I am starting a running note here, because keeping up with AI stopped being a monthly habit and became a weekly one. A week now holds what a year used to, and this past week is the proof. Five real shipments, from five different labs, and the ground moved again under everything I do for brand, content, and search.
OpenAIGPT-6 Astra
A new flagship model, out September 4. OpenAI called it the most intelligent in the world, then independent benchmarks scored it level with the model it replaced at more than double the price. For a brand team the move is not to rush out and rebuild your content or search stack on the shiny thing. Pick the model by cost per finished task, because the cheaper one often serves your audience just as well and protects the budget you would rather spend on the work itself. Read the evidence, not the launch copy. See OpenAI's announcement and the independent numbers.
AnthropicClaude 5.1
A price cut, not a new feature, on September 1. Anthropic dropped the cost of re-reading cached context by 75 percent, from a dollar to a quarter per million tokens. Dull on the surface, but when the cost of running content through AI falls, the campaigns and pages you once shelved as too expensive to produce quietly become worth doing again. That is how a small brand team starts to output like a big one. See Anthropic's pricing, and why that matters when AI belongs inside the work.
GoogleGemini 3.8 Flash
A new fast and cheap tier, out September 2, the third in six weeks. This is the workhorse you reach for a hundred times a day, the one behind the bulk brand work like product copy, metadata, alt text, and first drafts at scale. One thing to diary. Google holds this price only through December, then doubles it on January 1, so cost any workflow that leans on it against the January number now rather than the launch number. See Google's announcement.
MetaMuse Spark 1.3
A new challenger, also out September 2, that matched the frontier and undercut it hard, running more than 90 percent cheaper if you let Meta train on your usage, which is a trade a brand should weigh before taking it. It also asks before it does anything consequential, which is what makes it safe to hand an agent real work like drafting outreach or prepping a publish. The bigger point is that frontier AI is no longer only for big budgets, so a lean team can operate at a level that used to take headcount to reach. See Meta's writeup.
PerplexityHybrid Compute on Mac
A new privacy feature, shipped September 1. It runs the sensitive steps and any access to your private files locally on the Mac, behind an on-device filter that catches personal information before it can leave. That means you can point AI at the things you would never paste into a chat box, like unreleased campaigns, a customer list, or brand guidelines under NDA, and get the help without the exposure. It reopens a whole category of work that legal and security used to veto on sight. See how it works.
None of this existed on my radar eight days ago, and that is the job now. The brands that win the next year will not be the ones with the biggest AI budget. They will be the ones who notice what shipped this week and put it to work before their competitors finish reading the headline. That is why I file these every week, and why you should skim them the way you skim the market. The floor keeps moving. The only wrong move is to stand still on it.