AI News Roundup, Oct 14-20 2026: Opus 5.5 Tops the Index, Open Weights Close the Gap
The models shipped this month are now measured, not just announced. Here's what the scores say.
The headline this week is that the new models got graded. Artificial Analysis's October snapshot puts Anthropic's Claude Opus 5.5 at the top of its Intelligence Index at 57.6, with Claude Sonnet 5.5 at 56.0 and Claude Fable 5.1 at 53.4 close behind. OpenAI's GPT-6 Astra and Google's Gemini 4 Argon both land around 53.
On the open side, DeepSeek V4.1 Flash narrowed the gap between China's best model and the US leaders to about 3 percent, per a Bloomberg report on October 5. That's down from a 15 percent gap at the start of the year, and it changes the 'only closed models are good' assumption a lot of buyers still hold.
Microsoft pushed agent safety forward on October 7 by taking Execution Containers for AI agents (MXC) to general availability. It enforces a declared policy of allowed files and network destinations at runtime, with GitHub Copilot and OpenAI Codex already supporting it and Claude Code on the way. Arena, formerly LMArena, launched an Alignment Index built from real agent sessions that rates reliability on unauthorized actions and false attribution.
Meta released Muse Wonder and Muse Wonder Full on October 13, frontier-level models with what it calls active memory. But the separate Muse agent had already drawn fire for security flaws serious enough that Apple changed macOS settings to block it, a reminder that capability and trust are not the same project.
The throughline: the models are no longer just launching, they're being measured, governed, and in some cases opened. That's the stage where real adoption separates from demo hype.
Updated October 28, 2026: scores and prices below were verified against the original announcements and remain accurate as of publication. For this week's releases, see our newest roundup.
Frequently asked questions:
Q: Which model topped the Artificial Analysis Intelligence Index this week?
A: Claude Opus 5.5 at 57.6, with Claude Sonnet 5.5 at 56.0 and Fable 5.1 at 53.4 close behind. OpenAI's GPT-6 Astra and Google's Gemini 4 Argon both land around 53.
Q: How close are open models to the US frontier now?
A: DeepSeek V4.1 Flash narrowed the China-US gap to about 3 percent, down from 15 percent at the start of the year, per Bloomberg reporting - which changes the 'only closed models are good' assumption many buyers still hold.
Q: What did Microsoft ship for agent safety?
A: Execution Containers for AI agents (MXC) reached general availability on October 7. It enforces a declared policy of allowed files and network destinations at runtime, with GitHub Copilot and OpenAI Codex already supporting it.