# Anthropic Shutdown Order Hits Fable as Open Models Advance *By AI High Signal Digest • June 13, 2026* A U.S. export-control directive forced Anthropic to disable Fable 5 and Mythos 5 for all customers, just as Fable was posting leading benchmark results. Meanwhile, MiniMax and Moonshot pushed the open-model wave forward, and new work on agentic inference and automated discovery added technical signal. ## Top Stories *Why it matters: the clearest signal today was that model capability and model availability are now separate stories.* - **Anthropic had to shut off Fable 5 and Mythos 5.** Anthropic said a U.S. export-control directive suspended access for foreign nationals, including foreign-national employees, and that the practical result was disabling both models for all customers; other Claude models remain available, and Anthropic says it is working to restore access. Across Claude products, new sessions now use the selected default model or Opus 4.8, existing Fable 5 sessions end with errors, and platform requests to Fable 5 also return errors. [^1][^2] - **The shutdown landed just after Fable 5 posted some of the week’s strongest benchmark results.** It ranked #1 on FrontierSWE, which the benchmark authors called the biggest capability jump since release, and on many tasks it reportedly worked productively for close to 20 hours. Epoch separately reported 87% on FrontierMath Tiers 1–3 and 88% on Tier 4, while Artificial Analysis’s refreshed Coding Agent Index put Claude Code with Fable 5 at the top with a score of 77. [^3][^4][^5] - **Open models kept advancing at the same time.** MiniMax released M3 as an open-weight model with ~428B parameters (~23B activated), 1M context, native text/image/video input, and frontier coding and agentic positioning. Moonshot open-sourced Kimi-K2.7-Code with gains of +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, +31.5% on MLS Bench Lite, and 30% lower reasoning-token use than K2.6. [^6][^7][^8] ## Research & Innovation *Why it matters: the most useful research updates were about making agents cheaper to run, easier to evaluate, and more autonomous in discovery.* - **AA-AgentPerf** introduced a benchmark built around real long-context agentic coding workloads and production optimizations such as KV cache reuse and speculative decoding. Its first results show 61,354 Agents/MW on rack-scale GB300 versus 21,053 on B300, with GB300 about 3x more power-efficient than single-node Blackwell in this test. [^9] - **Recursive** unveiled an early automated discovery system, described as a milestone toward recursive self-improving superintelligence, and said it reached state-of-the-art results on NanoGPT speedrun, NanoChat, and NVIDIA’s Sol-ExecBench using ideas and code generated by the system itself. [^10] - **RandOpt** claims that adding Gaussian noise to pretrained LLMs and ensembling the results can match or beat PPO/GRPO on math reasoning, coding, writing, and chemistry across Qwen, Llama, OLMo3, and VLMs, without gradient-based training. [^11] ## Products & Launches *Why it matters: product releases kept pushing AI deeper into developer workflows, deployment infrastructure, and media generation.* - **OpenAI launched a docs agent** on developers.openai.com that answers questions about OpenAI products, links directly to relevant documentation, and can generate custom guides that open in Codex or export as Markdown for coding agents. [^12][^13] - **Claude Managed Agents** now support user-controlled sandboxes on a company’s own infrastructure or any chosen provider, with new setup guides for multiple partners. [^14][^15] - **Zyphra released ZONOS2**, a real-time TTS model with high-fidelity voice cloning that the company describes as the first open-source sparse MoE TTS system; it uses raw UTF-8 input and is released under Apache 2.0. [^16][^17][^18][^19] ## Industry Moves *Why it matters: partnerships and commercialization moves are increasingly shaping where AI gets deployed first.* - **Google DeepMind launched a Robotics Accelerator** with 15 European startups, offering access to its AI stack, Gemini Robotics models, and hands-on support over a three-month program. [^20] - **Cognition is putting money behind Devin adoption.** The company said its AI Productivity Guarantee is a $10,000,000 bet that Devin delivers engineering value. [^21] - **Baseten and Harvey are partnering on legal AI.** Baseten said the two companies are pushing open models toward frontier-level performance for legal applications and that LAB was central to their joint post-training work on legal agents. [^22] ## Policy & Regulation *Why it matters: this was not a normal product outage; it was a compliance event that immediately changed what customers and employees could access.* Anthropic said the directive suspended access to Fable 5 and Mythos 5 by any foreign national, including foreign-national employees, forcing the company to disable both models for all customers to comply. Other Claude models remain available, but developers are being told to move integrations off Fable 5 while Anthropic works to restore access. [^1][^2] ## Quick Takes *Why it matters: these smaller updates still add signal on benchmarks, tooling, and evaluation.* - FrontierMath Tiers 1–4 (v2) went live after an audit fixed errors in 42% of problems; GPT-5.5 (xhigh) now leads Tiers 1–3 at 85%, and Google’s AI co-mathematician leads Tier 4 at 76%. [^23] - Google Research’s **Gemini-SQL2** hit state-of-the-art results on the BIRD text-to-SQL benchmark; one team member cited a 2.5% improvement over the previous single-model SOTA. [^24][^25] - Ramp launched **Ramp SWE-Bench**, a private, production-grounded coding benchmark built from real internal engineering problems. [^26] - OpenAI is rolling out **Codex rate-limit reset banking**, starting Go, Plus, Pro, and Business users with one free reset. [^27] --- ### Sources [^1]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2065597531644743999) [^2]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2065597942602531163) [^3]: [𝕏 post by @ProximalHQ](https://x.com/ProximalHQ/status/2065184730279223410) [^4]: [𝕏 post by @EpochAIResearch](https://x.com/EpochAIResearch/status/2065511916035018943) [^5]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2065328920514515037) [^6]: [𝕏 post by @MiniMax_AI](https://x.com/MiniMax_AI/status/2065436935188058208) [^7]: [𝕏 post by @vllm_project](https://x.com/vllm_project/status/2065445059039031799) [^8]: [𝕏 post by @Kimi_Moonshot](https://x.com/Kimi_Moonshot/status/2065377579130142937) [^9]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2065559824230957190) [^10]: [𝕏 post by @RichardSocher](https://x.com/RichardSocher/status/2065094362774876232) [^11]: [𝕏 post by @yule_gan](https://x.com/yule_gan/status/2032482266773926281) [^12]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2065507724704858173) [^13]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2065507798268674366) [^14]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2065494480837583297) [^15]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2065494482460836183) [^16]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2065498869954490846) [^17]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2065498871728574820) [^18]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2065498881820156112) [^19]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2065498886521979222) [^20]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2065388989146628563) [^21]: [𝕏 post by @cognition](https://x.com/cognition/status/2065466791430615516) [^22]: [𝕏 post by @baseten](https://x.com/baseten/status/2065534055391658381) [^23]: [𝕏 post by @EpochAIResearch](https://x.com/EpochAIResearch/status/2065488154086568445) [^24]: [𝕏 post by @GoogleResearch](https://x.com/GoogleResearch/status/2065475343205740911) [^25]: [𝕏 post by @mirrokni](https://x.com/mirrokni/status/2065486426993553657) [^26]: [𝕏 post by @RampLabs](https://x.com/RampLabs/status/2065485806605619304) [^27]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2065225362544726371)