ChatGPT Images 2.5, Resolve adds agents: Sept. 9

ChatGPT can now revise campaign images faster, DaVinci Resolve can hand editing chores to an AI assistant, and two new agents want a place in your workflow. Set aside 30 minutes today. You’ll add an image approval checkpoint, isolate an editing test, and put hard limits around any agent that can act for you.

ChatGPT Images 2.5 speeds up controlled edits

OpenAI released ChatGPT Images 2.5 on September 8 and says it cuts generation delay by up to 50% compared with Images 2.0. The official announcement also names better subject consistency, more precise edits, transparent backgrounds, and two application programming interface models. An application programming interface, or API, is the connection that lets your other software request an image automatically.

The useful change is revision control. You can ask for one change without rebuilding the whole image. That can shorten the trip to an approved ad, but product details, prices, spelling, and usage rights still need a person.

Open ChatGPT, upload one previously approved image, and request a single change. Then place a comment on the exact area to revise. Compare the new file with the approved original at full size. Record the prompt, source file, reviewer, and approval in the same tracker used by your AI content approval workflow.

If images already pass through a human before publishing, test the faster route. If nobody owns final approval, fix that gap first. Browse the wider AI tool runbooks after the checkpoint is in place.

DaVinci Resolve 21.1 connects AI assistants

Blackmagic Design released DaVinci Resolve 21.1 on September 8 with support for Claude, Claude Code, and ChatGPT Codex. According to the 21.1 release announcement, an assistant can analyze a project, organize media, change settings, create highlight edits, remove unwanted clips, and start batches of exports.

That turns plain-language instructions into actions inside the video project. The gain is repetitive work, especially sorting recordings and producing several output files. The risk is scope. “Remove the bad clips” is too vague when an assistant can alter the whole timeline.

Update Resolve, duplicate one non-critical project, and keep the original untouched. Connect only one assistant. Ask it to organize media first, inspect the result, then try a highlight edit with a fixed length and named source folder. Export to a new folder rather than overwriting finished files. Anyone making speech-led videos should keep the review habits in the Descript editing workflow, even when the editor changes.

Use this for repeatable prep and export work. Keep final cuts, claims, captions, and brand decisions under human control.

Mercury 2.5 offers a cheaper fast model route

Inception released Mercury 2.5 on September 8 at a listed price of $0.20 per million input tokens and $0.75 per million output tokens, with launch pricing 80% lower. Tokens are the small pieces of text an AI model reads and writes. The Mercury 2.5 announcement reports a 260,000-token context window and 1,107 generated tokens per second on widely available NVIDIA hardware.

This is a routing story, not a reason to replace your main writing model. Mercury is a diffusion language model, meaning it refines chunks of an answer in parallel instead of producing every word strictly from left to right. Its speed suits behind-the-scenes jobs such as sorting enquiries, rewriting search terms, or formatting data before another step uses it.

Duplicate one low-risk automation and send only its classification step to Mercury through Inception, Baseten, or OpenRouter. Save 50 representative inputs and compare wrong categories, completion time, and total cost with your current route. Do not test on live customer replies. The model earns a place only if it passes the same inputs repeatedly, a standard covered in the automation tool comparison.

Meta Muse acts across connected accounts

Meta launched Muse on September 8 with a dedicated online computer, its own browser, background tasks, and connections to email, calendars, Instagram, and other services. Meta’s Muse product page says the agent asks for approval before actions such as sending an email or making a purchase and keeps an activity record.

An agent is software that can take several steps for you, rather than only answer a question. Muse may handle personal research and administrative errands. It should not become an unofficial business automation just because it can reach the company inbox. A broad connection can expose data or purchasing access the task never needed.

If you test Muse, create one narrow task with no customer data and no spending authority. Open its connection settings, grant only the account required for that task, and leave approval on for messages and purchases. Review the activity record after every run. Keep established lead follow-up in a documented CRM automation build, where ownership and failure paths are visible.

On the bench

• ChatGPT Sketch can turn a rough drawing into an image. Test it for layout direction, not final copy.
• Mercury Voice claims a response start under 170 milliseconds. Measure the pause on your own call script before considering a switch.
• n8n published a reliability checklist for AI agents. Use it to audit retries, logs, and human escalation in one existing workflow.

Get the next build from the workflow runbooks.