Images 2.5 is here.
Start with a graphic you need.

A better image model should make the correction easier

Give Images 2.5 a small, specific job you can actually inspect.

OpenAI introduced ChatGPT Images 2.5 on September 8. The company says it improves detail, targeted edits and consistency across repeated revisions, with faster generation. Those are release claims. The examples below show a small generation-and-edit workflow; there is no comparison against an earlier model here.

The update also brings templates, sketch-based input and comments on images. These are different ways to show what you mean when another paragraph of instructions isn't helping.

Where to start

In ChatGPT, ask for an image or choose More, then Images. You can upload an existing image and describe the change. In the mobile app, @Sketch lets you draw a starting layout. Templates are available from Images in the sidebar, but aren't yet available in Work mode.

The rollout covers all tiers. Some modes and features have different availability, and your plan's allowance still applies. Check what your account offers before planning a batch of graphics.

Try one graphic you already need

A resource cover is a useful first job because you can check the words, layout and readability. Here's a fictional workshop example you can adapt.

Read every word in the result. If the subtitle is wrong, ask for that specific correction and inspect the whole image again. OpenAI's help notes that edits can extend beyond the selected area.

After two correction attempts, decide whether to keep it, add the final words in your usual design editor, or use a simpler cover. That's a suggested work boundary, not a platform limit.

A graphic for this very Brief

For a second use, we made a share card for this issue using the Brief's name, date and cream, navy and terracotta colors. The title and date came through correctly. The illustration added a small landscape motif, although the prompt asked for abstract lines. That may suit this card, but it's another detail to inspect before accepting a result.

Library can bring more of your files into the conversation

Less repeated uploading helps, as long as you've selected the right source.

On September 10, OpenAI added Box, Dropbox and SharePoint alongside Google Drive in ChatGPT Library. Once connected, you can browse files and folders, add a source to a conversation and return to the original from citations.

It's a staged web rollout for eligible plans in Chat and Work. Your existing file permissions and workspace controls still apply. Having Library in your account doesn't mean every provider connection is available.

Use it on a source you already work from

Choose Add from Library or use an @mention to select the relevant file. If you're asked to connect or reconnect a service, read the permission request first. A folder is useful only when you actually need the files inside it.

If the outline contains two questions, the result should show those two. It shouldn't invent a third to make the handout look complete. And if you've got several versions, identify the current one before asking AI to work from it.

The useful part is halfway up the chat

You don't have to reopen every idea to return to one of them.

You work through a problem with AI, land on something useful, and leave the conversation to do the next thing. When you return, the chat has three possible directions, an abandoned outline and a suggestion that sounds familiar enough to feel like a decision.

Returning to parked ideas and repeating context came up in the Lab this weekend. It's a small set of conversations, but the friction is recognizable: you have to reconstruct the work before you can continue it.

A September 7 conversation on Using AI at Work, with Rana Gujral, asks a related question: are people making decisions, or simply approving what AI suggests? It's a practitioner discussion. I wouldn't turn it into a claim that AI use inevitably weakens judgment.

Recover one decision, with something to check it against

Start with the relevant, non-sensitive section of a chat you already know. A small excerpt is enough for this exercise.

Read the source before accepting the recovered decision. Then keep the accepted note with a link or excerpt that lets you check it again. A confident summary can still mistake a suggestion for something you approved.

Deep Research can end with something you can use

Start with the decision you need to make, then ask for the format that will help.

On September 9, OpenAI expanded Deep Research into ChatGPT Work and Codex. It can investigate questions across the web, supplied files and supported connected apps, then produce an editable document with citations. Depending on the tools available, you can also request a presentation, spreadsheet or Site.

For someone planning a workshop, that could mean comparing three venues and finishing with a short decision sheet. The useful part is being able to check the options against your requirements and keep the comparison you actually need.

Give the research a finish line

In Work, type @Deep Research or explicitly ask for deep research. Give it the question, your requirements and the output you want. You can answer clarifying questions and steer the research while it runs. For a single opening time or quick fact, an ordinary search is usually enough.

Before using the recommendation, open the sources behind the details that would change your choice. A listed room capacity doesn't confirm that the room is available on your date. If you requested a file, open it too; a promise to make the document isn't the finished document.

The pacing debate got more concrete

Watch what becomes independently verifiable.

Dario Amodei, Anthropic's CEO, is calling for a more measured pace of development for the most powerful AI systems. His proposal includes outside evaluators with ongoing access to check safety practices. Anthropic says it intends to bring in an external review team in the near future.

On September 9, OpenAI also called for mandatory national safety requirements tied to what AI systems can do, alongside industry standards. These are consequential positions from the companies building the technology.

They don't establish that the industry has agreed to pause, or that the proposed independent oversight is already operating.

Your useful document doesn't have to stay a PDF

Acrobat's new tools offer visual summaries, audio and a way to restyle plain documents.

Adobe announced new Acrobat capabilities on September 9, including interactive reports, summary slides, personal podcasts and audio summaries. The visual tools can turn dense source material into something easier to scan; the audio tools offer another way to review it.

The new Stylize feature also uses Adobe Express templates to reformat plain documents. Adobe says it preserves the source content as it changes the layout. That's a claim to check against your own original before using the result.

Start with a handout you already trust

If you've already got a clear workshop handout or service overview, try making a short visual summary for someone who needs the main steps at a glance. Choose one format first. A personal audio summary might help you prepare; a visual summary might help your attendee or client.

Compare the result with the source, especially anything an attendee needs to bring, do or understand. A cleaner layout is useful only if it still says what you meant. This is a suggested exercise; we haven't tested these Acrobat features for this issue.