May was another busy month in AI. New models arrived, developer tools improved, and everything kept getting faster and cheaper. We're seeing a shift towards more efficient architectures, more capable models, and greater availability.
As we do every month, we've picked out the highlights and summarized what resonated the most in AI.
Claude 4 is out
- It comes in two variants: Sonnet is faster and cheaper, Opus offers better reasoning.
- In our experience, it is currently the best for programming.
- Even so, it doesn't top every benchmark.
Useful resources: Anthropic News Claude 4
Artificial Analysis released a report on the state of AI models
A great summary even for people who don't follow AI daily.
Key findings:
- Most models are moving to the MoE architecture.
- The gap between open-source and closed-source models keeps narrowing.
- Most models already handle reasoning.
- Models keep getting faster and more efficient.
- Chinese AI models are gaining ground.
Useful resources: Artificial Analysis State of AI 2025
Google I/O conference
At this developer conference Google showed the new Gemini 2.5 Flash, among other things. It stands out mainly for its price, performance, and speed.
Its results match GPT 4.1, but it is currently 3 times faster and cheaper.
Qwen 3 is out
A family of several models from Alibaba.
- They all use the MoE architecture and work on reasoning principles.
- These are small models, yet they place very well in benchmarks so far.
Useful resources: Artificial Analysis Alibaba Launches
OpenAI can search your repository
- Given access to your repository, it treats your code as a source of truth.
- That makes it easier to navigate the code.
- It now uses GitHub as a search source, so it can help you find bugs, refactor, describe architecture, or write documentation.
Google and OpenAI released SDK libraries
- The libraries support the MCP server. After a quick setup, they can use tools from your server.
- They make it easier to build agents and offer a great developer experience.
JSON schema standardization is missing
- The under-specified protocol and the still-leaky SDKs mean tools for one LLM may not be compatible with another, even when using MCP.
May kept up the momentum, with major players releasing new models and improving developer tools. The shift towards MoE architectures and the narrowing gap between open and closed models point to an increasingly competitive and accessible AI landscape.
Want to stay one step ahead?
Don't miss our best insights. No spam, just practical analyses, invitations to exclusive events, and podcast summaries delivered straight to your inbox.