Easy AI Daily Digest | December 5, 2025: Gemini 3 Deep Think, GPT-5.1-Codex Max, Anthropic Acquires Bun

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily digest for December 5, 2025 covering major AI industry news. Google released Gemini 3 Deep Think mode for AI Ultra subscribers, scoring 45.1% on ARC-AGI-2 versus GPT-5.1's 17.6%. OpenAI launched GPT-5.1-Codex Max in the Responses API with VS Code and Cursor integration. Microsoft open-sourced VibeVoice-Realtime-0.5B, a lightweight real-time text-to-speech model supporting English and Chinese. Nous Research released Hermes 4.3 based on ByteDance Seed 36B, while Mistral Large 3 topped lmarena's open-source coding leaderboard. In business news, Anthropic acquired Bun as Claude Code revenue crossed $1 billion, Harvey raised $160 million at an $8 billion valuation, and Antithesis raised $105 million led by Jane Street. Research highlights include Google's Titans long-context memory architecture (2M+ tokens), TorchAI's MoE quantization support, and a fast ODE solver generating 4K images in 8 steps. OpenRouter's State of AI report analyzed 100 trillion tokens, finding Claude handles 60% of coding workloads.

📅 AI Industry Digest — December 5, 2025

Model Releases and Updates

#### Google Releases Gemini 3 Deep Think Mode Available for Google AI Ultra subscribers, Deep Think enhances complex reasoning via parallel thinking, scoring 45.1% on ARC-AGI-2 (surpassing GPT-5.1's 17.6%). Supports math, science, and other tasks.

Links: Google AI announcement | Google DeepMind details

#### OpenAI Launches GPT-5.1-Codex Max Available in the Responses API and integrated into the Codex agent harness. Supports IDEs including VS Code and Cursor, with improved code generation.

Links: OpenAI Devs announcement | Cursor integration

#### Microsoft Releases VibeVoice-Realtime-0.5B A lightweight real-time text-to-speech model supporting English and Chinese, open-sourced on Hugging Face.

Links: Hugging Face model page | Twitter announcement

#### Nous Research Releases Hermes 4.3 Based on ByteDance Seed 36B with near Hermes 4 70B performance, trained via the Psyche network with MoE support.

Link: Nous Research blog

#### Mistral Large 3 Tops Open-Source Coding Leaderboard Ranked #1 on lmarena, available via Ollama cloud, with community-confirmed coding strength.

Links: Mistral AI announcement | Ollama support

---

Research and Technical Advances

Google Titans long-context memory architecture: Combines RNN efficiency with Transformer performance, supporting 2M+ tokens; early results shown at NeurIPS. Google Research
TorchAO adds MoE quantization: New MoEQuantConfig enables quantization of mixture-of-experts models for faster inference. PyTorch PR
VATTENTION paper: First sparse attention mechanism with (ϵ, δ) guarantees, improving long-text processing. arXiv
STRAW: Sample-tuned rank-augmented weights mimic neuromodulation, dynamically adjusting model weights for better task adaptability. Substack
Fast ODE solver for diffusion models: Generates 4K images in 8 steps with quality comparable to 30-step DPM++2M SDE; open-sourced on Hugging Face. Space | arXiv

---

Industry News and Funding

Anthropic acquires Bun; Claude Code hits $1B: Anthropic acquired Bun as Claude's code-generation business surpassed $1 billion in annual revenue. Anthropic news
Perplexity receives investment from Cristiano Ronaldo: The football star invested in Perplexity, positioning it as "sparking global curiosity." Announcement
Harvey raises $160M Series F: $8 billion valuation, serving 700+ law firms, focused on legal AI. Source
Antithesis raises $105M led by Jane Street: Focused on deterministic simulation testing for AI-generated code. Source

---

Community Discussions

GPT-5.1 beats Gemini 3 at bug-finding: OpenAI Discord users report GPT-5.1 finds code bugs Gemini 3 misses. Discussion
Z-Image still filters sensitive content: Reddit users report filtering of gore/nudity with "maybe not safe" warnings despite claims of no censorship. Reddit
AI's impact on tech jobs debated on Reddit: Users argue AI will transform software roles rather than eliminate them. Reddit
LocalLlama tests VibeVoice-Realtime: Feedback on English/Chinese support, with some concerns about Mandarin accents. Reddit
Gemini 3 Deep Think benchmarks discussed: Reddit users debate the 45.1% ARC-AGI-2 score vs GPT-5.1's 17.6%. Reddit

---

Tools and Platform Updates

OpenRouter releases "State of AI" report: Analyzed 100 trillion tokens — 50% of open-model usage is roleplay, 50% of paid-model usage is coding, and Claude handles 60% of coding workloads. Report
Windsurf integrates GPT-5.1-Codex Max: Free trial for paid users with Low/Medium/High reasoning levels. Announcement
mcp-apps-sdk open-sourced: From General Intelligence Labs, enables embedding ChatGPT apps in other platforms. GitHub
tinygrad fixes train_step: PR fixes train_step not utilizing input tensors, improving training efficiency. PR
DSPy integration with Claude Code proposed: A user suggests native Claude Code support leveraging its Read/Write/Terminal tools. Discussion

---

Performance and Benchmarks

Gemini 3 Deep Think: 45.1% ARC-AGI-2 — 2.5x over GPT-5.1's 17.6%. Reddit
Mistral Large 3: #1 on lmarena coding among open-source models. Source
DeepSeek V3.2 serving metrics from Baseten: TTFT 0.22s, 191 tps, strong lmarena rankings in math/law/science. Source
GPT-5.1-Codex Max code generation: Users report improved code quality and efficiency in Cursor and other IDEs. Source

---

*Source: Easy AI educational project*