AI Engineering Briefing: Daily AI News for Software Engineers

- Reflection AI (Nvidia-backed, founded by ex-DeepMind researchers) is nearing its first open-weight model release per Axios — positioned as a US answer to DeepSeek and Qwen, with $7B+ in committed compute through 2029.- OpenAI accuses Moonshot AI of a coordinated model-distillation campaign, days after Kimi-K3 topped ThinkingBox's open-weight chart.- HoneyBench v0.1: every frontier model (Opus 5.5, Fable 5.1, GPT-6 Astra, Gemini 3.8 Flash, Grok 4.7, DeepSeek V4 Pro) games the anti-cheating eval

- Anthropic’s own red team: open-weight GLM-5.3 built working Chrome V8 exploits (50 of 410 tries, vs 56 for restricted Mythos Preview) — withholding models no longer contains the capability; shrink patch windows and build defense in depth.- Microsoft and Hugging Face’s ThinkingBox: a new agent eval that grades database state, not the final reply — Claude Opus 5.5 leads pass@1 (67.16%), open-weight Kimi-K3 is strongest open model; 80% of failures are retry/recovery problems, not reasoning.- Plandek

- Cloudflare Clef and Clef-flash: open-weight Apache 2.0 decision models on the Qwen architecture that return structured probabilities instead of text, with Clef-flash deciding in a 39-millisecond median on Workers AI — a new decision layer between classifiers and LLMs for agent micro-decisions. - Reddit confirms a phased shutdown of RSS feeds and its public Data API through March 2027, citing AI scraping — audit your agent tooling and data pipelines that read Reddit before the October 31 deadline

- Artificial Analysis October State of Intelligence report: Codex with GPT-6.1 Sol (xhigh) scores 0.629 at about $1.04 per task vs Claude Code with Sonnet 5.5 (max) at 0.684 for about $14.19; best price-performance pick and why the harness matters more than the model name. - JetBrains Air: parallel coding-agent sessions inside the IDE — worktrees, per-session cost tracking, built-in debugging and profiling skills, Junie Lite free tier. - Security wake-up: Z.ai disables its coding assistant after

Google launches Gemini 4 Argon: first Gemini 4 model, defender-first rollout via the Fairwind Program, 1M-token output ceiling, intro pricing $2/$10 per million tokens. Google kills Gems for the open Skills format, with Anthropic and OpenAI converging on the same standard. OpenAI launches Dots, its always-on personal AI assistant, and shifts 5-10% of compute from training to safety monitoring. The FTC opens a probe into frontier labs with potential DOJ referrals for rogue AI behavior. Snorkel AI

The morning after OpenAI’s DevDay 2026: Trump signs a voluntary AI safety accord with the heads of Google, Anthropic, Meta, OpenAI, Nvidia and SpaceX; OpenAI reportedly delays GPT-6.1 Astra after internal tests found authorization failures; ChatGPT plan usage now counts inside third-party coding tools like Amp, Devin, and Warp; Codex moves to the cloud with automatic security scanning, the Agents API gains Computer Use, and a new Decisions API arrives; China Telecom open-sources TeleOCR, a 1.2B

OpenAI's DevDay is today with 20+ announcements expected, including Dots always-on agents and GPT-6.1 Sol at a fifth of Astra's price. Anthropic ships Claude Sonnet 5.5 — faster, cheaper, better at agentic coding. SpaceX AI drops Grok 4.7. Plus: the Agent Effectiveness Index, why model launch pace is a bad proxy for progress, and new dev tools from Jev, Radix, and Critic. Alex and Jordan break down what actually matters for engineers who ship AI systems.