June and early July gave us a model release almost every day, which is exactly why you should not try to follow all of them. Most were incremental. A few will still matter next summer. Here is the month sorted the only way that is useful: by how long it will stay relevant.
Will matter in a year
A government switched off a frontier model
The biggest story was not a benchmark. On June 18, Anthropic shut down access to Claude Fable 5 and Mythos 5 after a US restriction on foreign-national access, then pulled the models more broadly. It is being called the first direct government intervention in who can use a frontier model. Fable 5 came back on July 1 with new safeguards, priced at $10 and $50 per million tokens. The specs were never the point. The precedent is: model access is now a policy lever, not just a pricing tier.
The price floor fell out
Open models kept dragging costs toward zero. Z.ai released GLM-5.2, a 753-billion-parameter mixture-of-experts model under an MIT license with a one-million-token context window. DeepSeek V4 landed at prices that still make Western labs wince. If your product margin depends on nobody noticing that near-frontier reasoning now costs cents, that is a problem you have this year, not next.
Might matter
Microsoft built its own models
At Build, Microsoft shipped seven MAI models, led by MAI-Thinking-1 (a one-trillion-total, 35-billion-active reasoning model trained from scratch) and MAI-Code-1-Flash, which went straight into GitHub Copilot. The interesting part is not the specs. It is the message to OpenAI: the biggest customer is now also a competitor.
OpenAI GPT-5.6
OpenAI began rolling out GPT-5.6 in early July. The upgrades are real. The naming remains a small act of violence against anyone trying to keep track. We will review it properly once it settles.
Probably a footnote
- Grok Imagine Video 1.5 makes video with native audio roughly twice as fast. Impressive in a demo, uneven in practice.
- Gemma 4 12B runs multimodal on a 16GB laptop, which is genuinely nice and will change exactly nobody roadmap.
- Nemotron 3 Ultra from NVIDIA, a 550B hybrid Mamba and Transformer model, matters mostly to people who already know why it might.
The one nobody is pricing in
xAI reportedly acquired the maker of Cursor for $60 billion in all stock. Whatever you think of the number, it tells you where the smart money thinks the value is: not in the model, in the thing wrapped around it. The model is becoming a commodity. The workflow is the moat. Keep that in mind the next time a lab tells you its training run is the real product.
If you remember one thing from June: the frontier stopped being a purely technical race and became a political and economic one. That shift outlasts any leaderboard.