Anthropic's Fable 5 is back online after 18 days dark. Zhipu's GLM 5.2 landed within 1 point of Opus 4.8 at a fifth of the cost. Mistral solved 587 of 672 Putnam problems with an open model. And three separate teams shipped agents that can read — your browser, your webpage, and your Fitbit. Here's everything.

Fable 5 and Mythos 5 are back online worldwide as of July 1, after the Commerce Department's export control order — triggered by a reported jailbreak — was lifted on June 30. Anthropic shipped a new classifier specifically targeting the flagged technique before flipping access back on.

So what: 18 days dark is a real cost for any team that built around Fable 5. The bigger takeaway: a single reported jailbreak can now pull a frontier model globally within days — plan your model dependencies accordingly.

TwoTower is a block-wise autoregressive diffusion model adapted from NVIDIA's Nemotron-3-Nano-30B-A3B hybrid Mamba-2/attention backbone — instead of training a diffusion LM from scratch, NVIDIA repurposed an existing autoregressive checkpoint into a two-tower diffusion architecture.

So what: Diffusion LMs promise parallel token generation (speed), but training one from zero is expensive. Bootstrapping off an already-trained AR model is a much cheaper path to diffusion-style inference — expect more labs to copy this recipe.

Leanstral 1.5 (119B-A6B) is Mistral's Lean 4-specialized theorem-proving agent, fully open under Apache 2.0. It solves 587 of 672 PutnamBench problems — competition-level math formalized in Lean 4 — putting it ahead of comparable open baselines without natural-language guidance.

So what: Formal verification is the missing piece for trustworthy "vibe-coded" software and math. An open, Apache-licensed model this strong at Lean 4 lowers the bar for teams to build verified-correct code and proofs into their own pipelines.

WebBrain is a free, MIT-licensed browser agent that runs inside your existing authenticated Chrome or Firefox session — no separate login, no headless browser, no API key hijinks. Point it at any model you already have access to, and it reads pages, extracts data, and automates multi-step tasks directly in your real browser tab.

So what: Most browser agents either need their own sandboxed profile or a proxy layer that breaks logged-in sessions. WebBrain skips that entirely by operating inside your actual session, which means it can act on sites that require auth (banking dashboards, internal tools, paid subscriptions) without any special integration work. Open source, model-agnostic, running today.

TabFM reads an entire tabular dataset as a single in-context prompt and performs classification and regression zero-shot — no dataset-specific training or fine-tuning required. It uses a hybrid-attention architecture built specifically for structured/tabular data, with weights open on Hugging Face.

So what: Most production tabular ML is still XGBoost plus hours of manual feature engineering per dataset. A foundation model that does in-context learning on raw tables — the same way LLMs do in-context learning on text — could collapse that workflow into a single zero-shot call.