On September 29, the AI field ushered in a wave of heavyweight updates. DeepSeek, Anthropic, OpenAI, and Zhipu AI simultaneously released their latest models, bringing stronger AI capabilities to developers and users.
DeepSeek-V3.2: Another Breakthrough in Open Source Large Models, Costs Cut by Over 50%
On September 29, DeepSeek officially released and open-sourced the DeepSeek-V3.2-Exp model on the Hugging Face platform.
Model Link: https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp
API Prices Slashed by Over 50%
The main feature of the new model is the significant reduction in service costs. DeepSeek immediately followed up by lowering its API invocation fees: the official API price has been reduced by over 50%, and the new pricing has taken effect immediately. This move will drastically reduce usage costs for developers.
Anthropic Claude Sonnet 4.5: Autonomous Programming Capability Leaps Forward
On September 29, Anthropic released the brand new Claude Sonnet 4.5 model.
Official Announcement: https://www.anthropic.com/news/claude-sonnet-4-5
Key Highlights
- Significantly Enhanced Instruction-Following: The model's understanding and execution of complex instructions are much more precise.
- Long-duration Autonomous Programming: It can continuously code autonomously for up to 30 hours, a more than fourfold increase compared to the 7 hours of Claude Opus 4.
- Available Immediately: It has already been launched in Claude Code, and developers can use it directly.
OpenAI Sora 2: The ChatGPT Moment for Video Generation
On September 29, OpenAI introduced Sora 2.0, its most advanced video generation model to date, and simultaneously launched a standalone iOS app (currently invitation-only).
Key Features
- Drastically Improved Physical Realism: The physical behavior of generated videos is much more realistic.
- Optimized Audio-Video Sync: The alignment between audio and video is significantly improved.
- Multi-camera Narrative Capability: Supports more complex storytelling methods.
- Social Interaction Features: Users can generate, browse, and share AI videos, and even "cameo" in others' creations as digital avatars.
OpenAI called this update the "ChatGPT moment for video generation," signifying a qualitative leap in video generation technology.
Official Introduction & Demo Preview: https://openai.com/index/sora-2/
You can get invitation codes shared by other users on the official Discord (connecting your OpenAI account is required to join the community): https://discord.gg/openai
Introduction to Sora by OpenAI (iOS APP)
Download Link: https://apps.apple.com/us/app/sora-by-openai/id6744034028
Currently, it supports downloading with Apple IDs from the United States, Canada, and Japan.
How to Access
OpenAI uses an invitation code referral mechanism to promote the Sora APP: each joined account can generate one invitation code, and each invitation code can be used 4 times.
Workflow: Download APP → Log in with OpenAI account → Enter invitation code to start using.
Product Positioning
Sora by OpenAI is a short-video social platform, positioned similarly to TikTok, with the core difference being that all content is AI-generated. Users can create, browse, and share AI videos, creating a brand new social experience.
Core Features
- Text-to-Video: Generate a 10-second video with a single sentence description.
- Draft Management: Generated videos are automatically saved as drafts, and users can choose to download them locally or publish them publicly to the platform.
- Cameo Feature: Scan a 3D face and clone the voice to create a personal digital avatar for the AI to use during video generation.
- Permission Control: Cameo supports flexible permission settings to protect user privacy.
- Social Interaction: When generating videos, you can reference other users' Cameos using
@. For example,@samacan make OpenAI's CEO appear directly in your video, enabling creative collaboration.
Zhipu GLM-4.6: Aligned with Claude Sonnet 4
Zhipu AI updated its flagship model GLM-4.6, achieving significant improvements across multiple dimensions.
Key Highlights
Significantly Upgraded Coding Capability
- A 27% increase compared to GLM-4.5.
- On public benchmarks and real-world coding tasks, its coding capability aligns with Claude Sonnet 4.
Expanded Context Window
- Increased from 128K to 200K.
- Able to better handle complex code codebases and agent tasks requiring long-term memory.
Enhanced Reasoning Capability
- Supports calling tools during the reasoning process.
- Multi-step reasoning is more stable and reliable.
Optimized Writing and Interaction
- Enhanced search and writing capabilities.
- Writing style and readability are closer to human preferences.
- Performs more naturally in scenarios like role-playing.
Benchmark Comparison
Official Documentation: https://docs.bigmodel.cn/cn/guide/models/text/glm-4.6
Online Experience: https://z.ai
The concentrated releases from these four companies demonstrate the intense competition and rapid progress in the AI LLM space. From price cuts to performance improvements, and from code generation to video creation, AI technology is accelerating its penetration into various application scenarios.
Comments