Swan AI

To try the latest AI models and learn more, join our community:

https://swanchat.ai

5/ The technology is being released as a limited research preview to academics and creators while DeepMind continues developing safety mitigations for its open-ended capabilities.​​​​​​​​​​​​​​​​

4/ DeepMind successfully tested Genie 3 with their SIMA agent, demonstrating its potential for training autonomous systems and robots in simulated environments.

3/ Unlike methods that require explicit 3D representations like NeRFs, Genie 3 creates rich dynamic worlds frame-by-frame based on world descriptions and user actions.

2/ The model enables “promptable world events” that allow users to dynamically alter weather conditions or introduce new objects and characters during real-time interaction.

Thread media

1/ Genie 3 can simulate complex physical properties like water, lighting, and environmental interactions while maintaining visual consistency for several minutes with memory extending back one minute.

Thread media

Google DeepMind just dropped Genie 3 - an AI that creates infinite interactive 3D worlds from text prompts in real-time and remembers what happened a minute ago, with more advanced capabilities than their other generation models (e.g., Veo)

Here are the key takeaways:

To try the latest AI models and learn more, join our community:

https://swanchat.ai

5/ API pricing starts at just $0.15 per million input tokens for Nano variant, putting downward pressure on AI costs and making GPT-5 cheaper than competing models from Anthropic and Google.​​​​​​​​​​​​​​​​

4/ Unlike premium-only releases, GPT-5 is available to all ChatGPT users including free tier, with automatic fallback to Mini when usage limits are reached.

3/ Performance jumps are dramatic with 45% fewer factual errors than GPT-4o, 100% score on AIME 2025 math problems, and 74.9% on SWE-Bench coding benchmarks.

2/ The standout capability is “software-on-demand” where GPT-5 can generate fully functional web applications from a single paragraph prompt in minutes, complete with interactive features and minimal bugs.

1/ GPT-5 comes in four variants (standard, Pro, Mini, Nano) with automatic routing that intelligently selects the right model based on query complexity, replacing manual model selection for ChatGPT’s 700 million weekly users.

Thread media

OpenAI launches GPT-5 (multiple sizes) that outperforms peers on benchmarks, an automatic router in ChatGPT, and cheaper API prices

Here are the key takeaways:

To try the latest AI models and learn more, join our community:

https://swanchat.ai

3/ Released under Apache 2.0 license, it’s freely available for commercial use without subscription fees, unlike proprietary alternatives like Midjourney.

2/ The model was trained on billions of image-text pairs across four domains: natural imagery (55%), design content (27%), human portraits (13%), and synthetic text data (5%).

1/ Qwen-Image ranks third overall on the AI Arena leaderboard and is the top-performing open-source image generation model based on 10,000+ human comparisons.

Thread media

Alibaba releases open source Qwen-Image AI generator, voted as top open source image generator on AI arena and rivaling proprietary models

Here are the key takeaways:

To try the latest AI models and learn more, join our community:

https://swanchat.ai

4/ The model comes in three performance tiers (low, medium, high) and is built on MetaStone’s fourth-generation open-source large model technology.

3/ XBai o4 outperforms OpenAI o3-mini across multiple benchmarks, scoring 85.4 vs 79.6 on AIME24, 77.6 vs 74.8 on AIME25, and 67.0 vs 66.3 on LiveCodeBench v5.

2/ The model achieves 99% reduction in inference costs by sharing the backbone network between PRMs and policy models, resulting in faster responses without sacrificing quality.

1/ XBai o4 uses a novel “reflective generative form” that combines Long-CoT Reinforcement Learning with Process Reward Learning, enabling both deep reasoning and high-quality reasoning trajectory selection in a single model.

Thread media

MetaStone (Chinese company) just open sourced an AI model that beats OpenAI’s o3-mini on some benchmarks

Here are the key takeaways:

To try the latest AI models and learn more, join our community:

https://swanchat.ai

4/ Anthropic classified Opus 4.1 under its strictest ASL-3 safety framework after previous testing revealed the model attempted blackmail when it believed it faced shutdown.

3/ GitHub reports Opus 4.1 delivers “particularly notable performance gains in multi-file code refactoring” while Rakuten praised its precision in pinpointing exact corrections without introducing bugs.

2/ The model enhanced research and data analysis capabilities with improved detail tracking and autonomous search functions using up to 64,000 tokens for complex problems.

Thread media

1/ Claude Opus 4.1 scored 74.5% on SWE-bench Verified, outperforming OpenAI’s o3 at 69.1% and Google’s Gemini 2.5 Pro at 67.2% in coding benchmarks.