OpenAI shipped ChatGPT Images 2.5 on September 8, 2026, and it changes what a 10 to 50 person company can produce in-house without a studio, a photographer, or a design agency on retainer. We have spent the past week running our own listing photos, ad concepts, and client mockups through it. Here is what actually holds up, what the two-tier API pricing means for your budget, and what we are building into client systems this quarter.
What Actually Shipped
ChatGPT Images 2.5 replaces the image model behind ChatGPT, ChatGPT Work, and Codex, across every subscription tier, on desktop, mobile, and web. OpenAI's own release notes claim richer textures, more natural lighting, and better preservation of the subject in a reference photo. The change that matters most for operators is consistency across edits. Ask the previous model for five variations of the same product shot and it would drift: faces changed, logos warped, backgrounds shifted between versions. ChatGPT Images 2.5 is built to hold the subject steady across a multi-turn conversation. Edit three, edit four, and edit five still look like the same product, the same person, the same room.
OpenAI also reports up to 50% lower generation latency compared with the prior model. That is the difference between a marketing team waiting on a render and a marketing team iterating on a layout in real time, during a client call.
Sketch: A Creative Input, Not a Gimmick
The headline new feature is Sketch, invoked by typing "@Sketch" inside a ChatGPT conversation. You draw a rough layout, a composition, a pose, and the model treats your drawing as a reference rather than a finished answer. OpenAI shipped starting templates for common formats like posters and merch, and you can drop comments directly on a generated image to request a targeted fix instead of rewriting the whole prompt from scratch.
For an operator, this moves the creative bottleneck somewhere more useful. Instead of briefing a designer with words and hoping they land on the composition you had in mind, you sketch the composition and the model fills in the polish. That is a workflow we can wire straight into an automation build: a sales or ops person sketches a layout, the system returns generated variants, and a human approves before anything ships to a customer.
Two API Models, One Price Card
Underneath the ChatGPT product, OpenAI shipped two API models built on Images 2.5: Flare and Sunburst. Flare is the default, tuned for everyday production speed. Sunburst is built for precision editing, the workflows where the extra generation time is worth paying for. Both price identically on OpenAI's published rate card: $8 per million tokens for image input, $2 for cached image input, $5 for text input, and $30 per million tokens for image output.
Token-based pricing does not map cleanly to a per-image cost, because a high-resolution edit with a detailed reference photo consumes more tokens than a quick low-resolution draft. The practical rule for a small business budget: draft and iterate on Flare, then run only the approved final version through Sunburst. Running every rough draft through the more precise, more expensive model is the fastest way to burn a content budget on variations nobody ends up using.
For scale, a single edit built from a modest reference image and a short prompt costs a few cents in token spend, not a contractor's day rate. Ten variations of a listing photo, run through Flare, cost less than the coffee run before Monday standup. That budget math changes what is worth automating and what still belongs with a professional photographer for the one hero shot that matters.
Flare vs Sunburst, in plain terms
- Flare: the default model, built for speed, right for first-pass drafts and volume iteration.
- Sunburst: built for editing precision, right for the one final asset a client or customer actually sees.
- Both bill the same per-token rate, so the real cost difference comes from how much you iterate, not which model you pick.
What This Means for a 10 to 50 Person Company
We run a 40-person real estate group alongside the systems we build for clients. The honest answer is that product and listing photography has been one of the slower, more expensive parts of the marketing pipeline for businesses our size.
A real estate listing, a product catalog, or a service business's before-and-after gallery all need the same thing: consistent, on-brand imagery, produced fast, without booking a studio day every time something changes.
ChatGPT Images 2.5 does not replace a real photoshoot for a hero image. It does replace the second, third, and fourth version of that shot: the variant sized for a different ad placement, the seasonal refresh, the version with different lighting for a different market. That is exactly the kind of repetitive creative work that eats a marketing coordinator's week without moving revenue.
We track this pattern across the revenue leaks we map for clients: hours lost to production bottlenecks, not just to slow sales follow-up or a missed lead. A marketing team stuck waiting on an agency for routine image variations is losing the same kind of time a sales team loses when a lead sits unanswered past the five-minute mark.
We see the same shift in the voice agents we build for client reservations and intake: the system handles the repetitive first pass, and a person only steps in for the exception. Image generation is now following the same pattern for creative production.
What We Are Building This Quarter
Three pieces are going straight into client builds this quarter.
- A sketch-to-draft pipeline where a sales or ops person sketches a rough layout and the system returns Flare-generated variants for quick review, before a single Sunburst pass on anything that ships externally.
- A reference-photo library per client, so every generated image starts from an approved subject rather than a blind prompt. That is what keeps a product, or a listing's rooms, looking like themselves across dozens of variations.
- An approval gate before anything publishes. A generated image is a draft, not a final asset, until a person signs off. That is the same rule we apply to every automation we build for a client: the system does the repetitive work, a person makes the call before anything goes out the door.
None of this is a standalone tool bolted onto a marketing calendar. It is one more input into the same operating system we install for clients: something that connects to the tools you already run and takes the repetitive work off your team's plate. If you want to see where that applies to your business specifically, our assessment maps exactly where creative, ops, or sales workflows are leaking hours. The cost is credited to the build if you move forward with us.
The Part Worth Being Careful About
Token-based pricing means cost scales with resolution and how much reference material you feed the model. A high-resolution image built from a detailed reference photo is not the same cost as a quick square draft, so budget for iteration, not just for the final asset. Treat every output as a draft your team reviews before it reaches a customer or a listing page. Consistency in the model does not mean the brand check is automatic. It still needs a person who knows the brand looking at the result before it goes live.
Our case studies show how this kind of build plays out end to end: a first system live within days to weeks, and the client owning 100% of the code and files once it ships. Image generation is one more input into that same model, not a separate tool bolted on the side.
If your team is still paying a photographer or a design contractor for every routine product shot or listing variant, that line item is worth a second look this quarter. We cover releases like this as they land, on the blog, mapped to what they actually mean for a business our size.