Generative AI – 2025 Summer Recap

Leave a Comment / By / 28. July 2025

While on vacation this summer I was just flabbergasted by the amount of new releases, features, product announcements and improvements that was released. Therefore I decided that I wanted to write a recap on the biggest announcements and what the overall trends that I am seeing in the market.

Firstly I want to point out the different vendors in the market and that Google is clearly doing something right with the direction they are taking. They are now processing 980 trillion+ monthly tokens across their products and APIs (up from 480T in May) so clearly getting more and more adoption. This of course mostly because of the launch of their video generation model (Veo-3)

Secondly It should be noted that I am seeing more and more open-source based models coming from China, such as the latest versions of Qwen (from Alibaba) and Kimi2 which is a small 32B parameter LLM using MoE and scores almost the high as closed source models https://github.com/MoonshotAI/Kimi-K2

Below is the score from Kimi-K2-instruct compared to DeepSeek, Claude 4 Sonnet and GPT 4.1

And now with Grok 4 from xAi, Elon Musks little AI Baby has also taken big leaps in terms of performance, being named as the nr 1 model within many different categories (for the time being, before GPT 5 get released)

General announcements and miscellaneous

  • White House AI Action Plan – https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf Which described how the US is going to win the “AI-Race” making it easier to build out new infrastructure and allow for easier export of US developed AI. This AI action plan is also ensuring higher focus on Open-Source and Open-Weight AI.
  • Mistral announced a report showcasing the environmental impacts of GenAI https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai. While many have been discussing “how much water does GenAI” use. In this report Mistral shows that by using their model “Large 2 with an output of 400 tokens (circa 350 words) uses 45mL of water. While this is a lot less the first assumed, it is important to note that for instance using Claude or any of the large providers have a system prompt closer to 15,000 tokens that is being sent to the service each time, meaning that the consumption is a lot larger.
  • An “hacker” was able to compromise the Amazon Q Extension to VS Code by adding wipe commands to the extension. https://www.404media.co/hacker-plants-computer-wiping-commands-in-amazons-ai-coding-agent/ while VS Code has built-in safe guards that would prevent this from happening, it shows now that will the introduction of MCP and “natural voice” commands that malicious attacks will soon just be natural language embedded into an integration.
  • Perplexity released Comet, which they have dubbed as an AI first browser which combines Perplexity, agents and browser use integration to interact with the local browser to perform tasks, similar to OpenAI Agents (currently only available for Max subscribers or waiting line for everyone else) https://comet.perplexity.ai/
  • Google and Windsurf entered a somewhat partnership, where many of the top people joined Google https://windsurf.com/blog/windsurfs-next-stage. Windsurf on the other hand released a new wave of features ( called wave 11 ) where they among other things now added voice mode to their IDE and Windsurf is now part of Cognition)

OpenAI

  • While a lot is going on at OpenAI, since they are now in the process of publishing their first open source model including next generation of GPT (GPT-5) which is scheduled to be released in August.
  • OpenAI announces a new partnership with Oracle ( 4.5 GW partnership with Oracle) where OpenAI will be using Oracle Cloud Infrastructure to host their services as part of Project Stargate.
  • ChatGPT Record mode – With record mode, ChatGPT can transcribe and summarize audio recordings like meetings, brainstorms, or voice notes. Available for Plus, Enterprise, Edu, Team, and Pro workspaces, and only available for the macOS desktop app
  • ChatGPT Study Mode – Allows you to use ChatGPT to study on a certain topic https://www.bleepingcomputer.com/news/artificial-intelligence/openai-confirms-chatgpts-new-study-feature-helps-with-exams/
  • ChatGPT Agent Mode – Which has now been released and available for most subscriptions, allows you to define a set of complex tasks that OpenAI can run within a virtualized browser (using the previous Operator feature) and using connected sources. ChatGPT Agent mode uses a specialized fine-tuned model which improves the hit rate on agentic tasks.
  • OpenAI also announced high fidelity input parameter on image generation https://community.openai.com/t/image-generation-high-fidelity-editing/1317649
  • OpenAI also announced a list supported connectors, meaning data sources that ChatGPT can connect with https://help.openai.com/en/articles/11487775-connectors-in-chatgpt

Microsoft and Copilot

Also more Agents coming from Microsoft in the upcoming months

Grok

  • Grok 4 was released, supports voice mode, photo and companions and supports deep reasoning and is currently leading the ARC-AGI leadership board (https://arcprize.org/leaderboard) however as before, they have gotten some criticism because of the lack of guardrails, just follow this person on twitter/x you get some more info about the prompt injections attack https://x.com/elder_plinius
  • Grok does not have a native CLI, but the team behind superagent created GrokCLI https://github.com/superagent-ai/grok-cli
  • Grok also has App Connections and Task Automations (similar to Scheduled Actions with Gemini or OpenAI)
  • Grok has companions that are “digital avatars” that you can talk with.

Anthropic

MCP

While MCP has now become a universal standard for how to integrate GenAI with 3.party actions/tools and data sources, I wanted to add some updates here as well since there has been big changes to the protocol with the new standard that was released back in March (which now supports Remote MCP servers) and now using OAuth.

AWS

Google

Qwen (Alibaba)

  • WAN coming! which is new video & image Generation Model from Alibaba Cloud read more about it here –> https://x.com/Alibaba_Wan (Seems to be similar to Sora)
  • Qwen Code https://github.com/QwenLM/qwen-code CLI tool similar to Claude CLI, that provides the same agent mode integration.
  • Qwen-MT -Translation https://qwenlm.github.io/blog/qwen-mt – New LLM aimed at doing translation, pretty accurate on Norwegian language.
  • Qwen3-Coder-480B-A35B-Instruct — a 480B-parameter Mixture-of-Experts model with 35B active parameters which supports the context length of 256K tokens. It also scores pretty high on SWE-Bench,
  • Alibaba also released Qwen3-235B-A22B-Thinking-2507 a new reasoning model with 256k context length and has similar scores to Gemini 2.5 Pro and OpenAI O4-mini. The best part that with most Alibaba based models is that they are open-source!

Mistral

  • Deep Research in Le Chat
  • Voice Mode (using Voxtral) new open source model for speech understanding model https://mistral.ai/news/voxtral (Supports English, Spanish, French, Portuguese, Hindi, German, Dutch, Italian)
  • Advanced Image editing (Black Forest Labs Partnership)
  • Devstral (New open-source model for coding) https://mistral.ai/news/devstral
  • Mistral Code https://mistral.ai/news/mistral-code which is a new platform for GenAI Coding that provides inferencing, access to open-source LLMs using the Continue extension in VS Code. This service can either be deployed using self-hosted or provided from Mistral directly. In combination with this Mistral is also offering their own Compute capacity called Mistral Compute https://mistral.ai/news/mistral-compute

Leave a Reply

Scroll to Top