The Signal: From Answers to Things You Can Make

September 10 AI news: Images 2.5, the Agents API, GPT-Live-1, business data and AlphaGenome Atlas, with practical experiments and primary sources.

September 10, 2026 / The next question is: what will you make?

A useful way to read this week’s AI news is to follow the output. An image becomes something you can revise. A conversation becomes a longer piece of work. A question about a business becomes a dashboard. This edition picks up after our September 7 briefing and reviews selected announcements through September 10. The product details come from primary sources; the experiments below are our suggestions, not independent performance tests.

Images: the second version matters as much as the first

On September 8, OpenAI announced ChatGPT Images 2.5, emphasizing more precise edits and better consistency across turns. Its announcement also describes drawing with Sketch, image comments, templates and shared prompts, with a rollout across ChatGPT, ChatGPT Work and Codex. These are provider descriptions; the interesting test is whether an edit preserves the details you asked to keep.

Try an imaginary place: a tiny reading room on an island, a rooftop greenhouse or a museum of forgotten inventions. Establish the scene, then request exactly one change. Keep the camera angle, windows and colour palette fixed while moving the entrance. Compare both versions. A beautiful new picture can still fail that request if it quietly changes the building.

We used image generation to make an original Curiosity Observatory concept for this site. Then Astra helped build a separate interactive 3D interpretation. The distinction matters: the artwork is an illustration, while the browser world has actual controls and downloadable output. You can inspect both and decide what you would change next.

Sources: OpenAI · ChatGPT Images 2.5, September 8, 2026

Open the concept and the working world

Agents: the surrounding software is part of the capability

OpenAI’s September 10 Agents API announcement describes a public beta that brings the Codex harness to developers, including sandbox choices, long sessions, tools and optional subagents. A harness is the software surrounding a model: it manages the work, the available tools and the ongoing context. Better answers alone do not explain whether a system can finish a project.

For a first experiment, make the job small enough to inspect. Give an assistant a fictional event brief and ask for a plan, a draft and a check against the supplied room capacity. Omit that capacity once. Does the workflow identify the missing fact, or confidently produce an invitation? Add the evidence and see whether the result changes for the right reason.

Our Agent Workbench turns that question into a local rehearsal. It reveals authored steps rather than calling a model, so it does not demonstrate the API’s performance. It does make the design issue tangible: a draft, a reviewed draft and an action in the world need different evidence and permission.

Sources: OpenAI · Agents API, September 10, 2026

Rehearse a small agent workflow

Voice: can you change your mind mid-sentence?

Also on September 10, OpenAI introduced GPT-Live-1 in the API. The announcement describes a voice layer that can listen and speak at the same time and delegate deeper reasoning or tool calls to a backend model. This is a developer release; it does not mean every existing voice application has adopted it.

Natural conversation is full of corrections. Ask a voice assistant to rehearse a fictional phone call, interrupt to change the date, then pause before completing your thought. Check whether it keeps the correction and gives you room to speak. Read the transcript if one is available. Pleasant delivery can make a missed detail harder to notice.

A useful family or learning project could be a character who helps you practise explaining a topic aloud. Keep the character’s knowledge bounded to material you can check. The lesson is not to make a voice sound certain; it is to make the interaction responsive while retaining an understandable record of what happened.

Sources: OpenAI · GPT-Live-1 in the API, September 10, 2026

Business data: ask to see how the answer was made

OpenAI’s new Data agent in ChatGPT Work connects approved business sources to analysis, interactive dashboards and approved actions. The announcement says administrators control connections and roles, and queries follow the connected account’s permissions. It also describes using business definitions and semantic layers, the shared rules that explain what a metric means.

Before making a real reporting workflow, build a tiny example you can calculate yourself. Include one missing value, a cancelled order and two possible definitions of a customer. Ask which rows were counted and which definition was used. An attractive chart is much more useful when you can follow the arithmetic and explain an exclusion.

The same habit applies beyond spreadsheets. In our fictional Evidence Atlas, you can remove a source and watch the support for a claim disappear. The interface makes the dependency visible. In a real analysis, ask for equivalent access to the source passage, calculation or query behind the conclusion.

Sources: OpenAI · Data agent in ChatGPT Work, September 10, 2026

Try following a claim back to its evidence

Science: a map of predictions opens different questions

Google DeepMind announced AlphaGenome Atlas on September 8: a large collection of predicted molecular effects for nine billion single-letter human DNA variants. The important word is predicted. A resource for investigating biological questions is not a clinical diagnosis or proof that every predicted effect occurs in a person.

This is a different kind of AI story from a new chat feature. Its value lies in what researchers can investigate with a structured prediction resource. For a general reader, the useful exercise is to separate three sentences: what was released, what was predicted and what was experimentally established. Those sentences should not collapse into one another.

When reading a scientific AI announcement, follow the linked research and look for the intended use and evaluation setting. A striking scale figure tells you how much material exists. It does not, by itself, tell you how accurate a particular result will be or whether it is suitable for your question.

Sources: Google DeepMind · AlphaGenome Atlas, September 8, 2026

Choose one thing to make this week

Start with an output you would actually enjoy keeping. A small world. A clear map of six claims. A workflow that stops at the right moment. Specify the controls, the evidence and the file you want at the end. Then inspect the result on a phone as well as a larger screen.

That is the idea behind the new Possibility Studio. We built original experiments for this site and wrote down the design choices so you can adapt them. They are browser programs made with AI assistance, not live intelligence concealed behind a button. Open one, change something, and use the walkthrough to begin your own version.

Step into the Possibility Studio

Take this into your next task

Read announcements for the new possibility, then choose a small output whose behaviour you can inspect. Keep sources, controls and permission visible.

Build a project brief.

Questions people ask

Were these products independently benchmarked here?

No. This briefing reviews primary announcements through September 10, 2026. Provider claims are attributed, and suggested experiments are not recorded test results.

Does the Studio call an AI when I use it?

No. Its interactive exhibits are local browser software with deterministic rules and fictional teaching data. AI assisted their creation.

Sources and editorial notes

Reviewed 2026-09-10. Product capabilities depend on the app, plan, region and workspace settings. Workflows and prompts are authored teaching material, not recorded model results or measured time-saving claims.

Keep going

Related Power of AI pages

Keep reading with Start here, everyday uses, the tool guide, the writing workshop, and sources and standards.