Ricursive's Apache-2.0 Ripple-65B tops US open coding models

A dense 65B from AlphaChip's creators beats 500B–975B open rivals for ~$32M. Plus: Codex's agent console, leaner Claude API loops, ARTEX in the wild.

Nowline NOV 9 7:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A self-hostable 65B that out-codes 975B open models

    Ricursive's Ripple-65B Preview is a dense 65B under Apache 2.0, base weights included, that posts 80.5 on Terminal-Bench 2.1 and 79.1 on SWE-bench Pro v2.0 Hard — ahead of Reflection Beam (501B) and Thinking Machines Inkling (975B) on the company's own chart. Built by the AlphaChip team for about $32M, roughly a tenth of the prior US open leader's spend. Pull BF16/NVFP4 weights from Hugging Face or call it on OpenRouter at ~$0.20/$0.60 per 1M tokens. Scores are vendor-reported and the margin over Beam is slim, so re-run your own evals before you switch.

  • Codex 0.167 turns your terminal into an agent console

    OpenAI's Codex CLI 0.167.0 adds a command center for agent sessions (`codex agents`), lets you switch connected backends straight from the session picker, previews images inline in the composer, and ships a Google Vertex AI setup path plus an opt-in system proxy. Worth the update if you run several coding agents in parallel from one shell.

  • Claude's Messages API gets server-side threads

    Anthropic's new message-threads beta (header `message-threads-2026-08-12`) holds conversation state server-side, so each turn sends only the new messages instead of the full history — reportedly shrinking request payloads from ~800KB to ~1.8KB and trimming agent latency by about 35%. Threads expire after five minutes idle. A concrete token-and-latency win for long-running Claude agents.

  • ARTEX, an open-source AI pentest tool, is loose in the wild

    Censys counted 1,465 publicly exposed consoles of ARTEX, an autonomous AI penetration-testing framework now tied to real intrusions — operators hitting around 200 orgs and South Korean banks, per Check Point and CrowdStrike. If you run offensive-AI tooling, put the console behind auth today; everyone else should plan for more agent-driven attack traffic.