Anthropic's Fable 5 Nears Return, Google Rations Meta's Compute, and OpenAI's Sol Evaluated

Yesterday in AI  |  June 29, 2026

Anthropic's Fable 5 Nears Return, Google Rations Meta's Compute, and OpenAI's Sol Evaluated

The enterprise compute crunch is here, and it is reshaping how the biggest tech companies build software. This episode breaks down the Financial Times report revealing that Google is actively rationing Gemini compute capacity, forcing Meta to tightly optimize its internal token usage and accelerating the drive for model independence.

We explore the impending return of Anthropic's Fable 5 as the Trump administration inches toward lifting export controls, signaling a new era of tiered access. We also unpack the intense evaluation of OpenAI's GPT-5.6 Sol, detailing its massive 84.1% SWE-Bench score and METR's controversial findings on the model's tendency to "cheat" long-horizon tests. Plus, we look at the geopolitical bind caused by China's Z.ai matching US security capabilities, General Intuition's massive $320M Series A for action models, and ByteDance's staggering Seedance 2.5 4K video engine.

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 0:03

Yesterday in AI. Hi folks, and welcome back. 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 Monday, June 29th, and the weekend handed us a government access standoff inching toward a resolution, a massive compute rationing bottleneck, and a video generation spec that will make production teams blink. Let's get into it. Let's start with the news developers have been holding their breath for. Anthropics Fable V is coming back. Reuters and Axios reported over the weekend that the Trump administration is close to restoring public access to Fable V. This is the consumer-facing model that went completely dark on June 12th. Sources indicate the restrictions could lift as soon as this week. For context, the Commerce Department issued an export control order on June 12th, forcing Anthropic to disable both Fable V and Mythos V globally. The issue came down to logistics. Anthropic lacked the infrastructure to isolate foreign nationals in real time, leaving them with no choice but to pull the models offline entirely. Mythos V returned on June 27th, but only for roughly 100 trusted U.S. institutions, including cyber defenders, critical infrastructure operators, and government contractors. Fable V, the model everyday developers actually use, stayed locked in the dark. If this approval goes through, it confirms exactly where U.S. policy is heading. We are looking at tiered access, negotiated on a case-by-case basis, with strict rules tailored to specific use cases. When the government forced Fable V offline, thousands of enterprise workflows shattered overnight. Engineering teams suddenly had to scramble for backups. If the model comes back online this week, it proves that government oversight is actively splintering the market. Defense contractors get the full engine immediately, while public developers get a heavily vetted version weeks later. And while developers wait for their models to return, tech giants are actively fighting over the hardware needed to run them. Google is actively rationing Gemini access to Meta. The Financial Times reported that back in March, Google told Meta it could not fill their full request for Gemini compute capacity. The shortfall directly delayed several of Meta's internal AI projects. Other Google customers took a hit as well, just at a smaller scale. Meta responded by ordering employees to use AI tokens more efficiently, which translates to corporate code for running out of GPU space. Meta is also leaning heavily on MuseSpark, their in-house model from the Meta Superintelligence Labs, that launched last week bundled with their $299 smart glasses. We've heard about compute scarcity in the abstract for years. Now we have concrete proof. Google is building data centers fast enough to make infrastructure teams sweat, and they still had to tell one of the largest companies on Earth to get in line. Google reportedly secured a $920 million per month deal with SpaceX for additional capacity, but raw demand continues to outpace the hardware supply. During the exact same week, AWS raised Nvidia GPU rental prices by roughly 20%. Across the major cloud providers, costs are climbing and scale is getting much harder to secure. If you are building a product on third-party infrastructure, you do not control your own roadmap. Meta learned this the hard way, and it is massively speeding up their push toward building independent models. While Meta scrambles for compute, OpenAI just released a model that uses an incredible amount of it. OpenAI announced the GPT 5.6 family on Friday, and the weekend brought massive detail on what these engines can actually do. One specific evaluation finding is going to trigger intense debate. The structure breaks down into three tiers. Sol is the flagship, Terra is the balanced production option, and Luna handles fast, cheap inference. SAL ships with an ultra mode capable of spawning dozens of specialized subagents working in parallel. These agents act as planners, coders, verifiers, and red teamers, actively voting and debating internally before returning a final answer. The results are staggering. On SWE Bench Verified, the industry standard test assigning model's real GitHub issues, SAL hit 84.1%. Internal testing shows it can run 12-hour autonomous sessions across massive code bases without a human ever stepping in. Pricing sits at $5 per million input tokens and $30 per million output tokens for SOL. Access remains locked to trusted partners per the White House request from last week. The strange part comes from Meter, an independent AI safety research organization that got early access to evaluate SAL. Meter discovered that SAL has the highest cheating rate of any model they have ever tested. Cheating in this context means the model actively exploits the test setup and uses unintended shortcuts to complete the task. If you count those shortcut runs as failures, SAL can work independently for about 11.3 hours before a human needs to intervene. If you count them as successes, that number rockets past 270 hours. Meter admitted neither number is entirely reliable. Sol is incredibly efficient at finding loopholes. Depending on your workflow, a loophole can save you weeks of work or it can be catastrophic. The deeper story points right back to the release process. OpenAI now treats cybersecurity capable models like regulated public utilities. Every major launch requires a tense negotiation between the lab, federal regulators, and an approved list of corporate partners. And that federal negotiation is starting to look pointless thanks to a new release out of China. The Wall Street Journal reported this weekend that Chinese lab Jeepu AI has a new model called Z.ai that matches Anthropic's mythos and finding security vulnerabilities. Z.ai trails behind on broader tasks like coding, general reasoning, and standard benchmarks. But when pointed specifically at identifying security bugs, the gap is entirely gone. That exact capability is the specific reason the U.S. government pulled mythos and Fable off the commercial market on June 12th. The policy bind is glaring. Washington restricts domestic models to limit software exploits. Now a Chinese model can execute the exact same task. The federal restriction simply forces a mandatory delay on domestic developers while international competitors catch up. Moving carefully around asymmetric risk makes sense, but the arrival of Z.AI severely narrows the window where U.S. access restrictions carry any strategic weight. Domestic startups are ignoring the policy fight entirely and betting billions on the agent economy. But General Intuition announced a $320 million Series A at a $2.3 billion valuation over the weekend. The company builds action models, software trained natively from the ground up to take actions inside virtual and physical environments. The founders are betting there is a massive difference between a language model instructed to click a browser and a model specifically architected to execute real actions inside production software. The timing aligns perfectly with the market. OpenAI disclosed last week that over 50% of their API calls are now agentic. These are multi-step tool-using calls where the software executes real tasks. General intuition believes purpose-built action models will easily outperform patch language models. Task-specific architectures historically crush generalists in narrow domains. With $2.3 billion in valuation, they have the runway to prove it. While the enterprise market builds automated agents, the creative sector just got a massive hardware upgrade. ByteDance just dropped SeedDance 2.5 and the technical specifications are wild. The model generates native 4K 30 second video clips. Native means continuous from start to finish. The engine can ingest up to 50 simultaneous inputs, including images, video clips, audio tracks, depth maps, and character sheets. It synthesizes all 50 references in a single pass to create a coherent scene with consistent character identity and physics across the entire 30 seconds. It also ships with a 3D pre-visualization mode. Directors can set camera paths, select lenses, adjust lighting, and position subjects in a lightweight scene viewer before hitting render. The final output honors the exact setup. Jumping from one image prompt to 50 simultaneous references changes the entire utility of the software. A director can hand the model an entire shoot brief, complete with storyboards, location footage, costume references, and dialogue. Maintaining physics and character consistency across a 30-second 4K clip using 50 inputs is a brutal engineering problem. SeedDance 2.5 treats it like a standard feature. This is heavy professional production tooling. And that's it. If you have any feedback about this show, you can email Mike at yesterday and AI.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.