Anthropic rolls back Fable 5 while Microsoft builds its own AI model | Now Shipping

Mike Belsito on this week's episode of Now Shipping covers three stories shaping how product teams build on AI.

We cover:
— Why Microsoft built its own code generation model despite investing $13 billion in OpenAI
— What the retirement of GPT-4.5 on 27 June means for product teams — and why model deprecation is now a product management problem
— How a multi-agent safety bypass led the US government to give Anthropic 90 minutes to pull its most powerful model, and what that means for teams building on single AI providers

Chapters
(01:37) Microsoft launches MAI Code One Flash
(04:22) What this means for product teams
(06:26) GPT-4.5 retirement on 27 June
(09:45) How to manage model dependencies
(10:59) Anthropic's Fable 5 pulled by US government order
(13:50) AI vendor risk as a product architecture decision
(15:38) Closing thoughts

Transcript

SPEAKER_00 0:00

I'm Mike Belsito, and this is Now Shipping, the weekly AI news show for product people. Every week I bring you three AI news stories that matter to you as somebody who's actually building software products for a living. Not the hype, not the noise, just the stuff you need to keep up with. Brought to you by the team at Mind the Product. This is Now Shipping. And we've got a pretty big uh new show here today. You could see I'm in a different setting right now. I'm not cooped up in my attic in Lakewood, Ohio. Um, I'm actually in London, England. I just yesterday, Mind the Product had MTP Con 2026. It was an awesome time. I'm exhausted, but I'm still excited to bring you the stories this week. And it's funny, things change so fast. That's why we made this show. One of the stories from last week, we have a follow-up this week. It's the third story we'll be covering. But first, story one Microsoft already pays OpenAI more money than any other company on earth. So why did they just build their own AI coding model? Story number two, GPT 4.5 is being retired on June 27th. If your product's running on that model, you might want to keep watching. And finally, yes, story number three is about Anthropic and the launch of Fable 5, or should I say, the rollback of Fable 5? Because last week the US government contacted Anthropic and gave them 90 minutes to pull their most powerful model. That is a lot, that's a big deal. We're gonna get into it all right now. All right, story number one is all about Microsoft. And I'll start by maybe asking you a question. If you were spending billions of dollars a year on buying AI from another company, and that company was getting ready to go public and maybe kind of compete with you in some areas and set their own prices to whatever they want, would you start building your own AI models? Because that's exactly what Microsoft just did. This week at its Build Developer Conference, Microsoft unveiled MAI Code One Flash, its first internally built code generation model. This is a model that's designed to take a written description and turn it into working code. And if that sounds familiar, yeah, that's exactly what Codex does. It's basically the same kind of thing, but now it's competing. So here's the thing Microsoft is probably OpenAI's most important business relationship. Microsoft's poured more than $13 billion into OpenAI. Azure is the infrastructure that OpenAI runs on. So when you're using ChatGPT, you're using Microsoft's cloud. It is deeply, deeply intertwined with OpenAI. And yet they just announced a model that directly overlaps with one of OpenAI's core developer tools. The official reason that Microsoft gave was pretty interesting. They said they want to reduce reliance on OpenAI and lower costs for developers. Now, think about that. Reduce reliance. They could have said we want to extend capabilities, uh, we want to, you know, offer more features. No, they said reduce reliance is a pretty direct statement. So I think Microsoft isn't necessarily going to be alone in this kind of thinking. Um, we're seeing with Google right now. Google's been building its own coding models. Anthropic is investing heavily into Cloud's coding capabilities, um, GitHub Copilot, which Microsoft owns, by the way. It already competes with OpenAI's tools in some ways. The entire AI ecosystem right now is simultaneously partnering and racing and hedging all at the same time. Now, as for Microsoft's model, MAI Code One Flash, it's a flash model, which in AI terms means it's built for speed and low cost. Not necessarily, you know, top-of-the-line performance. Think of it as a workhorse, right? The kind of model that gets deployed everywhere at scale of doing a lot of the heavy lifting behind the scenes. Microsoft tends to be pretty good at winning this kind of like workhorse category. So, what might this mean for you as a product person? It might mean a few things. I mean, first, platform stability and AI, it's not a given. I mean, hopefully that's obvious by now. I know we talk a lot about these platforms, like they're dependable as AWS for file storage. But this week is sort of a reminder the relationships between major AI players are complicated. They're not so straightforward, right? Like OpenAI and Microsoft, they're partners. OpenAI and Microsoft, they're also competitors. And the model you're building on today might come from a vendor who's actively trying to wean themselves off of the thing that you depend on. So if you've built deep into one provider's tooling, whether it's an API or a model family or an ecosystem or platform, it could be a concentration risk, right? And it's not a hypothetical, this is the kind of thing that's happening right now. Um, the second thing here is I think it'll be interesting to watch what Microsoft does next. MAI Code 1Flash is positioned as a cost reduction move for developers, but Microsoft has historically used developer tools as a Trojan horse for a deeper platform lock-in. Think GitHub, Visual Studio Code, Copilot. They're all ways of embedding Microsoft further into the daily workflow of anybody who's building software. This model's probably the next layer of that. So the question for product teams isn't just should we use the model? It's what does Microsoft want to do next? And is that actually where we want to go? And then the third uh takeaway from this story, I'd say, is if you're building developer tools or any product aimed at technical users, it's worth taking this pretty seriously. I mean, this coding model category is getting kind of crowded. And when the richest technology company in the world starts building in your space, even if they're saying it's just to reduce costs, it usually means something bigger's coming. So keep that all in mind as you move forward with your own products. That's story one. Let's move on to story two. All right, story number two is about the sunsetting of GPT 4.5 by OpenAI. And you might not think this is a product story, but I think this is one that most product people should actually be paying attention to, even if you're not using GPT 4.5 in your product right now. Um, okay, right now, but let's get to it, I guess, right? So open AI, yes, they're they are sun setting GPT 4.5. GPT 5.5 will now be the default replacement for most users, although OpenAI does have a range of GPT-5 models available depending on your needs. Now, if you are using 4.5 in your product right now, this is something you definitely need to act on, right? Like you need to make sure that you are moving forward and getting your product um off of the 4.4.5 model. Open AI has sent notices, they've updated their docs, they gave everybody 30 days to migrate. Now, for enterprise software, 30 days, not a lot of time, but it's also not nothing, right? So you did get notice about this. Yet there are going to be a lot of teams scrambling anyway. That's just how it goes. Um, AI model management, you know, at one point this might have been thought to be uh an engineering problem, but this is a product problem, right? Like if you've built features on top of GPT 4.5, even if you weren't the ones tracking the deprecation timeline, this is something you have to be paying attention to. So we've been talking for years about technical debt, right? The cost of building things fast and patching them together in ways to make sure that your system isn't so fragile over time. AI model dependencies, it's sort of like the new technical debt, right? And a lot of teams don't necessarily have a good plan for keeping up with it all. So think about what it actually means to build a product feature on top of a specific model version. We know the model's going to change. Now, hopefully it gets better in many ways, but it isn't just about improvement. These models behave differently. A prompt that worked perfectly on GPT 4.5, it might produce completely different outputs on GPT 5.5, whether it's something like the tone shifting or format shifting. Um, there might be edge cases that were handled one way and now get handled another way. Your users start experiencing inconsistency and they don't know why. And then eventually the model gets retired completely and you have to migrate or else your product could break. So this is going to keep happening. Open AI has already sunsetted lots of different models before. Anthropic is, they've sunsetted models as well. Like it's going to keep happening and then they're going to do it on their own timeline, not your timeline. So, what should you actually do about it? Well, first, if you're using any AI model in production, model version management has to be on your roadmap, not just tracked by engineering. The product team has to have their own plan for what happens when a model changes and goes away. Like, what do you test? What's your acceptance criteria for a migration? Who signs off on the new behavior to say whether it's good enough or not? These can't just be engineering questions. These have to be product questions too. And second thing is, you know, there is an argument for building abstraction layers between your product and the underlying model. Don't let the models API be the thing that your features call directly. You might want to have that middle layer that your team controls, something that can swap models without requiring you to rewrite your product logic. I know that could be over-engineering when you're trying to move so fast. It sounds like it could be the right call, though, when you're scrambling to migrate 10 different features in two weeks. Um, and then finally, just like know what your product actually needs from a model. I mean, if you've been using GPT 4.5 and you've never stress tested your prompts on GPT 5.5, you don't actually know how the migration is going to go yet, right? So now's the time to find out before you actually have to go through and migrate. So I think model deprecation is a new PM problem. This isn't just an engineering thing. And the sooner your product team owns that, I think the less surprised you'll be the next time a retirement notice comes in your inbox. And again, there's going to be a next time. So make sure that you're ready. All right, one more story to go. All right, story number three, and this is a big one. You might remember a story from last week's episode about anthropic launching Fable 5, their most powerful AI model, a model that's a part of their mythos class. And now we're here to talk about the pullback of Fable 5. Look, this is just illustrative of how fast things are moving in this AI everything world that we're in. So what actually happened? Well, the US government was the one that actually contacted Anthropic and gave them 90 minutes to pull Fable V back. Why? Well, on June 10th, a day after Fable V launched publicly, a well-known AI researcher goes by the name of Pliny the Liberator, published that he'd managed to bypass Fable V safety filters using a sophisticated multi-agent attack. Now, this wasn't a simple jailbreak. This was a coordinated technique that broke harmful queries into innocent-looking pieces, routed them separately through the model, and then reassembled the outputs. He published what he extracted, the word spread fast, and the government noticed. Why? Well, they got a phone call from the CEO of Amazon, Andy Jassy. He actually flagged a similar vulnerability to senior administration officials, including Treasury Secretary Scott Bessant. Now, this is pretty wild because Amazon has invested $13 billion into Anthropic with commitments for potentially much more, numbers up to $20 billion more. Now, what ended up happening? Well, again, yeah, the commerce department sent Anthropic a letter. They got 90 minutes to pull everything back. Anthropic leaders flew out to Washington for emergency talks. And as of this episode, there's no clear timeline for when Fable 5 can come back online. Although by the time this publishes, who knows? Maybe it'll be back online because again, things move so fast. What does this all mean for product people? Well, the easy story would be to talk about like, you know, AI safety drama or you know, government overreach. Um, and those conversations, maybe they should be had. I mean, I think that those are both legitimate things to talk through. But I'm more interested in like, what if you had Fable V in production for your products? What would you have done? I mean, a lot of teams would have no answer. According to a recent survey of over a thousand CIOs worldwide, 16% of companies have no contingency plan if key AI vendors become unavailable. Think about that. The numbers actually probably a lot more because companies these days they don't really have a definition of what an AI contingency plan even looks like. So we're saying 16%, which is significant, but it could be double, it could be triple. And what would this mean for you if you were one of those companies? Well, I think AI vendor risk is now a product architecture decision, or at least it should be. You would never build a production system where the only database was a single managed service with no failover, no backup, no alternative. That's a single point of failure, right? Now, AI model dependencies, it's kind of the same thing. Right now, most product teams have single points of failure baked directly into their architecture. Whether it's open AI or Anthropic or Google, one vendor, one model, no tested fallback. And also a multi-provider strategy, I think at this point, it can't be optional. I mean, it would be extra work, yes. But the alternative is what happened to Fable 5 users this week. You need tested fallbacks, not theoretical ones, not like, well, if something happened, this is what we could do. No, you need tested ones. You need ones where you can actually route real traffic through a backup model where you could check the output quality. You know what your users are going to experience. If the primary model goes down, I mean, that should be the bar. Now, this whole anthropic Amazon dynamic in this story, it's super interesting, right? Amazon invested $13 billion in anthropic, and yet they're the ones triggering the shutdown. I mean, it is pretty wild what's happening in this whole AI world right now. The people funding your AI providers, maybe the people that, for their own reasons, create problems for your AI providers. It's just the world that we're building in right now. But look, if your most important AI vendor went offline tomorrow, not for an hour, but let's just say for a week, what would happen to your product? If you don't have a clear answer to that, this is the week to figure that out. Now, I think that'll put a wrap on episode three of Now Shipping. I hope you found this valuable. If you did, please subscribe. Please tell a friend. Of course, if you have suggestions, leave a comment below. I will read every single one of the comments you leave. And trust me, we want to make this show even better in the future. And I'll be back with you next week with three more AI news stories that I think matter to you, the product people. So, once again, my name is Mike Belcito, and brought to you by the team at Mind the Product. This is now shipping.