Anthropic opened a research preview of Claude for Chrome on Tuesday to 1,000 Max-plan subscribers, who pay $100-$200 a month, with a waitlist behind them. The extension lives in a sidecar window with context on the page in view and can act on it — clicking, filling forms, moving between tabs using the logins already in your session.

Unusually, the company shipped its own adversary numbers alongside the product: unmitigated browser use fell to prompt injection in 23.6% of tests, reduced to 11.2% with its defenses in place. Financial-services, adult and piracy sites are blocked by default, and publishing, purchasing or sharing personal data requires explicit user permission. It follows Brave's finding that Perplexity's Comet agent was vulnerable to indirect prompt injection, making this the second consecutive week that browser agents have been the story.

Nous Research published Hermes 4 on Monday — hybrid reasoners at 70B and 405B built on Llama-3.1 bases, with the technical report covering a 14B Qwen3-based variant. Reasoning toggles with a think tag, so one checkpoint serves both fast answers and extended deliberation, and post-training scaled from roughly 1M samples and 1.2B tokens in Hermes 3 to about 5M samples and 60B tokens. Licenses are inherited from the bases: the Llama 3 Community License for the 70B and 405B, which carry Llama-3.1's 128K context windows.

The numbers are competitive with closed systems in reasoning mode: 96.3% on MATH-500 and 81.9% on AIME 2024 for the 405B, plus 57.1% on RefusalBench — the benchmark Nous wrote to measure willingness to answer — against 17.7% for GPT-4o and 17.0% for Claude Sonnet 4. For a 405B checkpoint the license inheritance is the real story: the weights are free to run, but Meta's Llama 3 Community License, not Nous Research, sets the commercial terms.

On Monday xAI and X Corp. filed suit against Apple and OpenAI in the U.S. District Court for the Northern District of Texas, alleging violations of Sections 1 and 2 of the Sherman Act and the Texas Free Enterprise and Antitrust Act. The complaint claims Apple and OpenAI colluded to maintain monopolies in smartphones and generative AI, and that Apple deprioritizes super apps and chatbot rivals such as Grok in App Store rankings while favoring OpenAI — the company it integrated ChatGPT with across iPhone, iPad and Mac in 2024.

OpenAI's response was a single line: the filing is consistent with Mr. Musk's ongoing pattern of harassment. Apple has said its App Store is designed to be fair and free of bias. The suit lands three weeks after Musk publicly threatened Apple with immediate legal action on 12 August, and it arrives with the judge in Google's separate search case still weighing a Chrome divestiture — the other live lever on who controls assistant distribution.

At 15:16 PT on Saturday Musk posted that the xAI Grok 2.5 model, which was its best model last year, is now open source; the weights appeared on Hugging Face as xai-org/grok-2. The repository ships 42 files totalling approximately 500 GB in FP8, and its own card states the checkpoint is tensor-parallel 8 — eight GPUs with more than 40 GB of memory each to serve it. It is a 2024-era mixture-of-experts model, and Musk added that Grok 3 will follow in about six months.

The model is not the story; the license is. Grok 2.5 is released under xAI's custom Grok 2 Community License, which engineer Tim Kellogg described to TechCrunch as custom with some anti-competitive terms. What xAI gave away is a one-generation-old system, well behind Grok 4 and the current open frontier set by GLM-4.5 and OpenAI's gpt-oss — so the practical effect is a very large downloadable artefact under non-standard terms rather than a competitive release.