Latest / UpNext AI / Anthropic’s Policy Reversal, Claude Science, and Agentic Persuasion Tests | UpNext AI – July 1, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Wednesday, July 1st, 2026, and here's what matters in AI today. First, the U.S. has dropped restrictions on Anthropic’s Mythos and Fable models. According to TechCrunch, the government lifted a requirement that Anthropic obtain a license before exporting those models abroad. That rule had effectively cut off public access to what the article describes as some of the most advanced AI models released so far. Anthropic said it would begin restoring access on Wednesday, July 1st. The earlier restriction had been added on June 12th, and TechCrunch reports that complying with it at scale proved impractical, which pushed Anthropic to end public access altogether. The bigger takeaway here is not just that access is back. It’s that the rules still look unsettled. TechCrunch says the Trump administration’s approach has left companies with little clarity on what will govern future model releases. The reporting also says Anthropic agreed to proactively detect and address security risks tied to the models, work with the U.S. government on protocols and standards for Mythos, Fable, and future releases, and inform the government of malicious activity. There’s also a competitive angle: TechCrunch notes that AI companies in Asia were beginning to release models approaching Mythos-level capability, adding pressure to ease restrictions so U.S. firms could still compete globally. So for now, the practical result is simple: access is being restored. But the broader policy environment for frontier model releases still looks fluid. Next, Anthropic is also making a very different kind of bet with Claude Science. This one is not a new model. As TechCrunch reports, Claude Science is a workbench for scientists that runs the same Claude models already available today, including Claude Opus 4.8. The idea is to give researchers one place to do computational work instead of bouncing between databases, pipelines, and separate tools. Anthropic’s pitch here is really about workflow. One main assistant acts like a project manager, connects to more than 60 scientific databases, and comes with prebuilt toolkits for areas including genomics, protein structure, and chemistry. That assistant can spin up sub-assistants for specialized tasks, and there’s also a fact-checking step that checks citations and calculations before work moves toward publication. TechCrunch notes an important caveat there: it’s still the same underlying model checking itself, not an independent source of truth. Anthropic also says the system is designed for reproducibility, including figures generated alongside the code that produced them, a plain-language description, and the full message history. Another practical detail: Anthropic says Claude Science can run on a lab’s own infrastructure rather than sending data to Anthropic’s servers. Claude Science is available in beta for Pro, Max, Team, and Enterprise subscribers, and Anthropic says it will support up to 50 projects with as much as 30 thousand dollars in credits. The broader point is that frontier labs are not just competing on model quality anymore. They’re competing to become the operating layer for specialized work. For the research section, a paper posted to arXiv yesterday argues that we’re testing social reasoning in AI the wrong way. The paper is titled “Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action.” The authors say most Theory of Mind benchmarks for large language models rely on passive question answering. But if models are going to act more like agents, that misses an important capability: whether they can use planning and actions to shape what another agent believes. In plain English, the paper says you can’t judge this kind of social reasoning just by asking models questions. You need to see whether they can actually carry out a plan that changes someone else’s beliefs. In the setup described in the paper, models were given belief-state goals and had to move objects or direct characters into rooms to achieve them. The researchers say they evaluated six frontier models across 600 task instances, alongside human participants. GPT-5 succeeded on about 80 percent of tasks in this agentic setting and was the only model to outperform human participants on the task, though it was still less robust than humans across contexts. The authors also found that both models and humans did better when the goal was to induce true belief states rather than false ones, which they frame as a positive alignment signal. Bottom line: if AI agents are going to act in the world, we need to measure not just what they can say, but what beliefs they can cause. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. In fact, you're listening to one of their voices - right - now. If you are a developer looking to elevate user experience with natural voice interfaces, this is your solution. Visit up next dot fm slash eleven to check out their latest offerings. ... Wayve has launched an 85 million dollar employee tender offer at an 8.5 billion dollar valuation. TechCrunch reports that the self-driving startup is giving employees a structured way to sell vested equity, part of a broader trend of AI companies using tender offers as a retention tool in a very competitive talent market. And Meta is adding usage limits and a soft paywall to some smart-glasses AI features. According to The Verge, the Conversation Focus feature will be limited to three hours per month unless users pay for Meta One Premium at 19.99 a month, and even premium subscribers will still face a higher monthly cap of 15 hours. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. If you enjoyed this episode, don't forget to subscribe, rate, and leave us a review! And that's your briefing for today. Full source links are in the episode notes, and we'll be back tomorrow with what's up next!