OpenAI DevDay 2026: Practical Takeaways for Business AI Adoption

by: Zeb Anderson, Head of Artificial Intelligence

This week I attended OpenAI DevDay in San Francisco, an invitation-only developer conference bringing together engineers, researchers, and technical leaders building with AI. My focus was on what these developments could mean for the businesses we work with at TriVista: which capabilities are becoming practical, how teams can use them effectively, and what needs to happen before they can be trusted with important work.

There was plenty to get excited about, but the things that stuck with me were a mix of big announcements and surprisingly practical advice. Faster models can shorten development cycles. More persistent agents can take on longer tasks. But useful outcomes still depend on clear instructions, security, and checking the results.

I also met some fascinating people whose work I had followed from afar, including Theo.

Faster AI Models Make Iteration More Practical

Ultrafast was an easy one to get excited about. The presentation highlighted speeds of up to 300 tokens per second, meaning the model can generate text and code much faster. That speed comes at a premium, so whether it is worth it depends on the work and the value of the time saved. Watch the budgets in the near term.

What this means broadly is that the loop between having an idea, trying it, finding the problem, and trying again is becoming shorter and shorter. Software is becoming easier to create, adapt, and replace.

GPT-6.1 Sol was another big announcement. OpenAI positions it as approaching Astra’s capabilities in coding, computer use, and professional work at one-fifth of Astra’s standard input and output token prices.

In my early testing, my unofficial “tokens per right answer” metric improved significantly. It reached a useful result in fewer iterations and appeared better at delegating smaller tasks to smaller models.

For me, that makes the question more interesting than “Which model is smartest?” It becomes “What can we now afford to do repeatedly?”

AI Agents Need Some Gumption

OpenAI also introduced dots: always-on agents with their own computers, designed to work toward ongoing goals within the permissions you set. They come with names and characters, too. I’m still deciding how I feel about that part.

The Agents API also added computer use, continuing the trend of bringing more of Codex’s capabilities into developers’ own applications.

One practical takeaway from Peter Steinberger, known as @steipete and for his work on OpenClaw, really stuck with me: queuing up nudges to keep a model going can deliver outsized value. The discussion about building more of that gumption into models and their surrounding tools was fascinating.

He described queuing a series of encouragement prompts, essentially “keep going, you’ve got this,” to maintain momentum. It was a surprisingly simple approach to helping an agent continue working through a task.

It also reinforced why I care about the harness: the software around the model that manages context, tools, approvals, and execution. The model is only part of what makes an agent useful. As models become more comparable in capability, the systems around them matter more.

My takeaway is to be more deliberate about defining “done,” giving the agent useful feedback, and checking the result. Persistence is exciting. Persistence toward the wrong outcome is not.

Build Security Into AI Development

SECURITY.md caught my attention. As I posted on X during the day: “Finally armor for the projects.”

It gives teams a place to document a project’s security requirements, threat model, and boundaries alongside other project guidance. That context can help security tools understand what matters in a particular codebase.

Codex Security Cloud was another relevant development, bringing repository scanning, investigation of potential vulnerabilities, and proposed fixes into a cloud workflow. Together, these capabilities could give defensive security teams more support against persistent threats.

I want security to be part of how we build, rather than something we remember after getting excited about a working demo. Faster development cycles make that feel more important to me, not less.

Share the Prompt, Not Just the Output

Another Peter gem was: “Show me the prompt.”

My reaction: sharing your prompts is awkward for a day and then legitimately helpful.

I’d like to see more teams share how they asked, what context they supplied, and what they changed, not just the polished output. Seeing the prompts and revisions behind a result can help others understand why an approach worked and where it fell short.

It also gives people who are particularly effective at working with AI a practical way to share what they have learned and help their teammates improve.

There was plenty more to explore, including cloud development environments, a refreshed Codex CLI, plugin extensions, ChatGPT Space and Pages, Slack and Teams integration, and collaborative slides.

Other announcements will take a little more time to understand and test. As we get hands-on experience with them, I’ll share follow-up posts on where they prove useful.

But the prompt-sharing point is the one I’d bring back to a team immediately. Compare approaches on a real task. Talk about what failed. Improve it together.

What Business Leaders Can Put to Work Now

For the companies we work with at TriVista, I’d start there: pick a workflow that matters, define a useful outcome, set the boundaries, and measure whether the work actually improves.

Look at the time required, the quality of the result, the cost, and how much human review is still needed. A faster model or a more persistent agent is valuable when it improves the work.

For some software tasks, iteration cycles are now measured in days or hours. That creates opportunities to test ideas quickly and learn where AI can deliver measurable value.

Overall, I’m excited for the future. I met incredible builders and security professionals who understand what is at stake. Seeing how seriously they take these efforts left me encouraged about the work ahead.

Explore how TriVista helps business leaders evaluate and implement AI.