Catch up on a year of AI models
TinySet builds a dataset of this year's frontier model releases. TinyResearch explains which ones mattered.
TinySet
Dataset · 90 rowsFrontier AI models released this year, with lab, release date and what's new about it
| Model name | Lab | Release date | What's new |
|---|---|---|---|
| Gemini 4 Carbon | Google DeepMind | 2026-10-28 | Google DeepMind's strongest model, with state-of-the-art cybersecurity capabilities and long-horizon agentic/coding performance. |
| GPT-6.1 Terra | OpenAI | 2026-10-26 | Fast, efficient model combining detailed visual reasoning and precise computer use with clearer writing; 50% cheaper API prices than GPT-5.6 Terra, over 3x faster than GPT-6 Astra |
| Kimi K3.1 Pro | Moonshot AI | 2026-10-23 | First model in Moonshot's K3.1 series and its most capable 3T-class open-weight model, with a 1M-token context window and substantial gains in coding, 3D creation, financial research, and agentic capabilities. |
| MAI-Base-2 | Microsoft AI | 2026-10-15 | An MoE foundation model with 80 billion active and 2.8 trillion total parameters that outperforms every open-weight model tested on held-out evaluations. |
| Nemotron 3.5 Super VL | NVIDIA | 2026-10-15 | An open 120-billion-parameter multimodal mixture-of-experts model delivering the fastest task completion and highest accuracy for specialized always-on agents, combining visual reasoning, tool use and code generation for agentic systems. |
| Muse Spark Open | Meta | 2026-10-15 | Meta's most capable US open-weight model to date: a natively multimodal 816B-parameter (42B active) mixture-of-experts with a 1M-token context window, built for coding, agentic work and long-horizon reasoning. |
| Gemini 4 Argon | Google DeepMind | 2026-10-13 | Google's latest frontier model, it excels at deep work in complex, real-world tasks across coding, finance, legal, cybersecurity, and science, ranking #1 on the Vals Index. |
| Muse Wonder | Meta | 2026-10-12 | Meta's most capable model to date, delivering state-of-the-art performance on personal and professional agentic tasks with Active Memory (carrying context across conversations) and Active Focus (persisting through hard tasks), available via the Meta Model API and Muse Code. |
TinyResearch
ReportWhat actually changed in AI this year? Focus on what mattered and skip the hype.
2026 was the year AI moved from impressive demos to genuine capability jumps — but also the year the gap between benchmark headlines and real-world deployment became impossible to ignore. The most consequential developments: frontier models crossed into research-grade scientific work for the first time; AI coding agents saturated their original benchmarks and required harder ones; Chinese open-weight models narrowed the gap with U.S. frontier labs to a few months while dramatically undercutting on price; global AI infrastructure investment approached $1 trillion; and regulatory frameworks either slipped deadlines (EU) or remained largely voluntary (U.S.). Enterprise adoption spread widely but scaled poorly, with most pilot programs failing to reach production. Safety practices improved at the margins while companies simultaneously weakened earlier self-imposed limits. The net judgment: AI's capability frontier moved decisively forward; the institutional infrastructure to govern and deploy it did not keep pace.
