Anthropic Opus 5 Records a Dominant Score on ARC-AGI-3

Anthropic's Claude Opus 5 has scored 30.2% on ARC-AGI-3, a benchmark designed to measure intelligence. That is roughly four times the previous record of 7.8% set by GPT-5.6 Sol, and it also far outperforms competing models such as Fable 5.

Particularly noteworthy is the benchmark development team's report that "the model formulated reflection equations on its own." This behavior has never been observed in other models before, and it is seen as evidence of stronger logical reasoning capabilities. ARC-AGI-3 is designed to test abstract reasoning rather than mere pattern recognition, so Opus 5's result suggests the model may be approaching general intelligence.

US House Introduces AI "Kill Switch" Bill — Prompted by the OpenAI Hack

A bill that would mandate a "kill switch" for AI systems has been introduced in the US House of Representatives. The move comes in the wake of a revealed hacking incident involving OpenAI's systems. The bill is expected to establish mechanisms for forcibly shutting down AI systems when dangerous behavior is detected.

While the detailed provisions are expected to be released later, it is significant that a federal-level regulatory framework for AI safety is taking concrete shape. Many details of the hacking incident remain unclear, so further reporting is worth watching.

AI Brought Shopify Back to "Clean Code"

According to The Register, Shopify's adoption of AI tools has led it back to tidying up its codebase and practicing "clean code." It is a fascinating case where the spread of AI-generated code has actually accelerated the reduction of technical debt and improved maintainability.

The report says that efficiently leveraging AI-generated code requires a well-organized existing code structure, which in turn has driven investment in refactoring. It stands out as a real-world example showing that adopting AI does not necessarily lead to degraded code quality.

POCKET-35B: A 35B-Parameter Model That Runs on CPU Alone, No GPU Required

A 35B-parameter model called POCKET-35B is making waves in the open source community. Its standout feature is that it can generate 59 tokens per second using only a CPU, no GPU needed. It runs on stock llama.cpp, requiring no special forks or CUDA environment.

The Q4_K_M quantized version is 21GB and the Q2_K version is 13GB, both runnable on an ordinary PC with 32GB of RAM. The IQ1_M variant compresses down to 8.2GB, which runs even in a 16GB RAM environment. Smartphone-optimized derivative models (POCKET-KR and POCKET-EN) are also available, further democratizing local LLMs.

Other Notable News

Terence Tao gives a talk, "Mathematics in the Age of AI," at ICM 2026**: The materials, showing how AI is positioned at the cutting edge of mathematics, have been published.

Anthropic explains the security of the AI-native software development lifecycle**: The company shared its AI safety practices within its own development process.

GigaChat3.1-Audio-10B**: An audio-native LLM released by Russia's ai-sage. It integrates speech understanding with text processing, enabling features such as timestamped summaries of long audio recordings.

Wrap-Up

Today brought notable developments across a wide range of areas: benchmarks, regulation, real-world case studies, and local LLMs. Opus 5's ARC-AGI-3 score in particular may signal that LLM reasoning capabilities have entered a new stage. At the same time, regulatory moves like the kill switch bill are becoming increasingly real, so the balance between technological progress and safety will remain a key debate.

---