US Suspends Anthropic Models Over Security as Rio Government Debuts Open-Source Giant
Government & Regulatory Actions
US Government Orders Suspension of Anthropic’s Fable 5 and Mythos 5 Models: The US government issued an emergency directive forcing Anthropic to suspend global access to its Fable 5 and Mythos 5 models due to national security concerns regarding a specific jailbreak vulnerability. Anthropic has complied with the order while disputing the severity of the risk and noting that the models' capabilities are comparable to existing systems like GPT-5.5.
- US government directive to suspend access to Fable 5 and Mythos 5
- Anthropic is suspending access to Fabel/Mythos for ALL users, not just non-Americans
- Anthropic forced to abruptly disable Fable 5 & Mythos 5 globally by US Gov over a jailbreak. This is exactly why we need local models.
- Statement on the US government directive to suspend access to Fable 5 and Mythos 5
- Anthropic says it’s complying with US government order to suspend Fable 5 and Mythos 5 access over jailbreak concerns
New Model Releases & Benchmarks
Rio 3.5 Open 397B Released by Rio de Janeiro Government: The city government of Rio de Janeiro has released Rio 3.5 Open 397B, an open-source fine-tune of Qwen 3.7 Plus that utilizes "SwiReasoning" for dynamic chain-of-thought capabilities. Benchmarks show the model outperforms its base version and competes effectively against leading models like DeepSeek V4 Pro and Kimi-K2.6.
Google’s Diffusion Gemma Prioritizes Speed Over Accuracy: Benchmark results for Google’s new Diffusion Gemma model reveal it is four times faster than the autoregressive Gemma4 but produces six times as many factual errors. The 26B model struggled particularly with niche topics, highlighting a significant trade-off between generation speed and factual reliability.
GLM-5.2 Open-Weight Model Set for Next Week: A new open-weight model, GLM-5.2, is scheduled for release under the MIT license next week, continuing the trend of high-performance releases from Chinese AI labs. While official benchmarks are still pending, early indicators suggest it aims to compete with top-tier frontier models like GPT-5.5.
AI reasoning Shows Massive Gains in Complex Mathematics: Data from Epoch AI highlights a rapid leap in AI performance, with models moving from solving almost no elite-level math problems to achieving high accuracy in just one year. OpenAI's GPT-5.5 currently leads the benchmark with ~75% accuracy on FrontierMath Tier 4, followed by Anthropic's Fable 5 at ~50%.
Model Capabilities & Use Cases
Claude Fable 5 Demonstrates Advanced Coding by Generating Minecraft Clone: A user successfully utilized Claude Fable 5 via the Computer platform to build and deploy a functional HTML-based Minecraft clone in just a few prompts. The project required the model to generate nearly 1,000 lines of code, showcasing its high-level orchestration and complex task-solving abilities.