AI News
AI community posts, news and release updates
Latest updates
- Discussion
Opus 5 Crushes Fable 5 in 3D Physics: All Scenes Pass at Half the Cost
Published Jul 24, 2026A user tested four models on creating HTML-based 3D physics scenes (tornado, wrecking ball, bridge collapse). Opus 5 nailed all three for $1.40, while Fable 5 cost $2.82 and failed on details. GPT-5.6 and Kimi K3 also had major physics errors.
- comparison
- Cost Efficiency
- Claude Opus 5
- 3D physics
Someone posted what appears to be the system prompt for Opus 5, sparking discussions about model instructions.
- Claude
- system prompt
- Opus 5
Anthropic released pricing for Claude Opus 5 and shared launch-day cost-per-task metrics for developers to evaluate value.
- Pricing
- Claude Opus 5
- cost-per-task
- News
DeepSeek founder: Nvidia digging its own grave in China, Huawei 950 can fully replace GB200/GB300
Published Jul 24, 2026According to a closed-door meeting transcript obtained by Tencent Tech, DeepSeek founder Liang Wenfeng says Nvidia is digging its own grave in China, and the Huawei 950 supernode can fully substitute for Nvidia's GB200/GB300 in performance, even at 100-200% higher prices. He reveals DeepSeek has about 16,000 Huawei cards.
- Nvidia
- DeepSeek
- Huawei
- 替代
- Liang Wenfeng
- News
Dead Internet Theory Was Right: AI Agents Are Eating the Web, Growing Nearly 8,000%
Published Jul 24, 2026A new report shows that traffic from AI agents has surged nearly 8,000% in the past year, validating the Dead Internet Theory.
- AI代理
- 增长
- 互联网
Running all 53 GAIA Level 1 tasks on the same Apple M4 Max with the same 4-bit Qwen-3.6-35B, Atomic Agent solved 37 tasks in 3h 12m, while Hermes solved 31 in 5h 10m — six more tasks solved, nearly two hours faster. Open-sourced under MIT license, a big win for open source.
- open-source
- benchmark
- Hermes
- Atomic Agent
- GAIA
A user highlights the rapid progress: Claude Opus 4.8 scored under 5% on ARC-AGI 3 two months ago, while Opus 5 now exceeds 30%. Additionally, Opus 5 offers comparable intelligence to Fable 5 at 26% lower cost per task. The pace of improvement is staggering.
- benchmark
- Claude Opus 5
- ARC-AGI-3
- performance jump
The article discusses how to avoid catastrophism amid rapid AI growth, advocating rational analysis and proactive action.
- AI发展
- 灾难论
- News
Atomic Agent crushes Hermes on GAIA: +11% accuracy, 1.6× faster, MIT open source
Published Jul 24, 2026In a side-by-side on 53 GAIA Level 1 tasks, Atomic Agent solved 37 in 3h12m vs Hermes's 31 in 5h10m. Examples: Audre Lorde poem indent in 7.6 min (Hermes blank), Vietnamese specimen city in 33s (Hermes 7.3 min no answer), dinosaur nominator in 57s (Hermes wrong after 11 min). Key tech: byte-stable prompt prefix for KV-cache reuse, compact tool-call JSON array, no-progress guard prevents spinning. All on same hardware/model.
- benchmark
- Hermes
- 效率
- KV cache
- Atomic Agent
- GAIA
On ChatGPT web, you can now create a shareable link for your custom pet and send it to a friend to adopt.
- ChatGPT
- custom pet
- shareable link