AI News Briefing — February 13, 2026

Google's Gemini 3 Deep Think update achieves 84.6% on ARC-AGI-2, a benchmark considered a frontier test of general reasoning. The model uses extended test-time compute ('thinking longer') and internal verification to solve problems previously requiring human expert intervention. This represents a major leap toward AGI-class reasoning capabilities.

Google DeepMind's official blog post announces the Gemini 3 Deep Think update, describing it as their most specialized reasoning mode for modern science, research, and engineering challenges. This is the primary source for the benchmark-shattering results.

Anthropic raised $30 billion at a $380 billion valuation, more than doubling its value from a $183B round just five months prior. The round was led by GIC and Coatue, with Anthropic reporting annualized revenue of $14B after tenfold yearly growth. This is among the largest private fundraising deals on record.

OpenAI released GPT-5.3-Codex-Spark, its first production model running on non-Nvidia hardware (Cerebras WSE-3), delivering over 1,000 tokens per second—roughly 15x faster than its predecessor. This marks a significant step in hardware diversification away from Nvidia's dominance. Available to ChatGPT Pro subscribers as a research preview.

MarkTechPost's detailed technical coverage of GPT-5.3-Codex-Spark explains the Cerebras Wafer-Scale Engine 3 architecture enabling the 1000+ tokens/sec performance. The model prioritizes extreme speed over deep reasoning, targeting near-instant developer workflows.

Building on yesterday's Reddit coverage, Z.ai launched GLM-5, a new state-of-the-art open-weights LLM with 744B parameters (40B active) trained on 28.5T tokens, described as Opus-class performance. The model integrates DeepSeek Sparse Attention and is part of a wave of Chinese open-model releases. This is a significant leap from GLM-4.5's 355B/32B active architecture.