic, and Google. Now a credible list includes XAI and multiple Chinese and open weight labs. And while we'll get into the implications for the leading labs in a minute, I think Nathan Lambert also gets at another part of the sentiment when he writes, The vibes shifting from Anthropic is so far ahead to model competition back to all-time highs took like four weeks. So let's talk Grok 4-6 first. The release appears to put SpaceX AI squarely back in the frontier model competition.

Starting point is 00:14:55 Regular listeners will know that I take any release benchmarks, not just with a grain of salt, but with an entire bowlful, but still, GROC's reported benchmarks are pretty hard to ignore. On GDP Val, which is the measure of how agentic AI performs on economic, economically valuable tasks, SpaceX AI claims to have overtaken both GPT-5-6-Sole-Sole and Fable 5 by a small amount, yes, but taken over them nonetheless. Coding performance is improving as well, with GROC scoring right between but in the range of 5-6-Sole and Fable 5 on CERC-BENCH, being a few points behind on both Deep Sway and Terminal Bench.

Starting point is 00:15:26 On the overall artificial analysis intelligence index, GROC 4.6 jumped a full five points from GROC 4.5's 56 to achieve an overall score of 61. that puts it ahead of Kimmy K3, tied with 56 soul, and just a point or two behind Fable 5 and Opus 5. What that means is that if this was a new model from either Anthropic or OpenAI, we'd probably be talking about how it's not quite state of the art and didn't push the frontier forward. But for SpaceX AI, who many had written out of the model race until fairly recently, this is a huge achievement, summed up by the broad sense that you can see across AI circles, that we once again have three frontier labs in the race.

Starting point is 00:16:06 Also, while SpaceX AI has massively improved GROC's performance from 4.5, it seems like they're still working from the same base model as GROC 4.5. Pricing remains the same at $2 per million input tokens and $6 per million output tokens, making it 60% cheaper than GPD 56 sole on a per token basis. Of course, as we know, comparing tokens to tokens is a seductive but ultimately fraud exercise given the massive differences in how many tokens different models might use to solve the same problem. But once again, artificial analysis's testing found that the model is pretty token efficient as well. It completed the benchmark run at 84 cents per task, putting it in line with Kimmy K3 and making it 32% cheaper than GPT-56 sole and 73% cheaper than Fable.

Starting point is 00:16:50 writes investor Davin Baker, absolute Pareto dominance for Grock and Cursor even after the OpenAI price cuts. Now, in terms of reactions for the community, for many folks, it was just gobsmacked at the achievement overall. Vitorio writes, So they just caught up in three years? How does Elon do it? Ben Davis writes, GROC 4.6 feels very good on first tests, very fast and capable and cheap, but time will tell as always. The cursor and SpaceX AI come back is glorious to watch. On Martin Casado from A16Z's highly technical tests, he found that it was strong. Pavel Hearn writes, tried Grock 4.6 on my bug bench an hour after release,

Starting point is 00:17:28 105 hidden bugs and two real repos judged blind. His conclusion? Looks like it may be my new default model, the best combination of time, value, and cost. And yet some folks did not have that same experience. Mejo-Mohan writes, Grogh 4.6 is not as good as GAPT-56 sole in my 30 minutes of usage. It does incomplete work, not incorrect, just incomplete.

Starting point is 00:17:50 Maybe it's the Grock harness? Justin Schroeder writes, early vibes on GROC 4.6 are not great. It's fast and is willing to do security work. I've already seen multiple instances where it makes dangerous mistakes and later tries to cover up poor decisions. It even gets defensive. Unfortunately, we cannot trust it. Entrepreneur Timmy McKeehan writes, Grogh 4.6 is one of the most oddly behaved models I've seen so far. It produces many times the output tokens compared to Terra or any similar intelligence model.

Starting point is 00:18:17 It is cheap and fast, but takes everything extremely seriously and always investigates unclear information. It values completeness above everything, including economics. The model seems to be designed to be economically viable, but acts differently. Now, when someone tried to clarify, if this is a positive or a negative sign, Timmy kind of shrugged and said, probably positive? Benjamin DeKracker tried to sum up, lots of people acting like GROC 4.6 just beat Anthropic and OpenAI when really it didn't. The Grock 4.6 numbers show that XAI is not out of the race, but also not at the top. It's in the middle topish against models that the competition is already getting ready to update. It shows that Grock still has a pulse, which is a good, but different thing.

Starting point is 00:18:55 He continues, or in sports terms, they advance past a critical wildcard game into the playoffs, but are mid-rank against tough competition. They prove they can still hang, not yet winning everything. And by the way, he clarified, this is not a slight against GROC 4.6, which looks solid, just a read of the actual rankings and situation. Now, of course, what Benjamin is referring to is the fact that 4-6 is being compared against GBT 5.6 and Fable 5, when both of those models are at this point several months old, and pretty much the only reason we don't have updates of them is that we're now past the threshold

Starting point is 00:19:27 where the U.S. government is going to be involved in every big new model release, and so state-of-the-art for us is very different from state-of-the-art at those top labs. However, it sounds like GROC 4.6 is itself just a waypoint. Elon Musk tweeted, GROC 4.7 is significantly better than 4-6 and should be ready in three to four weeks. Initial training is complete, and now we're adding a massive amount of SpaceX company data in supplemental training. This will be something special.

Starting point is 00:19:52 In another tweet, he said, GROC 4.7 will exceed all current models. That said, Anthropic is a great company and will probably release improved models soon. However, the SpaceX training corpus is so awesome and unique that I would be shocked if any model is better at real-world engineering than 4.7. Capturing the zeitgeist of credulity around these claims,

Starting point is 00:20:11 Chubby shared both those posts and said, I'm taking this seriously now. Grogh 4.6 was the leap I've been hoping for. If the 10-T model is still to come, then Elon's words can be taken seriously. it really could become the best model in general. Although, of course, Anthropic already has Fable 5.5-5 ready and just waiting to be released, that much is clear.

Starting point is 00:20:30 Nevertheless, the next few weeks will be exciting, and X-AI has shown just how much potential they possess. Leo at Synthwave DD adds, X-AI have made an incredible comeback. From the days of GROC 4 to 4-3, where they were trailing the frontier by far, they're now arguably the third best lab in the world, behind only Anthropic and Open AI. So where does this leave the rest of the field? Well, first of all, there's Google, the company that many feel, Anthropic has no overtaken as the definitive third place when it comes to state-of-the-art models.

Starting point is 00:20:59 After last week's departure of DeepMind CEO Demis Hesabas and longtime product leader Jeff Dean, many are basically counting Google completely out of the frontier AI race. The counterpoint, however, is that it appears that co-founder Sergei Bryn is back in the picture to spur a comeback for Gemini. Reuters reported that Brin has become a key cheerleader for Google's AI team in recent months, encouraging AI engineers to catch up in the AI race. He reportedly addressed a town hall after the release of Mythos, telling engineers it's time for Google to play catch up.

Starting point is 00:21:29 Sergei had, of course, been out of the picture for several years after stepping down his president in 2019. However, he returned to frequent work at Google in 2023 and stepped into his involvement with the AI team in 2024, just as they were getting back on track ahead of the release of Gemini 2. During last week's news cycle, we had already heard that Google was relocating AI training out of the Deep Mind office in London

Starting point is 00:21:49 and back to the main campus and Mountain View. That relocation would conveniently allow Bryn to play a more active role working day-to-day with key researchers. And of course, given what else we've heard about internal Google politics, one of the big benefits to having Sergey fully engaged is that presumably he's one of the few people that could effortlessly cut through that bureaucracy to get things done at Google.

Starting point is 00:22:10 According to the Reuters report that came out on Wednesday, that has already begun. Reuters writes, Bryn has used the implicit power he holds as Google's co-founder to push resource allocation towards specific areas such as recursive self-improvement. And to some, this is a good enough reason all on its own to not count Google out. Nick the CS guy from Google writes, don't mess with Sergey and definitely don't underestimate what he can do.

Starting point is 00:22:32 Still others think that Google is just temperamentally ill-suited to this particular race. Computer science professor Pedro Domingos writes, Hey, Sundar, getting deep mind to be an LLM lab is trying to shove a square peg into a round hole. You're destroying them and you'll still lose the race. let them focus on AI beyond LLMs, which is what they're good at, and create a nimble new lab to run the LLM race. Now, when it comes to what models we can expect next, I think at this point broad sentiment is that it would not be enough to recapture momentum by releasing a competent Gemini 3.5 Pro at this point. We're already a couple months behind when we expected to get it, and just catching up, I think, would be seen as a failure. According to Leo and some other leakers I've seen, the reports are that teams are instead shifting to work on these scaled-up Gemini 4, which while risky,

Starting point is 00:23:16 I think does make sense in context. Now, as Deirdre pointed out in that tweet at the top of this show, the top model labs question now has to necessarily include a bunch of entrance from China. And interestingly, just a few hours after GROC 4.6 launched, we got a significant leak out of China. Specifically, we got the benchmarks for the updated version of Deepseek V4 Pro, and they appear on paper at least to be very competitive. For example, these leaked benchmarks claim that the forthcoming model scored 87.9% on Terminal

Starting point is 00:23:44 Bench 2.1, putting it just zero, 0.1% behind Fable and 1.1% behind GPD 56 Seoul. It also claims to beat Fable by 0.2% on CyberGim, the main cybersecurity benchmark benchmark. Now, as always, there's the risk that this is just benchmark maxing and actual performance will feel a little flat. And unfortunately, almost as soon as these leaks started appearing, other information came out, suggesting that the model was more significantly behind than the benchmarks would have it seem. artificial analysis's benchmark run was pretty disappointing, with V4 Pro scoring just 53. That's only one point ahead of V4 Flash and trails behind Kimmy K3 and Muse Spark 1.2.

Starting point is 00:24:22 On the plet side, the model is pretty cheap, even after Deep Seek delivered a substantial price increase this morning, at a buck 32 per million input and $3.96 per million output, it's about one-twelfth the price of fable and slightly cheaper than Muse Spark. And people's first impressions also aren't that great. Lucky Faraday writes, Deepseek V4 Pro is benched max slop. I had high hopes for this model, but it's complete trash. This was supposed to be a fable-level model, and it can't even make a simple Minecraft clone.

Starting point is 00:24:48 Even DeepSeek V4 Flash did a better job. I know a Minecraft clone isn't a good test for a model, but come on, this is complete nonsense. And before the don't compare a less than $1 output model to Frontier model replies, they are the ones comparing themselves to the frontier, not me. Still, others pointed out that when we're discussing models in the second half of 2026, it is less about raw performance alone and more about where they fit in the model stack. Dax from OpenCode says DeepSeek is insanely good at inference, using about two times less GPU time. And Augustin LeBron writes,

Starting point is 00:25:19 I'm sure Kimmy K3 and Grok 4.6 and Deepseek V4 Pro are benched max more than Fable and GPT, but it doesn't matter. These models are an order of magnitude cheaper. As the frontier proceeds, fewer and fewer people need the bleeding edge and need it less often. And at first glance, Ramp's latest AI index seems, seems to provide some evidence of that. Ramps lead economist Ara Karazian writes, New from Ramp AI Index, disappointing adoption of Fable 5.

Starting point is 00:25:45 We've heard several reasons from businesses, mainly Fable 5 is just too expensive. A model so powerful it was briefly banned, and yet businesses don't think it's worth the price. Specifically, Ramp found that Fable 5 has made up only 6% of tokens that businesses purchased from Anthropic, and represented only $11.4% of dollars spent on anthropic models.

Starting point is 00:26:05 For comparison, they write, OpenAI's GPT-56 sole comprises 25% of OpenAI tokens and 23% of spend. In fact, they say Fable 5 is less popular with businesses than GPT 5.6 overall. Ramp argues that, quote, With Fable 5, we found a new upper bound to how much businesses are willing to spend on AI. Here, more performance is not worth the price tag. To encourage business adoption of the latest models, the labs will need to prove performance beyond even what Fable 5 is able to achieve

Starting point is 00:26:31 and simultaneously ensure that competitors aren't able to come reasonably close. That seems increasingly out of reach, especially as open source models catch up to being only a few months behind. However, I think that story is much less clear than they're letting on. First of all, as Simon Smith points out, ramp data overall suffers from selection bias and this data suffers from it even more. This data comes from their token and spend management product, meaning users are predisposed to focus on cost control. Fable simply isn't cost-effective for most tasks. In other words, this is an extremely enfranchised set of users who are specific. specifically using this in a product that is designed to manage spend and optimize spend away

Starting point is 00:27:10 from models that are more powerful than you need, rather than being a general assessment across a wide cross-section of businesses and business use cases. Still, to me, that isn't even the most damning thing, as perhaps one could argue that those companies in that type of spend management are a leading indicator of where others will get. I think the bigger and more obvious issue is that Fable 5 still comes with a 30-day data retention policy, and most businesses aren't willing to touch that with a 39-5-foot poll. Indeed, Aura actually came back to Twitter and retweeted himself to add this incredibly important detail saying, a lot of replies from employees who say they aren't allowed to use Fable because Anthropic

Starting point is 00:27:45 is required to retain prompts for 30 days for U.S. government safety checks. Look, it is absolutely the case that the more sophisticated buyers get, the less they're just going to smash on the state-of-the-art model at the highest effort level for every single prompt. But the data retention policy really makes this not a particularly clear comparison. Now, lurking behind everything we've discussed in today's show is the fact that anthropic and open AI both have more advanced models more or less ready to go at this point that are being held back by a combination of government pressure, internal concern, or is simply the fact that because nothing else is caught up, they don't really have pressure to move things forward faster. Still, even if on the one

Starting point is 00:28:24 hand we are seeing a slowdown in the speed with which anthropic and open AI specifically are dropping models, I think it's pretty hard to look around the model landscape right now, and not feel like we have increasingly more rather than less choice. Anyways, friends, some fun new treats to try for the weekend, but that is going to do it for today's AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace.

The AI Daily Brief: Artificial Intelligence News and Analysis - Grok 4.6 Shows How Fast Your AI Options Are Expanding

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.