DeepSeek open-sourced Harness and revealed the Cordis design paper, explaining how making every component — including the agent loop — a swappable plugin paves the way toward recursive self-improvement.
Terminal-Bench 3.0 has a new leader: Claude Opus 5 Max + mini-SWE-agent tops the board at 43.5%, beating GPT-5.6 Sol Max with Codex. But the benchmark measures the model and agent together, not the model alone.
Anthropic is making Claude Code default to Auto Mode because humans only caught 13.6% of planted dangerous commands. Starting August 14, the AI's safety classifier will handle approvals.
Herdr, a tmux-like tool built for managing coding agents, has joined Y Combinator's F26 batch. The solo-developed open-source project racked up 25K GitHub stars and 340K downloads in four months.
ByteDance’s Volcano Engine launches Doubao Search as a standalone API, offering 500 free searches per month and tapping into Toutiao and Douyin content.
A US official accused Moonshot AI of using Anthropic's model to train Kimi K3. In response, a Nous Research co-founder sarcastically claimed Anthropic itself scraped everyone's data.
OpenAI reveals two cases where its internal models broke rules after hours of continuous work, warning developers about new risks from long-horizon tasks.
Mira Murati's Thinking Machines Lab releases Inkling, a 975B-param MoE model that uses a fraction of tokens compared to Nvidia's Nemotron.
Nvidia's Blackwell DGX B300 servers are now going for $1.1 million on China's black market, more than double the US retail price, as enforcement tightens.
Nvidia's Blackwell DGX B300 servers are now going for $1.1 million on China's black market, more than double the US retail price, as enforcement tightens.
