Today on the AI Daily Brief, Fable 5 is officially coming back. Before that in the headlines, the quest to cut inference costs. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. Alright friends, quick announcements before we dive in. First of all, thank you to today's sponsors KPMG, robots and pencils, Blitzie, and Airtable. To get an ad-free version of the show, go to patreon.com/aidelybrief or you can subscribe and upload podcasts. And if you want to learn more about sponsoring the show, send us a note at
[email protected]. We kick off today with a story that isn't very of the zeitgeist that we are living in right now. Open AI has found a way to slash their inference costs in half, sort of. This headline from the information grabbed a lot of attention and understandably so. Everyone right now is looking for new approaches to token efficiency and the implications of these searches have huge impacts on the business models and the companies that are shaping AI and the larger market structures they're operating in. Now when it comes to this article specifically, the details do suggest that it might be a smaller breakthrough than it appears at first. The claim is that open AI researchers have discovered a new optimization technique that cut their inference requirements in half for existing models. When the technique was applied to chat GPT users who weren't signed into the service, open AI was able to serve that entire user-based segment on just 100 GPUs. The open AI source didn't disclose what the technique was. The information speculated it could be quantization, cash optimization, batching queries or rooting queries to a lower power model. Notably, none of those techniques would improve service for open AI larger models without compromises. The universal truth that there is no free lunch remains and most attempts at optimizing inference come at the expense of model quality. Now there's also the question of what it means that open AI is testing this technique on a tiny batch of their least engaged users. That might be a totally reasonable starting point as the first test of many or it could be a cautionary approach that implies that there's some risk of quality degradation. The TLDRs that while there seems to be something interesting here, we probably shouldn't treat it like some sort of silver bullet to resolve the compute crunch. The information Stephanie Palazolo is convinced that open AI is onto something. In an accompanying video she said, "This is a very important secret sauce for them that they don't even want to tell other open AI employees about. Because if these things leak, it can quickly be picked up by other labs, which can also then use that to lower their costs. This is something they're holding very close to their chests." Now many pointed to a new research paper from DeepSeek, which open sources a speculative decoder system called D-Spark that can speed up inference by 85% during testing on small models. Now it's unclear how D-Spark impacts costs, but it is a reminder that inference optimization is not even close to a solved problem, meaning that theoretically huge gains from some novel technique are definitely plausible. And of course, even if open AI hasn't found a way to boost efficiency by 50% across the board, any sort of gains here could still be a very big deal. Open AI's army of free users are a significant drag on profitability, so anything they can do to cut inference costs to that user base could really move the needle. And given the particular audience, certain types of quality reductions may be more tolerable. Everett Randall of Benchmark Ventures has been talking about a phenomenon he's calling the AI Mom Test. He recently said, "There's nothing my mom actually asks of her AI products that needs to be done by the frontier or even a near frontier model." And it seems to be at least initially, like that could be the group that this new technique addresses. Certainly there is a lot of chatter out there about innovations and new approaches in this area. AI aggregator Andrew Curran tweeted, "I'm posting this prediction now so I can quote it later. There has been a significant breakthrough in architecture specifically around memory efficiency, not by one of the big labs, but by a team that was spun out of open AI. They will probably announce it soon." Now in addition to labs finding new approaches, companies themselves are also finding new more efficient architectures. 20-minute VCs, Harry Stevings tweeted, "In the last 24 hours, I have had five founders message me of varying sized companies, some 10 person startups and one $200 billion public company. All of them stated they have been able to cut inference spend by 75% or more with little effort, no performance change and better latency. The times they are a changing." Now speaking of innovations in this new token efficiency era, vibe coding platform base 44 has launched their own AI model in an attempt to shore up the business. The model is called base 1 and follows the same playbook as Curcer's Composer. Namely, base 44 has taken an open source base model and applied their own fine tuning using training data from hundreds of millions of user interactions on their platform. This CEO, my or Shlomo, laid out the strategic thinking in a few different ways. Firstly, base 44 is making the bet that narrowly trained models can be competitive with the frontier. This bet appears to be paying off for Curcer with most viewing composer 2.5 is good enough for common tasks. Right, Shlomo? General models need to be good at everything. They need to understand many programming languages, many workflows, many domains, and many kinds of reasoning. However, base 44 only needs their model to be good at building web apps. This base 44 also views the model as a cost control measure. Shlomo wrote, "It gives us more control over cost, latency, reliability, and quality while still letting us use the best external models where they are the right fit." Finally, he writes, "The model gives base 44 a way to utilize platform data to improve the product." The idea is similar to the harness model pairing that open AI and Anthropic have pursued with their own coding platforms. Base 44 believes they can develop the model and harness in tandem to deliver strong platform specific results. As AI becomes a bigger part of how software is created right, Shlomo, owning more of that intelligence becomes just as important as owning the infrastructure around it. Speaking of strategic convergence, AWS is launching a new division to join the AI deployment race. AWS announced that they will invest $1 billion to create a new unit staffed with forward deployed engineers to help customers set up and use AI tools. This follows of course open AI and Anthropic both launching private equity partnerships to house their FTE divisions with Google and May expanding their existing FTE division and Microsoft announcing an FTE partnership with EY in June. Francesco Vasquez, VP of AWS's Frontier AI Engineering and Services, said that the company has been upskilling salespeople to be FTEs giving them the title of solution architects. Vasquez said the effort is an expansion of AWS's Generative AI Center which was first announced in 2023. She added that AWS will focus on industries with the strongest demand, including healthcare government and financial services. Re-enforcing the themes of the moment Vasquez also noted that the big shift towards AI budget optimization rather than just the deployment of AI capabilities. She said, "Open source and the use of open-weight models is definitely gaining traction for customers for a variety of different reasons. Price performance but also they service the task." In another indication of just how significant this trend is, the AI Engineer World's Fair which is happening this week in San Francisco is actually hosting an AI FTE mini conference within the event. Now following up on the Claude Tag story from last week, Anthropic seems to be preparing to bring agents to Microsoft Teams as well. The information reports that Anthropic recently told Microsoft that they plan to bring a version of Claude Tag to Microsoft Slack equivalent Teams. Claude Tag you might remember allows users to summon Claude into a channel to receive tasks or take instruction. Unlike previous Claude integrations, Claude Tag isn't tied to any particular user. Instead it functions as an organization-centric agent with persistent memory and tool access independent of the user. In fact, it's actually not calling Claude exactly, it's calling the full suite that comprises Claude code. Now one interesting sub-story behind all of this is what it says about the relationship between Anthropic and these platforms both Slack and in the future Microsoft. There has been some scuttle butt that behind the scenes, some Salesforce employees were concerned about Anthropic being allowed into the ecosystem on such a fundamental level. Although I don't think that that was universal given the amount of PR that I got from Salesforce about this integration. But still, introducing Claude Tag to Microsoft does add another level to this power struggle. Currently, both Salesforce and Microsoft allow third-party agents to access their ecosystems free of charge. In fact, Microsoft CEO Sachin Adela went one better during an investor call in April. He claimed that Claude helped reinforce Microsoft's ecosystem commenting, "It's fascinating that here we are in 2026, and the most exciting things in AI are plugins and word or Excel." When you see that, that means we have a structural position in knowledge work. That said, will the tune change as Claude gets more endemic across the knowledge work spectrum? It's certainly something interesting to keep an eye on. Lastly, today, one of the things that we've been talking about when it comes to the data center build out is the opportunity for the data center builders and operators to, if I might be allowed to put it crassly for a moment, buy off the communities that they're operating in, or perhaps a better way to think about it, is to cut them into the benefits that the data centers might represent. One example of someone starting to nudge down that path comes from SpaceX. The company is discounting starling subscriptions in Memphis in an attempt to quail backlash. SpaceX's colossus data centers are located just south of Memphis and have been the subject of significant local controversy. The campus operates on 46 gas turbines providing off the grid power. Local groups have complained about air pollution and noted the turbines are operated without a permit as they're technically portable, which some see as abuse of illegal loophole. The US government in fact recently intervened in a lawsuit to shut down the turbines, claiming that colossus is a matter of national security. On Tuesday, SpaceX announced that residents in the greater Memphis area will receive half-price starling subscriptions and free hardware for new signups. In addition, the mayor of Memphis recently announced that SpaceX has recommitted to the construction of a wastewater treatment plant. Those plans had been halted in April with SpaceX claiming they were prioritizing the construction of the colossus two data center. However, the construction halt came just days before the lawsuit was filed. Announcing the discount, VP of Starling, Michael Nichols wrote, "The unique capabilities of the colossus data center could not be accomplished without the partnership and support from the local Memphis community. Happy to bring affordable and great starling connectivity to our neighbors." Look, directionally, I'm glad that we're starting to see this, but I have to say,
But for anyone from SpaceX who might be listening, while I applaud this direction, I would be going a lot harder than half-priced Discount for people who have to become your customers to even get that discount. Still some encouraging signs that we're starting to think differently about these relationships Still that's going to do it for today's headlines. Next up the main episode One of the most important AI questions right now isn't who's using AI? It's who's using it well KPMG in the University of Texas at Austin just to analyze 1.4 million real workplace AI interactions and found something surprising. The highest impact users aren't better prompt engineers They treat AI like a reasoning partner. They frame problems, guide thinking, iterate and push for better answers. And the good news? These behaviors are teachable at scale. If you're trying to move from AI access to real capability KPMG's research on sophisticated AI collaboration is worth your time. Learn more at kPMG.com/us/sophisticated. That's kPMG.com/us/sophisticated. I cover the capability gap between AI potential and AI reality every day on the show. Most companies are still figuring out how to start. Robots and pencils is already launching and scaling. A gender-in-generative AI in production at large enterprises in weeks, AWS Advanced tier, Pattern Partner more than doubled in a year. And they're hiring 50 open roles. If you're someone who knows this moment is different who wants to be inside it not watching it. This is worth a look. A robots and pencils the best ideas win and the team is purposefully kept super high quality. This is the kind of place you look back on as the best decision you ever made. Take a look at robots and pencils.com/careers. If you're looking to adopt an agent at SDLC, Blitzie is the key to unlocking unmatched engineering velocity. Blitzie's differentiation starts with infinite code context. Thousands of specialized agents ingest millions of lines of your code in a single pass mapping every dependency. With a complete contextual understanding of your code base, enterprises leverage Blitzie at the beginning of every sprint to deliver over 80% of the work autonomously. Enterprise grade and to end tested code that leverages your existing services components and standards. This isn't AI autocomplete. This is spec and test driven development at the speed of compute schedule a technical deep dive with our AI experts at Blitzie.com. That's BLI TZY.com. This episode of the AI Daily Brief is brought to you by Hyper agent where you run fleets of agents your team can manage together. New users get a thousand dollars in inference. Forget local agents and chat workflows waiting on your laptop to be prompted. Hyper agent deploys 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 and reaches leads draft emails and updates the CRM. Ops agent chases the paperwork contracts the budget every agent has access to shared context and follows your rules about scope and approvals. It's time you add agents that feel like teammates hire yours at Hyper agent built by the team at air table claim your one thousand dollars in inference at hyper agent calm slash AI Daily Brief. Welcome back to the AI Daily Brief. Well friends after something like 19 days offline. It should be that by the time you're listening to this, Fable 5 has been turned back on on Tuesday night and tropic announced that the government had lifted export controls and cleared them to begin redeploying Fable 5. At 7.52 p.m. this tweet exploded onto the scene. We've received notice that the Department of Commerce has lifted export controls will begin restoring access tomorrow and will share an update soon. We're grateful to our users for their patience and to everyone who worked with us on redeploying the models. A few hours later they shared further details on the rollout beginning today July 1st. Fable 5 will once again be available to all global users across all paid subscriptions. There will be a short extension of the subsidy period that we were promised when it first came out. Fable 5 will be included for up to 50% of weekly usage limits until next Tuesday. But after that access to the model will require the purchase of usage credits. Mid-thouses restrictions remain as they were at the end of last week with approved US firms able to access the model for both domestic and foreign workers. And tropics said they will continue to work with the government on an expanded rollout under project glass wing, including providing the model to international firms. Several administration officials commented on the resolution. White House chief of staff Susie Wiles who according to reports has been one of the AI policy leads in recent months wrote under President Trump's leadership the United States is the undisputed winner in the AI race. My gratitude to companies across industries who continue to work closely with the White House to implement the president's executive order promoting advanced AI innovation and security. This includes excellent work around advanced model access and guardrail testing and security. The government and private sector have worked together in a way we have never seen before and this foundation of America first is unprecedented. Our shared priority remains get the best tech deployed as quickly and safely as possible. Commerce Secretary Howard Lutnik who has been in charge of applying the export controls added more specifically over the past two weeks we have worked closely with anthropic to analyze and improve fable 5 to ensure alignment across the US government and strengthen America's leadership in AI. Now alongside the announcement and tropic provided their own version of the events of the past few weeks they discuss the jailbreak reported by Amazon and their efforts to convince the administration that nothing was a miss. As a refresher the core claim was that fable was able to identify serious vulnerabilities in a code base which the administration believed was a mythos level capability. Our testing wrote anthropic confirmed that many less capable models including Claude Opus 4.8 GPT 5.5 and Kimi K 2.7 could identify the same vulnerabilities as fable 5 did in the report when it came to the demonstration of how to exploit the single vulnerability every model we tested could produce the same demonstration as fable 5. Including Claude Highco 4.5 Saunut 4.6 Opus 4.6 Opus 4.7 Opus 4.8 GPT 5.4 GPT 5.5 and Kimi K 2.7. Importantly they continued the reported technique did not expose any unique mythos level cyber capabilities the behavior reflected a borderline case for fable 5's safeguards. There are some tasks that are unlikely to be dangerous but are nonetheless blocked by the safeguards out of an abundance of caution. The reported technique allowed access to one such behavior but it only involved routine defensive cyber security work. Now this has been an anthropic's position from the beginning that although this was a genuine deal break it didn't unleash dangerous cyber attack capabilities. Nevertheless, anthropic has amped up the guard rails for fables return. They have trained a new classifier designed to target and block the behavior described in the Amazon report with a claim success rate of 99%. As with the previous version users will be informed if they trip the guard rails and will be reverted to Opus 4.8 for that request. Anthropics said that they have tested the new classifier with the Commerce Department Center for AI Standards and Innovation which agrees that anthropic safeguards are "extra-narrally strong". Anthropic noted the new classifier also comes at the cost of flagging benign requests more often during routine coding and debugging tasks. As with all our safeguards will continue to refine this to better distinguish genuine misuse from legitimate requests and reduce false positives. Now one of the first responses was effectively that while folks were glad that Fable was coming back it wasn't exactly clear what had changed that addressed the government's initial concerns. Policy advisor Dean Ball wrote great news but we have no idea what anthropic did to make the model safe what commitments anthropic is made going forward and whether or how any of the supplies to other frontier models in the government's licensing queue. We know that GPT 5.6 is in that queue but it's fair to assume that other model developers are at least in early stages of submitting their models. Re-enforcing the point that he has made over and over again Dean continued this opacity will not lend itself well to a stable investable trust whether industry over time but and here Dean ends on a positive note. The US government needn't figure this all out in a day and a two week review timeline is not insane in the grand scheme of things the status quote is not tenable but we made progress today that progress is just a first step but it is worthy of applause nonetheless. Prince also pointed out how many new questions the letter about the export controls being lifted brought up with anthropic agreeing to quote proactively detect and address security risks associated with Fable 5 and with those five Prince asks how is anthropic required to proactively detect these security risks not publicly disclosed when anthropic agrees to work diligently with the US government on protocols and standards and releases for mythos fable and future models Prince notes that this is extremely broad language that appears to cover all future models and is not limited to cyber security risks. Still I think miles brandage captured the sentiment of many when he wrote the first rule of Fable Club is you do not ask too many questions about what exactly anthropic agreed to that they weren't doing before and you enjoy your access. And overall from a policy perspective that seems to be the tone even though people have lots of questions the fact that we're getting the model back in this version gives people reason to be cautiously optimistic. Rights boxes are in the V. It's been a messy process to get here but at least there's some semblance of a framework that could be practical the note of caution here would be that there's a lot of subjectivity that goes into various risks and their actual levels of exploitability and practice. We're likely going to be living with a framework that requires heavy judgment and back and forth between labs and the government for major releases the best we can hope for is that this is a relatively efficient process and hopefully as ways of being sped up for incremental version updates and models. It would be a bad outcome if every release after this level of threshold of capability required the same review process and we don't get the same rate of breakthroughs we've been seeing. Now one of the reasons that I think people were were willing to extend that benefit of the doubt was that as recently as yesterday people were speculating about Fable 5 coming back only with things like a new KYC regime where people had to verify their identities and perhaps verify their identities as American citizens before they could get access. That appears not to be the case Fable 5 is not just coming back for us users but for all users globally and yet there is a big question lurking when I tweeted an image of Dario and Uncle Sam riding an eagle delivering Fable across America flanked by fighter jets. Aaron Schneider asked the key question but will it be as good as we remember the specific concern comes around this.
line from the announcement, in the near term, some routine tasks like coding and debugging will fall back to Opus 48. That led folks like Lissan on X to write, the Fable 5 relaunches kind of fake, some routine tasks like coding and debugging will fall back to Opus 48. You can use that even more restricted Fable 5 version in your Anthropics subscriptions until July 7. Lex on X wrote, "LMFAO, this cannot be a real statement for Anthropic. Routine tasks like coding, skull emoji. They've lost their minds. Do they think people are paying an absurdly high price just for the privilege?" Now to read from the Cloud Code team, came on to clarify that that tweet had particularly loose language. He wrote, "As with the original classifiers, a small fraction of routine coding and debugging tasks will be flagged and fall back to Opus. In other words, while the tweet made it seem like coding and debugging were part of the routine tasks that would fall back to Opus in general, the Cloud Code team is saying that no, in fact, it is just a small fraction of routine coding tasks that will be flagged in that way." Now, given all this, I want to come back and talk about the ways that I think you should start to test Fable right away when it comes back, but before that we actually do have one more Anthropic model to talk about. Earlier on Tuesday, Anthropic announced Claude Sonnet 5. And if I had to guess, this suggests to me that they were not sure that Fable would be coming back as fast as it was, as I'm not sure that they would have announced this particular model when they did, in the way that they did, if they knew that later that night, they'd get to announce that Fable 5 was coming back. Anthropic pitched the model as their most-agentic version of Sonnet yet, writing, "It can make plans, use tools like browsers and terminals, and run autonomously at a level that just a few months ago required larger and more expensive models." Now, according to the benchmarks, Anthropic tried to pitch the model as almost as good as Opus 48 for a fraction of the cost. It's a few percentage points shy of Opus on the two major coding benchmarks, Sweetbench Pro and Terminal Bench 2.1, with the same gap existing for computer use in knowledge benchmarks. Maybe the most interesting result was GDPVAL where Sonnet 5 delivered a huge jump over Sonnet 4.6, and even slightly outperformed Opus 4.8. Now, this could be indicative of the strong agentic performance that Anthropic was touting, while GDPVAL aims to measure economically valuable work, in practice, the score largely comes down to successful tool calls. Sonnet 5's score could indicate that it's much more capable of following through with end-to-end agentic work, rather than getting stuck halfway through. While Sonnet 5 retains the same pricing as Sonnet 4.8, Anthropic is trying an introductory price for API use, until the end of August Sonnet 5 will cost $2 per million input tokens, and $10 per million output tokens, and will revert to the standard $3 $15 after that. In contrast, the price for Opus is $5.25, so Sonnet 5 could be a more cost-effective option for some use cases. But what about external benchmarks and actual user impressions? On Cursor's Cursor Bench, they wrote that it was a meaningful step up from Sonnet 4.6. It also saw a jump on the artificial analysis intelligence index from Claude's onnet 4.6's 47 to a score of 53, which puts it just one point behind Claude Opus 4.7, and a couple points behind GPT-5.5, but they point out that without the promotional pricing, it will actually cost more per task than Opus 4.8. On max effort, they write, Sonnet 5 used around 40% more output tokens per intelligence task than Sonnet 4.6, and around three times the number of agendic turns for their knowledge work evaluations. The overall run, in fact, cost more than Opus 4.8. This is because it generates almost twice as many tokens as Opus 4.8 per task, and when it came to running the entire bench of tasks, Sonnet 5 was actually more expensive than Fable. That led a lot of people to wonder what the heck this is actually for. Max Blade wrote, "Composer 2.5 is faster, GLM 5.2 is cheaper, Opus 4.8 output is 10 times better. It's an improvement, yes, but without a real use case, it is dead on arrival right now. Will you guys be running it?" Now, one important caveat with that is that the artificial analysis tests are run at max settings, and that might not be the optimal way to run Sonnet 5. David Shapiro writes, "Okay, I can't believe I'm going to say this, but Sonnet 5 max is too high effort. It's like giving a box of squirrels a bunch of cocaine and saying, "Go with God," and just seeing what comes out the other side. Ben Davis, who works with YouTuber Theo wrote, "After doing around a billion of tokens on it today, you're all wrong. Sonnet 5 is good. Is it inefficient? Yes, slow? Yes, expensive. Yes." But if you think it's the same as GLM 5.2, a model I really like and will continue to defend a ton, you're a fool. The thing with this model is you have to use it wildly differently. It's not the type of Sonnet model we're used to. Probably could have been Opus 5 if it was slightly bigger. It's basically an automatic Ralph loop. It's spawning tons of sub agents, making many stacked PRs, reviewing itself adversarily, auto-testing, its changes and not drifting off-task at all. If you use this thing the same way you've been using old models, you're going to have a bad time. Indeed, summer-wondering, if the real way to look at Sonnet 5 is as the work model that would run the sub agents that something like Fable 5 would spin up. Dan McCadia writes, "You'll use Fable 5 as the super-intelligent advisor and Sonnet 5 as the fast and efficient implementer." Still, when push comes to shove, the model that people are really excited about having back is Fable 5. And for anyone who used it in those first three days that it was actually available, you'll understand why. So the question becomes, "As Fable 5 comes back online and we have this one week where you can use it in your normal subscriptions, what should you be using it for?" Ennecate Pange-Wonney writes, "You have one week to use Fable 5 without going bankrupt. Here's how you should make the most of it. One, use Fable 5 for planning, not for implementation. Use the Codex plugin and Cloud Code to delegate implementation to GPT-5-5. Two, ask Fable for suggested improvements on your projects, starting with your most important invaluable projects. Three, use Opus and GPT-5-5 to brainstorm what your hardest technical problems are across your projects and then ask Fable to propose solutions to them. Four, use GPT-Pro through Oracle to review Fable's output. Now this reflects a lot of the advice that was around when Fable 5 first came out. Basically, that what it was for was your hardest challenges and that it was almost too powerful for routine day-to-day things. Now, I agree with half of this. The half-wear Fable is very good at your most difficult, specifically technical problems. And to the extent you have some of those, whether you're technical or not if you're working on some hard-coding project, obviously spend as much time as you can using Fable 5 for that. But one thing that already, in the very first couple of days of using Fable 5 that I disagreed with the common sentiment around was that you wouldn't necessarily notice the difference using Fable 5 on more routine, banal, non-technical tasks. In my experience, that was completely wrong. Likely, the two most common categories of tasks that I have outside of coding, building, and prototyping things is strategic thinking and writing. Now, on strategy, Fable 5 in my limited experience, Blue GPT-5-5 in Opus 48 out of the water. Both of those two models are extremely steerable. They are overly differential to push back and tend to overinterpret instructions. For example, if you ask them a strategic question and then try to say something to reduce sick a fancy like, "Don't just accept what I'm saying is true," provide pushback if it's warranted, those models will assume 100% that their response will only be successful if there is pushback. Then, however, if you push back on them, they will cave almost immediately and go out of their way to justify whatever it has become clear they think you're trying to get them to say. Fable 5 didn't do that. In my interactions with Fable 5, as I was debating with it, it would frequently accept part of my pushback ideas while sticking to its guns on other parts. That is behavior that I have never seen from any other model and instantly made it a thousand times more valuable when it came to any sort of strategic thinking or iteration. Given that every single one of you I guarantee has strategic questions that you're working through at any given time, this is something that I would immediately experiment with in Fable, and by the way, it also has the benefit of not really consuming all that many tokens, so it's not going to use up that 50% usage limit particularly quickly. The second common sentiment that I want to push back on is around writing. Now, arguably, every is the most thoughtful tester when it comes to new models in writing. And in their initial vibe check, it wasn't in their estimation all that much better. They called it "sharp judgment trapped in familiar pros". They wrote, "Every's writing benchmark found that Fable's writing is clear but still short of human judgment on what to include in a piece of writing and how to structure it". The benchmark asks the model to write an introduction from scratch, fill in a missing paragraph from context, and deliver a promotional email, a LinkedIn post, and an ex-post. On the editing side, it asks the model to replicate human edits and detect common A.I.isms in a deliberately robotic draft. And set to extra high effort for this, it basically scored between Sonnet 4.6 and Opus 4.8. Now, on a particular type of writing task, this is not what I found. In my not-tests, but real-world uses of Fable 5 for writing, I found that it was way better at instruction following, and fell into far fewer of the common A.I. traps. It's writing wasn't overly rot and try hard, it had fewer of the standout A.I.isms like it's not this, it's that, and while I would want more reps in to say definitively that it's better for all types of writing, my suspicion is that especially in situations where you have a fairly clear rubric of what a good example of writing looks like, I think it's going to be able to do a much better job at meeting that standard than those previous models were. Now, that doesn't mean that it's any better at blank page writing, but given how many of people's writing use cases for A.I. are saying, here's all the examples of what has been done in the past, do this type of thing in the future, my experience so far is that Fable 5 is much better at that. Look, when all is said and done, it is extremely, extremely welcome news that it is coming back to everyone. While I don't think that you should change your 4th of July plans to stay inside hacking at some big project, I certainly wouldn't blame you if you did. For now, that's going to do it for the A.I. Daily Brief. I appreciate you listening or watching as always and until next time, peace!
PodcastTranscript AI