Model Halts, Agentic Git, and Open Reasoning

Anthropic switched off Fable 5 and Mythos 5 worldwide after a US Commerce Department order barred all foreign nationals from the models. The trigger, as reported by Fortune and others, was a warning from Amazon CEO Andy Jassy that Amazon researchers had got Fable 5, the safeguarded version of Mythos 5, to help carry out cyberattacks. That account is contested: some reporting frames the research as defensive and questions whether it was a true "jailbreak".

The detail: The letter reached CEO Dario Amodei on Friday 12 June at 5:21pm ET with about 90 minutes to comply, and it is the first US export control aimed at an AI model rather than at chips. In its own statement, Anthropic called the move an overreaction, describing the flaws as "minor" and "relatively simple" and no worse than those of other public models. The pair had launched days earlier at under half Anthropic's prior token prices, and Microsoft had just blocked staff from using Fable 5. Amazon is both an investor in Anthropic and one of its cloud hosts.

Related: Anthropic says the Mythos line can find and exploit software flaws better than human experts; its Project Glasswing flagged more than 10,000 high- or critical-severity vulnerabilities, 1,726 of them confirmed true positives. The White House summoned Amodei in April; the Pentagon demanded access to Claude Gov, then designated Anthropic a supply-chain risk; and Trump barred federal agencies from using the company.

GitHub co-founder Scott Chacon used AI coding agents to rebuild Git from scratch in Rust, a project called Grit.

The detail: By Chacon's account, agents in Claude Code, Cursor, and Codex worked over about two to three weeks to produce more than 360,000 lines across 500+ pull requests and 7,000+ commits, at a cost of $10-15k in tokens (roughly 45 billion), with the result passing 99.3% of Git's 42,000-test suite. Grit is MIT-licensed, library-first and memory-safe, aimed at embedding Git in editors or agent runtimes and at WASM edge builds. He adds that the agents "cheated" by shelling out to the real Git, that parallel agents corrupted shared state, and that it is not production-ready.

Related: Other recent agent-written code includes Stripe's "Minions", which merge 1,000+ pull requests a week with no human-written code, and Reco.ai's AI rebuild of its JSONata engine for a reported 1,000x speedup; one analysis warns the output is creating "verification debt".

Hugging Face published open-r1, a fully open reproduction of DeepSeek-R1's reasoning. A reproduction means it rebuilt the training data and recipe, not just the weights, so others can retrain a similar model rather than only run DeepSeek's.

The detail: The release includes Mixture-of-Thoughts, a 350,000-example dataset of step-by-step reasoning traces distilled from R1 across maths, code, and science, and OpenR1-Distill-7B, a 7B model fine-tuned on that data from Qwen2.5-Math-7B. Hugging Face says it reproduces DeepSeek's distilled 7B on reasoning benchmarks such as AIME 2024 and MATH-500.

Related: It follows other open releases around DeepSeek's models: V4 matched frontier quality at lower cost, V4 Pro outscored GPT-5.5 Pro on precision, and GuppyLM showed a full training recipe in the open.

---