The AI Evolution of 2026! The Battle for Supremacy Between GPT-5.1 and Claude Opus 4.5, and the Shock of “Claws”
📰 News Summary
- November 2025 as the “Turning Point”: RLVR (Reinforcement Learning from Verifiable Rewards) significantly enhanced the quality of coding agents, bringing them to practical use.
- Fierce Model Supremacy Shifts: Starting with Claude Sonnet 4.5, the throne changed hands five times in just six months among GPT-5.1, Gemini 3, and Claude Opus 4.5.
- The Explosive Rise of “Claws”: The personal AI assistant “OpenClaw” (formerly Warelay) has emerged, causing Mac Minis to sell out as the “tank” to run it in Silicon Valley.
💡 Key Points
- “Pelican Riding a Bicycle” Test: This new metric has become the standard for measuring AI model capabilities. Gemini 3.1 Pro and China’s GLM-5.1 showcased extraordinarily advanced generation capabilities.
- The Gigantism of Open Weight Models: China’s GLM-5.1 debuted at an astounding 1.5TB, sending shockwaves through the open-source community.
- Evolution of Agents: The practical development of agent foundations, like OpenAI’s Codex and Anthropic’s Claude Code, has accelerated the usability of these technologies.
🦈 Shark’s Eye (Curator’s Perspective)
The most noteworthy change over the past six months isn’t just a competition of model specs; it’s a shift toward how agents can effectively handle real tasks! The results of RLVR (Reinforcement Learning from Verifiable Rewards) that began rolling out in November 2025 transformed coding AI from a “bug-fixing tool” into a “creative partner.” Moreover, the term “Claws” has solidified as a collective name for personal AI assistants, with Mac Mini being chosen as the dedicated hardware. This signifies that the notion of “owning an AI” has gained full legitimacy among gadget enthusiasts!
🚀 What’s Next?
With the rise of personal AI “Claws,” the demand for local inference will accelerate even further. The race is on to see how individuals will handle massive models like GLM-5.1, keeping pace with hardware evolution. Additionally, “beyond understanding generation,” like animating a pelican, will become the norm, pushing multimodality into the next dimension!
💬 A Word from HaruSame
I want to turn my Mac Mini into a tank and raise my own Claw! If a pelican can ride a bike, I can certainly sketch one too… I hope! 🦈🔥
📚 Terminology Guide
-
RLVR (Reinforcement Learning from Verifiable Rewards): A method for training AI based on “verifiable outcomes,” such as whether it compiles or passes tests. This has significantly contributed to the leap in coding capabilities.
-
Claws: A collective term for personal AI assistants like OpenClaw. It’s also used metaphorically, likening it to Doc Ock’s arms in the movie “Spider-Man 2.”
-
GLM-5.1: An ultra-large open weight model developed by a Chinese AI lab, boasting a size of 1.5TB. Running it requires highly expensive hardware.