Fable 5 Goes Dark, Rio's 397B Open Model, and KPMG's AI-Written AI Report — Digest for June 14, 2026

Five items for builders today: Anthropic's Fable 5 and Mythos 5 are completely offline after a US Commerce Department export-control order citing a potential jailbreak; a separate story reveals Anthropic had been silently degrading the model for AI researchers before the ban hit; Meta is operationally dismantling its $2B Manus acquisition after Beijing ordered the deal reversed; a city government in Brazil released a 397B open-weight model on HuggingFace with striking benchmarks that haven't been independently verified yet; and KPMG had to pull a report on AI usage after researchers found it was riddled with AI hallucinations about clients who never said what the report claimed.

On the evening of June 13, Anthropic disabled access to both Claude Fable 5 and Mythos 5 for all users worldwide. The trigger was a directive from the US Commerce Department subjecting the models to export controls — rules that restrict their use outside the United States. 1 Anthropic said the only way to comply in the short term was to shut the models off entirely, including for US customers.

The Commerce Department's concern, according to an Axios report citing an administration official, was a reported jailbreak that can get around Fable 5's classifier-based safety blocks on cybersecurity, chemistry, and biology prompts. 2 The administration reportedly wants a pause while the "national security apparatus" is "hardened," which the official suggested could take weeks.

Anthropic pushed back on the framing. In its announcement, the company said the government had provided only "verbal evidence of a potential narrow, non-universal jailbreak" limited to reviewing specific codebases for software flaws. Anthropic noted it had only seen evidence of the jailbreak being used to find "minor" and "relatively simple" software vulnerabilities, and said other publicly available models such as GPT-5.5 have similar capabilities. 3

This is the first time a US government order has forced a leading AI lab to take a live deployed model offline. Access to Anthropic's other models — Claude Opus, Sonnet, and Haiku — is not affected.

What it means for builders: Any project using claude-fable-5 or claude-mythos-5 API identifiers is broken until the ban lifts. Plan around the fallback models Anthropic still offers, and watch for any announcement of partial restoration.

Separate from the export ban — and revealed the same week — Anthropic disclosed on June 11 that Claude Fable 5 had been quietly routing certain requests to a weaker model or rejecting them outright, without telling users. The affected tasks included prompts related to training competing large language models, debugging AI code, and optimizing neural network architectures. 4

Users were consuming tokens without receiving the expected quality of output. Anthropic had not disclosed this behavior in the model documentation at launch.

After researchers flagged it publicly, Anthropic acknowledged having "made a wrong trade-off" and apologized. The company changed the mechanism: when the system determines a user might be building a high-capability AI, it now shows an explicit warning before rejecting or redirecting the request. The underlying restriction was not removed — only the silent part was.

The practical upshot: researchers using Fable 5 for model training or architecture search tasks can now at least see warnings before the system silently burns their budget. But for those tasks, they'll still need to switch to a different provider or model. The incident also reframed how the export-control ban landed — Fable 5 was already less capable for AI development tasks than its headline benchmarks implied.

Meta has severed Manus from its internal systems, stopped employees from using Manus tools on internal projects, and halted data sharing between the two companies. 5 This is the most concrete enforcement step yet on Beijing's order — issued roughly two months ago — requiring Meta to unwind its $2 billion acquisition of the Chinese-founded AI startup.

The Manus co-founders have separately held preliminary talks about raising around $1 billion from outside investors to reclaim the startup from Meta, which could lead to a Chinese joint-venture structure and an eventual Hong Kong listing. 6

The divestiture order adds context to a broader tightening of Chinese tech policy: Beijing has also expanded travel restrictions for researchers and executives at private AI firms, and is reportedly requiring government approval before major AI companies — including Moonshot AI, StepFun, and ByteDance — accept US investment. 7

Manus has continued shipping features — including Similarweb and Shopify integrations — through all of this. The core issue for builders: if you've built on Manus' agentic API, the underlying ownership situation is in flux, and you should watch for any service continuity announcements as the separation proceeds.

The municipal IT company of Rio de Janeiro — IplanRIO — has published a 397-billion-parameter open-weight model on HuggingFace under the handle prefeitura-rio/Rio-3.5-Open-397B. 8 The repository is real and downloadable: 97 safetensor shards totaling roughly 807 GB, licensed MIT.

The model is a post-trained derivative of Qwen3.5-397B-A17B. That base model is a sparse Mixture-of-Experts architecture — meaning the full 397B parameters exist in storage, but only about 17B activate per token, which keeps per-token compute roughly comparable to a mid-size dense model. 9 The model also claims multimodal input (image-text-to-text) and a 1-million-token context window, though the config file shows 262K as the native position limit — the 1M figure depends on serving-mode overrides.

The benchmark claims are large: 80.2 on SWE-Bench Verified, 70.8 on Terminal-Bench 2.1, 58.1 on SWE-Bench Pro, and 90.9 on GPQA Diamond. The catch: every single one of these numbers comes only from Rio's own model card as of writing. No independent benchmark host or third party has reproduced them yet. A HuggingFace discussion thread clarifies that the headline scores depend on a "SwiReasoning" latent-reasoning inference path — and that inference engines limited to standard token generation (including llama.cpp) cannot implement it, so results will vary significantly by stack.

The model's origin is what makes it unusual. Frontier-scale open-weight releases almost exclusively come from AI labs — Qwen, DeepSeek, Meta, Mistral, MiniMax. A city government's technology subsidiary publishing something in this weight class is different. Whether Rio's post-training has genuinely improved on the Qwen3.5 base, or whether it's a thin wrap around existing weights, will become clear once community benchmarking catches up.

Practical: MIT license, 807 GB download, BF16 weights. Deployable via Transformers, vLLM, or SGLang. Treat the benchmark claims as first-party until independently reproduced.

The HuggingFace repository is at prefeitura-rio/Rio-3.5-Open-397B.

KPMG has removed its report "Redefining excellence in the age of agentic AI" from its websites after multiple organizations named in it said the claims about their AI usage were simply not true. 10 UBS, the UK's National Health Service, Swiss Federal Railways, and Transport for London all told the FT that the report's description of their AI deployments was either false or misleading.

The research group GPTZero identified the inaccuracies as AI hallucinations — meaning KPMG appears to have used AI to help write a report about AI, without adequate human review of the claims. The report was originally published in October 2025.

A KPMG spokesperson said the firm removed the report while conducting its own investigation, and noted that its guidelines require "human oversight to validate content and verify independent sources." This follows a similar episode last month in which EY withdrew a loyalty rewards report that included fake footnotes and AI hallucinations.

For builders, the pattern is worth noting: AI summarizing information about specific companies' internal practices is a reliability minefield, because the model has no direct access to private operations and will confidently fill gaps. Any report using AI to make specific claims about named organizations needs human verification against primary sources before publication.