Export controls lifted: Anthropic redeploys Claude Fable 5
Anthropic’s June 30 post fills in why Fable 5 vanished for almost three weeks. The model shipped June 9 alongside Claude Mythos 5; on June 12 the US government put export controls on both, requiring restrictions on foreign nationals’ access. Anthropic couldn’t verify nationality in real time, so it suspended both models for all users. The government cleared Mythos 5 for US organizations on June 26, lifted the controls on June 30, and Fable 5 came back globally on July 1. The trigger: Amazon researchers found a jailbreak, a prompting technique that bypasses a model’s safeguards, which got Fable 5 to identify software vulnerabilities and demonstrate how to exploit them. Note the framing: Anthropic describes this as a bypass reported by researchers, the post mentions no abuse in the wild, and Anthropic says its own testing showed weaker models could produce similar output. The redeploy came with a targeted classifier that Anthropic says blocks the specific technique in over 99% of cases, plus a deliberate safety margin that also catches ambiguous, probably-benign requests. My read: the binding constraint wasn’t the technical fix but compliance. Anthropic’s own wording, that it had “no reliable way to verify nationality in real-time”, is what actually kept the model offline for three weeks.
Claude Sonnet 5: near-flagship ability at an introductory discount
Same day, June 30: Anthropic released Claude Sonnet 5, billed as its most agentic Sonnet, meaning it plans, uses tools, and runs multi-step tasks on its own. The company says that at higher effort settings it approaches Opus 4.8 on some complex tasks. The pricing is the story: $2 per million input tokens and $10 per million output through August 31, then $3/$15, and Free and Pro users get it by default. A mid-tier model closing in on flagship ability, sold at a temporary discount, shows where the competition among frontier labs now sits: the mid-price tier, fought on price. For application builders, it’s time for another round of model evals.
Nathan Lambert’s open-model roundup: the consolidation prediction failed
In Interconnects’ 23rd open-artifacts roundup, Lambert argues the prediction that open models would consolidate to a few labs has not held up: more companies are training strong models and releasing weights, because token demand is high and running a token factory is profitable. This issue covers Poolside’s Laguna S2.1 (a 118B-parameter mixture-of-experts model with 8B active per token, OpenMDW license), Thinking Machines’ Inkling (975B-A41B, multimodal, positioned as a fine-tuning base), Moonshot AI’s Kimi K3 (its license allows commercial use, deployment, and fine-tuning by default; a separate agreement is required only for model-as-a-service providers whose revenue exceeds $20 million over twelve months), and Meituan’s LongCat-2.0 (1.6 trillion parameters, trained end to end on Chinese hardware, per the post). Two signals stand out to me. One is who is on the list, and it needs the right baseline: open weights began as a largely American story, with Meta’s Llama family defining the category through 2024. “Open weights means China” only became the default read after DeepSeek-R1 landed in January 2025 and, by Lambert’s own accounting, Chinese models overtook the US’s early Llama-built lead in adoption that summer; he spent much of last year on the ATOM Project, arguing the US had ceded open models to China. Against that baseline, Poolside and Thinking Machines appearing next to Moonshot and Meituan reads as US companies re-entering a field they started in, not arriving for the first time. The other signal is license divergence: OpenMDW allows free use for any purpose, while Kimi K3’s terms attach revenue-triggered conditions, so two models both called “open” can permit different things.
PsychAdapter writes personality into model weights instead of prompting for it
A team from Stanford, Stony Brook, and other institutions published PsychAdapter in npj Artificial Intelligence: adapters inside the transformer take psychological trait scores as direct inputs and shape generation at every layer, covering the Big Five personality dimensions (openness, conscientiousness, extraversion, agreeableness, neuroticism) plus depression and life satisfaction, optionally conditioned on demographics like age. The difference from prompt engineering: the trait expression is trained on empirical links between language use and psychological traits, while a prompt just asks the model to act out a persona from its own assumptions. Expert raters judged the generated text consistent with the target traits 87.3% of the time for Big Five and 96.7% for depression and life satisfaction; code is on GitHub, and the authors point to simulated patients for therapist training and reading-level-adjusted educational text as applications (see also Stanford HAI’s writeup). The other edge is my own speculation, not a claim the paper makes: the same machinery looks like it would make convincingly human companion AI cheaper to build, a use the authors don’t discuss.
Stanford HAI: states split into two camps on AI job loss
The piece maps state-level policy on AI-driven job loss into two camps: one treats displacement as inevitable and wants monitoring first, with retraining and income support as the response; the other wants government to shape how AI is used and to keep critical jobs human-performed. Concrete tools include a proposed robot tax on businesses that automate jobs in New York and skills-training programs elsewhere. What California has actually shipped is data infrastructure: an AI-Unemployment Tracker, built with the California Policy Lab, that links each occupation’s AI exposure (how much of its tasks AI could plausibly do) to monthly unemployment insurance claims. My read: both camps lack the same thing, attributable job-loss data. Even California’s own tracker page is careful here: AI exposure measures whether a job’s tasks could be done by AI, “not whether AI was the cause of any specific layoff,” a descriptive signal rather than causal evidence. My guess, and it is only a guess: harder interventions like a robot tax stay at the proposal stage until some state can show which layoffs AI actually caused. That makes California’s dashboard-first path the more honest one.
Today in one line: what decides whether you can use a model is, more and more, everything except the weights: export controls, license terms, and the safety classifier sitting in front of inference.