May AI news

May was again very busy in AI. New models appeared, tools for developers improved, and it continued to get 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 summarised what we think resonated the most in AI.

Release Claude 4

• It comes in two models Sonnet, faster and cheaper, and Opus with better reasoning capabilities.
• In our experience, it is currently the best for programming.
• Still, in some benchmarks it doesn't perform at its best.

Useful resources: Anthropic News Claude 4

Artificial Analysis released a report on the state of AI models

• This is a great summary even for those who don't do AI on a daily basis.
• For example, the main outputs are as follows:
  • most models are moving to MoE architecture
  • the gap between open-source and closed-source models continues to narrow
  • most models already handle reasoning
  • progress in model efficiency and speed continues
  • the growing potential of Chinese AI models

Useful resources: Artificial Analysis State of AI 2025

Google I/O Conference

• Developer's conference, where for example the new version of Gemini 2.5 Flash was presented. This version is very interesting especially in its price, performance and speed ratio.
• It is equal in performance to GPT 4.1, but is currently 3 times faster and cheaper.

Release Qwen 3

• A family of several models from Alibaba.
• All of them use MoE architecture and all of them work on reasoning principles.
• These are small models, but so far they are doing very well in benchmarks.

Useful resources: Artificial Analysis Alibaba Launches

OpenAI can search your repository

• With access to your repository, it can use your code as a source of truth.
• This makes it easier to navigate the code.
• It now uses GitHub as a search resource. This can help you with bug finding, refactoring, architecture description or documentation.

Google and OpenAI released SDK versions of libraries

• Libraries support the use of the MCP server. They can use tools from your server after a very short setup.
• They make it easier to build agents and provide a great developer experience.

Lack of JSON schema standardization

• The under-specified protocol and so far leaky SDKs lead to the fact that tools for one LLM may not be compatible with the other even when using MCP.

---

May showed continued momentum in the AI space, with major players releasing new models and improving developer tools. The shift towards MoE architectures and the narrowing gap between open-source and closed models signal an increasingly competitive and accessible AI landscape.