September was the month “thinking” stopped being a metaphor. OpenAI shipped a model that visibly reasons before it answers, Meta pushed capable multimodal models into the open, and California’s marquee AI safety bill died on the governor’s desk. Underneath the headlines, the real story for operators is that the menu of what you can build with—and what it costs—changed meaningfully in thirty days.
Reasoning Models Go Mainstream
OpenAI released o1-preview and o1-mini on September 12. Unlike GPT-4o, these models spend time generating an internal chain of thought before responding, trading latency and token cost for sharper performance on math, science, and multi-step coding problems.
Why it matters for builders: o1 is not a drop-in replacement for your existing chat model—it’s slower and pricier, and it’s overkill for summarization, classification, or customer support. The right pattern is selective routing: send genuinely hard, multi-step reasoning tasks to a model like o1, and keep fast, cheap models for everything else. Treating “smarter” as “use it everywhere” is how you blow up both your latency budget and your bill.
Open Models Caught Up on Vision
Meta launched Llama 3.2 at Connect on September 25. The release added its first open multimodal models (11B and 90B vision) plus lightweight text models (1B and 3B) small enough to run on phones and edge devices—all with a 128K-token context window.
Why it matters for operators: this narrows the gap between “open” and “frontier” for a large class of practical tasks—document understanding, image reasoning, on-device assistants. If you’ve been defaulting to a closed API for everything, it’s worth re-running the build-vs-buy math. For workloads with privacy constraints, predictable volume, or latency requirements, a self-hosted Llama model may now be both cheaper and good enough. The answer is workload-specific, which is exactly the point.
A Glimpse of Genuinely New Product Surfaces
Google’s NotebookLM shipped Audio Overviews in September, turning your uploaded sources into a surprisingly natural two-host podcast-style conversation. It went semi-viral for good reason: it’s a clear example of generative AI creating a product experience that simply didn’t exist before, rather than bolting a chatbot onto an existing one.
Why it matters: the durable wins from AI aren’t “add a chat box.” They’re new interfaces that wouldn’t be possible without the model. When you’re prioritizing your roadmap, ask which features become possible now—not just which existing workflows get a marginal speedup.
Regulation Stalled—For Now
On September 29, Governor Newsom vetoed California’s SB 1047, the most ambitious US attempt to regulate frontier AI models. The bill had cleared the legislature in late August; Newsom argued it focused on model size rather than actual deployment risk.
Why it matters for founders: don’t read the veto as “regulation is off the table.” It signals that scrutiny is shifting toward how AI is deployed and where—high-risk use, sensitive data, consequential decisions. Building basic governance now (knowing what data flows into models, logging AI-assisted decisions, keeping humans in the loop where stakes are high) is cheap insurance against whatever lands next.
Leadership Churn at the Frontier
In late September, OpenAI’s CTO Mira Murati resigned, alongside its chief research officer and a research VP—a notable exodus during the company’s shift toward a more conventional corporate structure.
Why it matters: your vendor’s org chart is a real dependency. None of this means abandon OpenAI, but it’s a reminder to avoid hard architectural lock-in. Abstracting your model layer so you can swap providers is no longer a theoretical nicety—it’s basic resilience.
The Takeaway
September’s lesson wasn’t any single launch—it was that capability, openness, and cost are all moving at once. The teams that win aren’t the ones adopting every release; they’re the ones making deliberate choices about what to use where.
If you’re weighing how these shifts fit your roadmap, let’s talk about what to build, what to buy, and what to skip.