AI NEWS: GPT-4o Major Updates, Gemini 2.5 Pro, New DeepSeek, MCP Everywhere, New Image Models

Here's a rundown of the latest AI news, touching on significant updates to major models, new benchmarks, and interesting industry trends. We cover how older models are getting surprisingly powerful upgrades, new contenders are shaking up the coding and image generation scenes, and a crucial protocol for connecting AI agents to tools is becoming the standard everywhere. Plus, there's a peek into the booming financials of leading AI companies and some cool open-source developments from overseas.

• GPT-4o Got Some Serious Muscle: GPT-4o received major updates, making it surprisingly powerful again. It's now considered the best model for image generation and, according to independent benchmarks (like Artificial Analysis), the leading "non-reasoning" model for coding, surpassing strong competitors like Claude 3.7 Sonnet and Gemini 2.0 Flash.
• Why Update an Older Model? The reason OpenAI is pushing updates to 4o seems linked to a concept called Javon's paradox – as things get cheaper, demand increases. Even OpenAI, a giant with massive funding, is facing GPU shortages needed for training newer, larger models like 4.5. So, they're optimizing and enhancing the existing 4o, which is likely more cost-effective to run at scale.
• More 4o Improvements: Beyond coding and image generation, the updated GPT-4o is better at following complex instructions with multiple requests, handling technical coding problems, and shows improved intuition and creativity. On the lighter side, it uses fewer emojis now.
• Availability & Issues: The updated GPT-4o is currently available to all paid users, with free users getting access over the next few weeks. However, its image generation capability is already seeing rate limits due to unexpectedly high demand, and the overall speed for normal queries feels "unusably slow" right now, which is a problem because speed is a crucial factor for a preferred model.
• Gemini 2.5 Pro is a Coding Beast: This was another huge news item. Gemini 2.5 Pro is incredibly good at coding and, importantly, it's "super fast." It's called a "full thinking model" and is considered the "best coding model ever used" by the speaker, who finds speed especially vital for agentic and coding tasks.
• Massive Context Window: A key feature of Gemini 2.5 Pro is its massive one million token context window, about 10 times larger than Claude 3.7's. This is exciting because a larger context window should allow the model to understand entire codebases better, which is being actively tested.
• Gemini 2.5 Pro Availability: Good news for developers using specific tools: Gemini 2.5 Pro is now available in Windsurf and Cursor, making it easier to test and integrate into coding workflows.
• New DeepSeek V3 Checkpoint: It was a big week for models! A new version (checkpoint) of DeepSeek V3 was released quietly. While not a completely new model, this update significantly improves its performance, especially excelling at coding, math, and logic tasks.
• DeepSeek V3 is Fast & Open: DeepSeek V3 is also fast and, importantly, it's open-source and uses a very permissive MIT license. This is great because it allows anyone to download and potentially run it (though it's a large model, which might be a challenge locally) or use it via inference providers.
• DeepSeek V3 Benchmarks: Benchmarks show the new DeepSeek V3 performing extremely well against frontier models like the previous DeepSeek V3, QwenMax, GPT 4.5, and Claude Sonnet 3.7, especially dominating in math tests like AIME 2024, despite many of those competitors being closed-source.
• ARC-AGI2 Benchmark Released: The ARC Prize organization launched ARC-AGI 2, their new benchmark designed specifically to test models' "AGI-ness." These tasks require abstract reasoning and the ability to apply understanding from one context to another, something humans find relatively easy but AI models struggle with.
• AI vs. Human Performance on AGI Benchmarks: The current scores highlight the gap: The best AI models (like O3 Low) score very low on ARC-AGI 2 (only 4%), whereas humans achieve a perfect 100%. This significant difference is seen as proof that the benchmark is effectively testing true generalization capabilities beyond current AI strengths. There's still a million-dollar prize for solving the ARC AGI tasks.
• MCP is Becoming the Standard: The Model Context Protocol (MCP), a way for AI agents to easily connect and use various tools (like Zapier actions), is rapidly gaining traction and seems to be becoming the industry standard.
• Widespread MCP Adoption: Zapier (which offers connections to 10,000+ tools) announced its adoption of MCP, allowing agents to directly access its vast library of actions. OpenAI also adopted MCP as part of its agents API, enabling agents to use tools via the protocol. Microsoft is integrating MCP into Copilot Studio. This widespread adoption means wherever you run your agents, you'll likely be able to use MCP to give them tool access.
• Anthropic's Influence: The speaker notes that while MCP is an industry standard, Anthropic gets credit for setting it, giving them some influence in its development.
• Text-to-Image is Hot Right Now: This week saw significant advancements in text-to-image generation beyond GPT-4o's improvements.
• Reve Image 1.0: Reeve AI launched its text-to-image model, Reve Image 1.0, which looks really good based on examples shown, featuring accurate text rendering and diverse styles. It ranks highly in quality based on user votes in artificial analysis rankings.
• Ideogram 3.0: Ideogram also released 3.0, which looks phenomenal. While Ideogram claims the highest ELO rating for quality, the key takeaway is the high degree of control offered through features like remixing, upscaling, and style preferences, allowing users to create beautiful, hyper-realistic images with lots of customization.
• OpenAI is Making Bank: Financially, OpenAI is booming. Sources report they expect revenue to triple to $12.7 billion this year, although they are still currently losing money overall.
• AI is Not a Fad: This massive revenue growth is seen as strong evidence that AI is definitely not a fad and that the significant investment in the field is resulting in substantial value. The speaker feels this personally, using AI tools extensively, and believes the main barrier to wider adoption is people not knowing what's possible or how to use the tools effectively – a problem the speaker is personally trying to solve through education.
• OpenAI Leadership Changes: Some recent C-suite changes were noted: Sam Altman is shifting focus away from daily operations to concentrate more on research and product, while operating chief Brad Lightcap will take on a larger role overseeing business and day-to-day activities.
• Massive Valuation: SoftBank is reportedly set to invest $40 billion in OpenAI, valuing the company at an astounding $260 billion, which would make it one of the most valuable private companies ever.
• Quen Qvq Max - Visual Reasoning Powerhouse: Chinese company Quen released Qvq Max, an open-source visual reasoning model that can understand and reason with information from images and videos. This open-sourcing trend from Chinese companies is seen as a positive development, allowing users to customize and run the AI themselves.
• Quen Qvq Max Capabilities: This model can handle complex tasks like analyzing relationships between scenes in multiple images, solving math problems, and generating code or artistic creations based on visual input. It is a "thinking model with vision capability."
• Availability Challenge: The main drawback for US users is that Quen's services often require a Chinese phone number. However, there's hope that this powerful model will soon be available via third-party inference providers that US users can access, or potentially through quantized versions that are easier to run locally, as the full model is quite large.