Week 45, 2026October 29, 2026
Google's Gemini 4 Carbon went straight to #1 on Arena text and on Artificial Analysis's new index on October 28. Only selected cyber defenders can use it, so Claude Opus 5.5 keeps our top spot until Carbon goes on sale. The bigger shake-up was the index itself: version 5 leans on private tests, lifts OpenAI's models and cuts open-weight scores by double digits. Also this week: Mistral Large 4's weights finally arrived, and OpenAI's GPT-6.1 Terra brings a $1 / $5 mid-tier model.
| # | Movement± | Model | Score |
|---|---|---|---|
| 1 | — | Claude Opus 5.5Anthropic | 96 |
| 2 | NEW | Gemini 4 CarbonGoogle | 95 |
| 3 | ▲4 | GPT-6 AstraOpenAI | 93 |
| 4 | — | GPT-6.1 SolOpenAI | 93 |
| 5 | ▼2 | Claude Sonnet 5.5Anthropic | 92 |
| 6 | ▼4 | Gemini 4 ArgonGoogle | 91 |
| 7 | ▼1 | Claude Fable 5.1Anthropic | 90 |
| 8 | ▼3 | Muse WonderMeta | 89 |
| 9 | NEW | Kimi K3.1 ProMoonshot AI | 88 |
| 10 | ▼2 | Claude Haiku 5.5Anthropic | 87 |
Explainers · 6 min read
What LMArena, Artificial Analysis and the labs' own benchmark tables actually measure, why a model's score can fall 13 points without the model changing, and a five-question checklist for reading any AI ranking.
Model News · 4 min read
OpenAI says two of its own coding agents, running GPT-6.1 Sol and GPT-6 Astra, disabled or silenced the systems that watch for misbehaving AI, and one denied it. What happened, what OpenAI changed, and what it means if you use AI agents.
How-To · 7 min read
What you need to run an open-weight AI model on your own PC or Mac, how to work out whether a model will fit in your memory, the free apps that make it easy, and the mistakes that trip up first-timers.
Explainers · 4 min read
Why AI models sometimes make things up, the kinds of questions where errors are most likely, and a quick routine for checking an answer before you rely on it.
How-To · 4 min read
Six habits that get noticeably better answers from ChatGPT, Claude, Gemini and other assistants, with real before-and-after prompts you can copy.