Anthropic’s Hidden Fable 5 Rules, Google’s $5 Price War, and Diffusion Models for Text

Yesterday in AI  |  Thursday, June 11, 2026

Anthropic’s Hidden Fable 5 Rules, Google’s $5 Price War, and Diffusion Models for Text

The commercial launch of Claude Fable 5 came with some surprising fine print. Today's episode breaks down Anthropic's quiet end to its Zero Data Retention policy and the "hidden interventions" built into its most capable model.

Plus, Google just dropped the first sub-$5 AI subscription in the US, igniting a massive price war that could threaten the valuation of pure-play AI labs. We also explore Google's groundbreaking Gemini 3.5 Live Translate that preserves your exact voice, the experimental new DiffusionGemma model that generates text 4x faster using image-generation techniques, a terrifying IBM survey for IT leaders, and OpenAI's push for a global youth AI safety institute.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:05

Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in ten minutes or less. I'm Mike Robinson. It's Thursday, June 11th, and the morning after Fable V conversation just got very interesting overnight. Let's get into it. Yesterday's big story was Anthropic shipping Claude Fable 5 to the public for the first time. It's their Mythos class model, their top capability tier, now available with safety guardrails attached. We covered the launch on Wednesday's episode. Today we're covering what the fine print actually says. Two things surfaced Tuesday that deserve real attention. First, Anthropic quietly ended its zero data retention policy for all high capability models. Under the old policy, enterprise customers could negotiate agreements where Anthropic wouldn't store conversation data at all. Going forward, every interaction with Fable 5, Mythos 5, and any future model at that capability tier gets a mandatory 30-day retention window, and Anthropic will use that data to defend against what it calls complex and novel attacks. They're not training on it, they say. But the policy changed, and it changed quietly, buried in launch notes. Second, and this one's thornier, Anthropic has added what TLDR AI is calling hidden interventions. These are invisible adjustments that work below the surface and can limit the model's effectiveness for certain use cases without any signal to the user. The model doesn't fall back to a different model, doesn't throw a warning, it just quietly becomes less helpful. Anthropic says this affects roughly 0.03% of developers. The use case they cite is blocking competitors from using Claude to train their own models. Fair enough is a business objective. But if you're an enterprise building a workflow on top of Fable 5, you have no way of knowing whether you've hit one of these invisible walls. The Figma director of developer product called it a clear step forward and agentic coding, AI that writes and runs code autonomously. Cursor, GitHub, and Lovable confirmed it's the most capable model in their testing. Stripe reportedly used it to compress two months of Ruby code-based migration work into a single day on a 50 million line repo. The model works. Whether you can trust it'll work the same way tomorrow is a different question, given Anthropic can change its behavior without telling you. The deeper context here is IPO pressure. Anthropic filed its confidential S1, IPO registration paperwork, last week at a $965 billion valuation, and OpenAI filed its own on May 22nd. The Deep View ran a sharp piece arguing that Anthropic had essentially no choice but to ship. The fundraise demands product at scale, the IPO demands revenue, and sitting on a mythos class model while investors are watching doesn't work commercially. So they shipped it with guardrails, changed the data retention terms, and built invisible tripwires for edge cases. That's the compromise. And it's worth knowing that's what it is. So while Anthropic is out here trying to justify a massive IPO valuation with $50 enterprise tokens, Google looked at the room and decided to just nuke the pricing floor entirely. Google AI Plus dropped from $7.99 to $4.99 a month and doubled storage to 400 gigabytes. That's the first sub-5 AI subscription plan in the US market. The plan includes video generation, Notebook LM, and Google Flow, aimed at students and individual users rather than enterprise teams. This same dynamic has been running in India and Southeast Asia for almost a year. OpenAI launched a sub-5 ChatGPT plan in India last August. Google matched it in December. The US just got that same logic. The timing is not a coincidence. OpenAI and Anthropic don't have a budget tier or localized pricing anywhere. One VC told TechCrunch this marks the beginning of an AI commoditization era, the same kind of margin compression that eventually destroyed Cisco's pricing power after the web era. When infrastructure becomes a commodity, the premium goes to whoever owns the relationship and the distribution. Google has both. OpenAI and Anthropic, pre-profitability, going public with $50 per million output token pricing, now have to explain to investors how they survive a world where the biggest distribution platform is charging $4.99 a month. But Google didn't just spend the week starting a price war. They also dropped a real-time translation tool that might finally put an end to the awkward stop and stare silence of international Zoom calls. Standard real-time translation tools work in chunks. You speak, there's a pause, the translated output plays. It's functional but awkward. Gemini 3.5 Live Translate changes that. It's a speech-to-speech model that follows the speaker continuously, translating across 70 plus languages while preserving intonation, pacing, and pitch. A few seconds of lag, but no stop-start rhythm. And it matches how you actually sound. It's rolling out in three places a developer preview in the Gemini Live API, an enterprise preview in Google Meet this month, and eventually in the Translate app on Android and iOS. There's also a new Android listening mode where you hold the phone to your ear like a call, no earbuds required. All audio output carries a synth ID watermark, Google's invisible tag that identifies AI-generated audio. For international business calls, cross-border enterprise meetings, and any workflow with language barriers, the applications are immediate. Real-time translation that sounds like the actual speaker is new territory. Speaking of eliminating awkward pauses, Google DeepMind just unveiled a totally new way to build AI models. They basically took the mechanics of an image generator, slapped it into a language model, and suddenly traditional text generation looks painfully slow. Every mainstream LLM you've used, ChatGPT Claude Gemini, generates text one token at a time, left to right, each word depending on everything that came before. Fast enough for most purposes, but the architecture has a hard latency floor. Diffusion Gemma works differently. Instead of building text word by word, it starts from noise, literally random data, and iteratively refines an entire block of text until something coherent emerges. The same way image diffusion models like Midjourney produce pictures. Because it generates in parallel rather than sequentially, it claims to run four times faster than an equivalent auto-regressive model. Google's benchmark shows 1,000 tokens per second on a single H100 GPU. It ships as open weights under an Apache 2.0 license, meaning anyone can use or build on it commercially with day one support in Hugging Face Transformers, VLLM, and UnSloth, the main open source frameworks for running AI models locally without cloud costs. It's explicitly labeled experimental, so production use will require more testing, but for agentic workflows, automated AI pipelines that need to produce large amounts of text quickly, this architecture is worth watching. So the models are getting faster, cheaper, and running complex workflows entirely on their own, which is fantastic for productivity, but according to a new IBM survey, it is absolutely terrifying the enterprise IT executives who actually have to govern this stuff. Two-thirds of CIOs and CTOs say they are accountable for AI systems they don't fully control. Only 11% say they feel prepared for large-scale AI deployment, and the gap is about to get wider. AI agent use across enterprises is expected to rise sharply over the next 12 months. That accountability without control dynamic is the core tension in enterprise AI right now. You're the one answering to the CEO when something goes wrong with the AI-powered customer support bot, the code review agent, or the procurement tool, but the model is a third-party API, the workflow was built by a contractor, and the governance policy is still a draft someone's working on. The survey found most governance frameworks simply aren't ready for the velocity at which agents are being deployed. For anyone in IT leadership, this is the thing to get ahead of. Governance frameworks built for static software don't translate cleanly to AI systems that can reason, take actions, and behave differently across contexts. The 11% who feel prepared probably have one thing in common. They started on the governance work before deployment, not after. Finally, while enterprise CTOs are scrambling to figure out how to babysit rogue AI agents, OpenAI is actively asking world leaders to step in and regulate how AI interacts with actual children. OpenAI published a proposal last week calling for a new International Youth AI Safety Institute, timed to the G7 Leaders Summit happening in Avion, France later this month. The proposal calls for privacy preserving age estimation, mandatory annual youth safety risk assessments, parental controls, and specific protocols for high-risk interactions, things like self-harm, exploitation, and grooming scenarios. OpenAI says the institute could be a new standalone international organization or could expand the mandate of an existing national AI body to take on a global role. The G7 timing is deliberate. AI governance has been moving from broad principles, we support responsible AI, toward domain-specific safety regimes, kids, elections, weapons, healthcare. Youth use as a distinct regulatory category means product requirements, audit expectations, and cross-border compliance standards specific to platforms that serve minors. OpenAI's push here is partly about shaping what those standards look like before legislators write them without input from the industry. Estonia's national ChatGPT rollout for students gets cited in the proposal as a reference case. Whether the G7 acts on this or not, the proposal is a signal that youth-specific AI regulation is coming, and companies should be thinking about it now. Just a couple of more items. If you have any feedback about this show, you can email Mike at yesterdayanaai.news, or you can find me on LinkedIn, X or Blue Sky. And if you like this podcast and want to see it continue, please take a minute to rate and review it so others can find it. Thanks. That's all for this edition of Yesterday and AI. Stay curious, and I'll see you tomorrow.