The frantic pace of AI releases

Published on

A quick reflection: do you have any idea how many AI models (and versions) were released in 2025 alone? Listeners of the IA Sob Controle podcast will certainly have an idea, but inspired by a post from Nate Jones, I decided to put together this summary:

January:

  • OpenAI: Operator
  • Google: Gemini Flash Thinking
  • DeepSeek AI: DeepSeek-R1
  • Alibaba: Qwen 2.5-Max
  • Barcelona Supercomputing Center: ALIA-40B

February:

  • xAI: Grok 3
  • Anthropic: Claude 3.7 Sonnet
  • OpenAI: GPT-4.5 and Deep Research (using o3)
  • Alibaba: QwQ-Max
  • Google: Gemini 2.0 Flash and Gemini 2.0 Pro Experimental
  • Microsoft: Phi4-mini and Phi4 Multimodal

March:

  • Google: Gemini 2.5 Pro and Gemma 3
  • OpenAI: GPT-4o Image Generation
  • Alibaba: QwQ-32B and Qwen2.5-VL 32B

April:

  • Meta: Llama 4
  • OpenAI: GPT-4.1, o3 and o4-mini
  • Google: Gemini 2.5 Flash
  • Alibaba: Qwen 3

May:

  • Microsoft: Phi-4 reasoning
  • Anthropic: Claude 4 Opus and Sonnet
  • OpenAI: Codex
  • Google: Gemini 2.5 Pro, Jules, Gemini Diffusion and Gemma 3n
  • DeepSeek AI: DeepSeek-R1 (updated)
  • Maritaca AI: Sabiá-3.1

June (so far):

  • Google: Gemini 2.5 Pro (updated)
  • OpenAI: o3-Pro

And we’re still in the middle of June! And that’s only counting the top companies!

It’s a little crazy to think about it, and about how much capital (intellectual and financial) we’re seeing poured into this area. The release speed is frantic. We’re talking about an average of 3-5 “top” models per month, not to mention smaller updates to them.

You can also spot some trends. Practically everyone started focusing on “reasoning” models (now called LRMs), and more recently also on hybrid models, which know when to answer directly or when to “reason”.

It’s also super cool to see many of these companies (mainly the Chinese ones, but also Google, Meta and soon OpenAI) with state-of-the-art open-source models.

For those who work with AI or just follow the field, keeping up with this pace, staying updated, can be a… pretty arduous task. Being in the loop is super important, I’d say essential (since it’s possible that a new model solves that specific business problem of yours), but it’s also important not to get lost in unnecessary hypes.

It’s a task that Marcus Mendes and I try to make easier for you every week with the IA Sob Controle podcast. And I have to say it’s been a pleasure doing it :)