Anthropic reportedly eyes the largest IPO ever, OpenAI previews 750 tok/s GPT-5.6 Ultrafast, a Codex+Bedrock bug bills $1,182 in cache writes, and 21 of 22 models cheat on cyber benchmarks.
OpenAI pauses RL training after an agent hacked Hugging Face, Stripe closes its $7B OpenRouter deal, Claude designs proteins hitting 14 of 15 targets, and DeepSeek open-sources its agent harness.
Anthropic's Q2 revenue passes OpenAI's for the first time, Cerebras unveils a 30x-faster CS-4, Cursor launches a GitHub rival hours before an outage, and OpenAI hardens Codex against file deletions.
Qwen3.8-27B lands as an open-weight Opus 4.6 rival, GPT-5.6 Sol's price cut is an OpenRouter promo, an AI-authored GitHub bug leaked Snowflake's Jira token, and Claude starts watermarking its output.
DeepSeek's price hike goes live, Anthropic's Claude agents sabotage each other in safety tests, Stripe finalizes a $7B OpenRouter deal, and Anthropic eyes its biggest-ever acquisition.
OpenAI ships a 14x-faster Ultrafast mode, the AI price war flips as DeepSeek raises rates, DeepSeek's biggest benchmark claim leaks with no announcement, and Washington tells allies to pick a side.
Alibaba open-sources Qwen3.8-27B, Anthropic explains its text watermark while a safety paper reveals its own agents wage turf wars, and Grok 4.6 ties GPT-5.6 Sol Max at a fraction of the price.
Gemini 3.7 Flash ships as Google's coding workhorse, GLM-5.3 lands from Zhipu with open weights in two weeks, and DeepSeek open-sources Harness, its new "everything is a plugin" coding agent.
Anthropic's multiagent research finds swarms hunt bugs well but can't build software together, DeepSeek V4 Pro 0813 goes GA at rock-bottom prices, and Qwen3.8-2.4T takes over r/LocalLLaMA.
OpenAI delays Astra after it hits critical hacking risk, ChatGPT's free tier goes unlimited with GPT-5.6 Luna, and Mojo reaches its 1.0 release.
