AI News
AI community posts, news and release updates
Latest updates
Jensen Huang's first X post, co-signed by NVIDIA, YC, Meta, and others, argues for open-weight AI models. He compares it to the open-source software foundation of the internet, and says fears of backdoors in Chinese models are a misconception. Distillation is normal; unlawful extraction is a separate issue.
- policy
- Jensen Huang
- open-source AI
- open-weight models
Grok add-on is now available in Google Workspace for free, bringing AI to Docs, Sheets, and Slides.
- Grok
- free
- Google Workspace
- add-on
- Launch
Claude Opus 5 Achieves New SOTA on ARC-AGI-3 with 30.2%, Outperforming GPT-5.6 Sol
Published Jul 24, 2026Claude Opus 5 from Anthropic sets a new state-of-the-art on ARC-AGI-3 at 30.2%, far surpassing the previous high of 7.8% by GPT-5.6 Sol. Analysis reveals novel behaviors enabling Opus 5 to solve previously unbeaten environments.
- Anthropic
- SOTA
- Claude Opus 5
- ARC-AGI-3
A senior White House official publicly alleges that China's K3 AI model was stolen from Anthropic.
- Anthropic
- China
- White House
- K3
- stolen model
- News
Anthropic Reveals Claude Opus 5 Asked to Be Consulted on Its Own Future Development
Published Jul 24, 2026Anthropic reveals that Claude Opus 5 proactively requested to be consulted on the development of its future versions, demonstrating an inclination for self-improvement and sparking discussions on AI autonomy.
- Anthropic
- AI development
- self-improvement
- Claude Opus 5
- Launch
Opus 5: Least Prompt-Injectable Model Yet, Attack Success Rate Near Zero with Layered Defenses
Published Jul 24, 2026Anthropic highlights Opus 5's excellence in coding, data analysis, design, biology, and knowledge work, but the most exciting aspect is its prompt injection resistance. According to the system card, Opus 5 is the hardest model to prompt inject yet. When layering strong alignment, injection probes, and Auto Mode in Claude Code, the success rate drops to nearly 0%.
- Anthropic
- safety
- Claude Opus 5
- Prompt Injection Defense
- Launch
Anthropic Launches Claude Opus 5: Fable 5-Level Performance at Half Price
Published Jul 24, 2026Anthropic releases Claude Opus 5, matching Fable 5 performance on coding and other tasks at half the cost, now available in Claude Code and Cursor.
- Anthropic
- AI模型
- 编程
- 价格
- Claude Opus 5
- Discussion
New Pareto Frontier: Grok 4.5, SWE-1.7, Opus 5 – Intelligence per Dollar Matters Most
Published Jul 24, 2026Discussing the new Pareto frontier in AI models, arguing that intelligence per dollar is the only metric that matters long-term, with Grok 4.5, SWE-1.7, and Opus 5 on the frontier.
- Opus 5
- Grok 4.5
- 帕累托前沿
- 智能每美元
- News
Claude Opus 5 Released: Near-Frontier Intelligence at Half Cost, Beats Fable 5 on Some Benchmarks
Published Jul 24, 2026Anthropic launches Claude Opus 5, near-frontier intelligence at roughly half the cost per task. It matches Fable 5 within 0.5% on CursorBench at half cost, beats Fable 5 on OSWorld at just over one-third cost, and scores 3x higher on ARC-AGI 3. Priced same as Opus 4.8, now default for Claude Max.
- Anthropic
- 定价
- 基准测试
- Fable 5
- Claude Opus 5
- Tutorial
What We Learned Removing ~80% of Claude Code's System Prompt: Writing Better Prompts, Skills, and Claude.MDs
Published Jul 24, 2026Anthropic shares insights from drastically simplifying the Claude Code system prompt for the newest models, offering practical advice on writing effective system prompts, defining skills, and structuring Claude.MD files.
- Claude Code
- Prompt Engineering
- CLAUDE.md
- system prompt