July AI news

While in previous months we have mainly seen new model releases, July suggests a different trend: AI systems are starting to profile themselves more as tools for specific tasks. Agents, learning modes, more autonomy.

This month, in addition to the already confirmed news, we've got a few rumours that are currently circulating in the AI world, and it's only a matter of time before they actually happen.

OpenAI Agent

A new tool from ChatGPT that can autonomously perform actions across connectors.

• The agent has its own workspace that you can open directly in the chat.
• For Pro subscription users and above, it has access to all the connectors you have connected. For example, Google Drive, Calendar or GitHub.
• It can perform specific actions, such as creating a calendar event, ordering goods, analyzing documents, etc.
• It is semi-autonomous.
• The inspiration for the agent structure was a model that placed very well in the International Mathematical Olympiad (IMO).
• This is actually an iteration of OpenAI Operator, which basically worked similarly, but with a much worse UX.
• Useful links: Introducing ChatGPT Agent, Connector listů, List of possible actions

ChatGPT Study Mode

In addition to the agent, OpenAI also introduced a new Study Mode, which you can enable in GPT chat.

• ChatGPT takes on the role of a teacher or tutor.
• The mode allows you to go through the material step by step, at a pace that suits you.
• The GPT does not give you the answer right away, but tries to guide and engage you in thinking, and it can also test the user on the material.
• Useful links: ChatGPT Study Mode

Grok 4 and Grok 4 Heavy

Elon Musk's xAI released new versions of its Grok 4 and Grok 4 Heavy in July.

• Both versions are only available in the paid version.
• According to published data, they achieve the highest scores in the so-called Intelligence Benchmark Index (Artificial Analysis).
• Useful links: News Year 4

Gemini 2.5 Deep Think

Google has released an extension to its Gemini model line.

• The version is focused on better reasoning and multi-step planning.
• Deep Think should enable parallel testing of agents - that is, scenarios where multiple separate processes are running at the same time.
• Only available to Ultra subscribers.
• Useful links: GoogleAI Status

Extremely fast models from Cerebras

Cerebras hosts extremely fast language models that are significantly ahead of the competition in processing speed.

• The platform can host very large models.
• But it is also one of the most expensive options.
• In terms of generation speed and response, it is among the absolute top.
• Despite the high cost, it can be an interesting choice for companies that solve demanding tasks.
• Useful links: Artificial Analysis Status

And what are no longer verified news, but still worth mentioning:

Open-source model from OpenAI

From several hints and leaks, it seems that OpenAI is working on an open-source release model. Details are not yet confirmed, but:

• There is talk of a smaller language model that could be an alternative to open-source models like LLaMA or Mistral.
• The release could be a response to community pressure and competition - Meta, Mistral or Cohere already offer open models.

ChatGPT 5 available during autumn

• OpenAI hasn't officially announced anything yet, but has hinted several times that a "major update is near".
• Improvements in reliability, reasoning abilities and long-term memory are expected.

Meta offers astronomical sums to AI engineers

• You can reportedly get up to $1.25 billion for a four-year contract. That works out to some $312 million a year.
• Even so, many candidates reject them, but these are mainly employees of competing companies such as OpenAI.
• Useful links: Tom's Hardware Article

July showed a shift from new models to their real use.
AI is becoming more practical, autonomous and closer to everyday scenarios.