TL;DR: AI News Recap (Feb 6–9, 2026) — OpenAI’s Codex push, Claude Opus 4.6 momentum, RLMs, and MoE systems work

Major Highlights:

• OpenAI pivots user-facing focus to “builder tooling” with GPT‑5.3‑Codex
  • A Codex-centric Super Bowl ad framed “You can just build things” as the mainstream interface for frontier models—shifting mindshare from chat UIs to hands-on creation. GPT‑5.3‑Codex is rolling out across Cursor, VS Code, and GitHub with phased API access and is flagged as OpenAI’s first “high cybersecurity capability” model under its Preparedness Framework. Early adoption signals are strong (1M+ Codex App downloads in week one, 60%+ weekly user growth), with a stated intent to keep a free tier (possibly with tighter limits).
• Claude Opus 4.6 consolidates lead in interactive/agentic use; coding gap narrows
  • Opus 4.6 is repeatedly called the strongest “agentic generalist,” while Codex 5.3 gains ground in coding. Opus tops both Text and Code Arena leaderboards (with Anthropic models holding 4 of the Code Arena top 5 in one snapshot). In niche tests like WeirdML, it excels but is notably token-hungry (avg ~32k output tokens; sometimes hitting a 128k cap), highlighting serving-cost tradeoffs.
• Recursive Language Models (RLMs) emerge as a practical long-context strategy
  • RLMs introduce a second, programmatic “context pool” (files/variables/tools) alongside token context, letting models selectively materialize information into tokens. Authors released an open-weights RLM‑Qwen3‑8B‑v0.1 showing marked gains; engineers are already prototyping RLM-like recursion within agents (e.g., Claude Code using bash/files) to improve long-horizon tasks without massive context windows.
• MoE systems: new comms pattern vs. growing skepticism
  • Multi‑Head LatentMoE with Head Parallelism claims O(1) comms w.r.t. activated experts, deterministic traffic, better load balance, up to 1.61× speedups vs standard MoE expert parallelism and ~4× less inter‑GPU communication at k=4—pushing feasibility of very large expert counts. Yet practitioners voice doubts about top‑k routing pathologies and call for next‑gen conditional compute beyond classic MoE.

Key Technical Details:

• Models: GPT‑5.3‑Codex (OpenAI), Claude Opus 4.6 (Anthropic), RLM‑Qwen3‑8B‑v0.1 (open weights).
• Adoption: 1M+ Codex App downloads in week one; >60% weekly user growth.
• Performance: Opus 4.6 leads Text/Code Arena; WeirdML outputs average ~32k tokens (can hit 128k).
• Serving economics: Reports on Opus “fast mode” behavior; token-hungry outputs stress throughput/latency and cost.
• Systems: Multi‑Head LatentMoE + Head Parallelism—up to 1.61× faster, ~4× less inter‑GPU comms (k=4), deterministic comms; community sparsity tracking across GLM/Qwen/DeepSeek/etc.
• Availability: Codex rolling out across Cursor/VS Code/GitHub with phased API; Cursor notes 5.3 feels “noticeably faster than 5.2.” Some VS Code rollout pauses acknowledged.

Community Response/Impact:

• Strong “permissionless building” zeitgeist; developers porting apps (e.g., to iOS/Swift) and shipping agentic tooling fast.
• Friction: 5.3 can be overly literal in UI labeling; hiccups in rollouts; partner dynamics (e.g., Cursor/OpenAI) debated.
• Evals: Movement toward a “post-benchmark” mindset—workflow design, tool choice, and harness quality often trump headline scores.

First Principles Analysis:

• The interface shift from chat to builder tooling reflects a deeper product thesis: lower activation energy for creation compounds user value and model stickiness.
• RLMs reframe “long context” as a control problem: use structured external state plus recursion to turn sprawling tasks into tractable subproblems—improving effective context without scaling token windows linearly.
• MoE advances remain attractive for throughput but are constrained by routing stability and training dynamics; the next leap likely blends conditional compute with smoother, differentiable allocation to avoid MoE’s chronic collapse modes.
• Net: Models are “good enough” that systems design, eval realism, and serving economics now dominate differentiation.