01
Zhipu ships GLM-5.3 with 50% coding jump, open weights held back for cyber safety review
Z.ai released GLM-5.3 built on the same base as GLM-5.2, with all gains from extended post-training. Terminal-Bench 3.0 climbed from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and CyberGym reached 84.5%, slightly surpassing Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol. Open weights are delayed roughly two weeks pending a safety evaluation, a decision Z.ai attributes to the model's strong cybersecurity capabilities.
02
Alibaba's Qwen surpasses 3 billion downloads, claims over 50% of open-source model share
Alibaba's Qwen family has been downloaded more than 3 billion times over six months, with 460+ open-sourced models and 300,000+ derivative versions. According to a Hugging Face report cited by Alibaba, Google recorded 418 million downloads and Meta 227 million over the same period, putting Qwen well ahead of both.
03
DeepSeek launches V4 Pro and open-source Harness, a rival to Claude Code
DeepSeek released V4 Pro 0813 with stronger agentic capabilities and coding upgrades, alongside Harness, an open-source coding agent positioned as a competitor to Anthropic's Claude Code. V4 Pro is available on API with higher pricing than its preview predecessor, and DeepSeek published its own benchmark results showing gains in multi-step tool-use workflows.
04
Anthropic watermarks Claude-generated text to comply with EU law
Anthropic began watermarking text generated by Claude to comply with EU regulations. The company will release an API with "keys" to decode the watermark and verify whether Claude generated a given block of text. The watermark applies even when Claude is used only for proofreading, a scope that tech analyst Ben Thompson called "clearly absurd."
05
Vals AI raises $40M from a16z after finding frontier models fail 52% of real finance analyst tasks
San Francisco startup Vals AI closed a $40 million Series A at a $400 million valuation, led by Andreessen Horowitz. The company's testing found that frontier AI models fail 52% of real-world finance analyst tasks, highlighting what Vals AI calls an "Akerlof problem" where labs know far more about their models' weaknesses than the enterprises buying them.
06
Nvidia partners with Wall Street for $500 billion AI infrastructure financing
Nvidia announced partnerships with Apollo, BlackRock, Goldman Sachs, KKR, Brookfield, and Blackstone to finance up to $500 billion in AI infrastructure, making AI factory compute an "investable asset class." The deal aims to fund data centers, power capacity, and chip deployments for Nvidia customers, though analysts warn it could increase enterprise AI costs and worsen chip shortages.
07
Grok 4.6 launches with 500K context and $2/$6 pricing, matches GPT-5.6 on benchmarks
SpaceXAI released Grok 4.6 as a post-training upgrade with 500K context window, priced at $2 per million input tokens and $6 per million output tokens. The model matches GPT-5.6 on benchmark scores and is tuned for long-running agents, coding, and knowledge work, reportedly needing half the steps of Claude Opus 5 to complete equivalent tasks.