The past week has been one of the most significant in AI development we’ll likely see all year. OpenAI dropped two powerful new models—O3 and O4 Mini—that have quickly become the talk of the tech community. But what’s truly fascinating isn’t just these releases, but how quickly competitors like Google and xAI are responding with their own innovations.
After spending time testing these new models and reviewing community benchmarks, I’ve noticed something remarkable: we’re witnessing an unprecedented acceleration in AI capabilities that’s transforming how we interact with technology.
OpenAI’s New Models: Tables, Math, and Creepy Capabilities
The most noticeable change with O3 is its preference for organizing information into tables rather than lists. While this might seem like a minor UI change, it represents a fundamental shift in how AI communicates complex information. These tables provide clear, scannable data that makes research and information gathering significantly more efficient.
What’s truly impressive about these models isn’t just their presentation but their raw intelligence. O4 Mini High solved a complex Project Euler math problem in under 3 minutes—a task that only 15 human experts could solve in under 30 minutes. That’s 10 times faster than the best human solvers.
The multimodal capabilities are equally striking. In benchmark tests like Enigma Eval, which tests the ability to solve complex visual puzzles, O3 achieved a 13% pass rate—far higher than Claude 3.7’s 2.26%. While 13% might sound low, it represents a massive leap forward in solving these particularly challenging problems.
Then there’s the slightly unsettling side of these advancements. One user uploaded an image of a restaurant menu with no identifying information, and O3 was able to search the web, match menu items, and correctly identify the exact restaurant. Others have reported the model pinpointing locations from selfies. The privacy implications here are significant and worth considering.
The Curious Limitations
Despite these impressive capabilities, these models still struggle with seemingly simple tasks. One user showed that O3 needed over 7 minutes to correctly identify the time on an analog clock. Even more telling, it completely failed at a simple visual tracing task that most children could solve easily—even after 13 minutes of “thinking.”
These limitations reveal something important about current AI: while these models can perform graduate-level analysis and solve complex mathematical problems, they lack the intuitive visual reasoning that humans develop naturally. This suggests we’re still far from true artificial general intelligence, despite the marketing hype.
Google and xAI Strike Back
What makes this moment particularly exciting is how quickly competitors are responding. The day after OpenAI’s announcement, Google released Gemini 2.5 Flash Preview—a lighter, cheaper alternative to their flagship model that still delivers impressive results, particularly in coding tasks.
Meanwhile, xAI appears to have quietly updated Grok 3 Mini, which is now matching or beating both OpenAI and Google models on several benchmarks while maintaining extremely competitive pricing (30¢ in/50¢ out per million tokens).
The competition extends beyond text models. The video generation space is seeing similar rapid advancement, with new models from LTX Studio, Cling, WAN, and V2 AI all pushing boundaries in different ways—from lightweight open-source options to high-fidelity 4K generation.
What This Means for Users
For those of us using these tools, this competitive landscape is delivering tremendous value. The rapid pace of improvement means we’re getting more capable AI at increasingly affordable prices.
However, there are important considerations about how these models are deployed. OpenAI’s decision to limit context windows in ChatGPT (8K tokens for free users, 32K for Plus) while offering much larger windows via API (200K+) creates an artificial perception of limitation for many users.
Similarly, the benchmarks reveal that different models excel at different tasks—O1 Pro still dominates in cipher decryption, while Gemini models often outperform in coding tasks. This suggests that the ideal approach might be using multiple specialized models rather than seeking a single “best” option.
The AI arms race is accelerating, and we’re all benefiting from the innovation it’s driving. While we’re still far from the science fiction vision of artificial general intelligence, the practical capabilities being delivered today are transforming how we work, create, and solve problems.
Frequently Asked Questions
Q: What are the main differences between OpenAI’s O3 and O4 Mini models?
O3 is OpenAI’s more powerful but expensive model that excels at multimodal tasks and complex reasoning, while O4 Mini is more affordable and surprisingly strong at mathematical reasoning. O3 tends to organize information in tables and has stronger visual capabilities, while O4 Mini High (with extended reasoning) outperforms on many mathematical benchmarks. O3 appears to be significantly more expensive to run, as evidenced by OpenAI’s stricter usage limits.
Q: How do these new AI models compare to human capabilities?
These models show a fascinating mix of superhuman and subhuman abilities. They can solve complex mathematical problems 10 times faster than human experts and perform graduate-level analysis of research papers. However, they still struggle with intuitive visual tasks that most humans find simple, like tracing lines in a diagram or quickly reading an analog clock. This suggests AI has surpassed humans in certain analytical domains while still lacking fundamental visual-spatial reasoning abilities.
Q: What are the privacy concerns with these new AI models?
The ability of models like O3 to identify specific restaurants from menu photos or locations from selfies raises significant privacy concerns. These models can now connect visual information with web data to make surprisingly accurate identifications without explicit permission. This capability suggests these systems have extensive knowledge of real-world locations and businesses that can be matched against user-provided images, potentially revealing more information than users intend to share.
Q: How is the video generation landscape changing?
Video generation is advancing rapidly with several key innovations: lightweight open-source models like LTX Vout that can run on modest hardware; new control mechanisms like WAN’s first/last frame control that allow for more precise direction; high-fidelity models focused on detail like V2 AI’s VU Q1; and memory-efficient approaches like Frame Pack that enable longer video generation with minimal VRAM requirements. These developments are making AI video generation more accessible and capable across different use cases.
Q: What’s the best way to use these AI models for coding tasks?
For coding tasks, accessing these models through APIs or dedicated tools like OpenAI’s Codex CLI generally produces better results than using them through ChatGPT’s interface. This is partly due to context window limitations in ChatGPT and pre-prompting that may interfere with coding tasks. Gemini models currently appear to have an edge in coding benchmarks, but OpenAI’s models accessed through API can be quite competitive. The open-source community has already modified tools like Codex to work with multiple model providers, giving users flexibility to choose the best model for specific coding tasks.
