Product Photos, Social Posts, Offers: What the New AI Image Model Means for Your Marketing Budget

On September 8, 2026, OpenAI rolled out ChatGPT Images 2.5, its new flagship image model – in ChatGPT, ChatGPT Work and Codex. Before the announcement had even been fully digested, industry newsletters and LinkedIn were already circulating figures like "twice as fast and 72 percent more efficient." As a business owner, though, the more interesting question is a more sober one: what does this model leap actually mean for your marketing budget – for product photos, social posts and offer visuals? And which of the numbers going around are even accurate?

In this article, we break down what OpenAI actually changed, where the widely cited "doubling" claim really comes from, what an AI image actually costs today – and what you're now required to consider under the EU labeling requirement before using AI-generated images in your advertising. Because in this case, there's a lot more research standing between a sensational headline and a sound budget decision than most newsletters would have you believe.

What's new in ChatGPT Images 2.5?

According to OpenAI, ChatGPT Images 2.5 is noticeably better than its predecessor in three main areas: more natural lighting, richer textures, and more reliable preservation of people and objects from reference photos. That last point in particular matters for marketing purposes – if you upload a product photo as a starting image and want to generate variants for different channels, the model is meant to recognize the product more consistently across multiple editing rounds than before.

On top of that come three new features that directly affect the workflow: "@Sketch" turns a rough sketch plus a text instruction into a finished image, ready-made templates like "Poster" and "Merch" speed up recurring formats, and focused image comments allow targeted revisions to individual areas of an image instead of regenerating the whole thing. For developers and agencies integrating image generation into their own workflows, OpenAI has also released two new API model variants: "GPT-Image-2.5 Flare" – small, fast, with quality roughly on par with GPT-Image-2 but at about half the latency – and "GPT-Image-2.5 Sunburst," the base model, which surpasses GPT-Image-2 in image quality but works more slowly.

This split into two variants is the genuinely practical part of the announcement for your marketing budget: going forward, you can deliberately choose whether to use the cheaper, faster Flare model for a quick batch of social media variants, or the slower but more detailed Sunburst model for a high-quality hero image in your online shop.

Faster, but how much faster? A fact check

Many summaries of the announcement have been circulating the claim that ChatGPT Images 2.5 is "twice as fast as GPT-Image-2 and 72 percent more efficient in GPU usage." On closer inspection of the sources, though, this figure can't actually be attributed to OpenAI. It comes from Microsoft AI – specifically from a statement by CEO Mustafa Suleyman on X and Microsoft's official blog post about its competing model, "MAI-Image-2.6-Flash." Microsoft's own precise wording there is: "2.8 times faster than GPT-Image-2-Medium" and "72 percent more GPU-efficient" – with the explicit baseline being the mid-tier quality level of OpenAI's previous model, not OpenAI's own new top-tier model.

OpenAI itself cites a latency reduction of "up to 50 percent" for ChatGPT Images 2.5 compared with the predecessor, Images 2.0 – not the doubling in speed that's been circulating. The 2.8x and 72-percent figures refer to Microsoft's separate competing model, MAI-Image-2.6-Flash, not to OpenAI's own update.

Why does this distinction matter to you beyond mere pedantry? Because it actually explains the real market dynamic better than the misattributed figure does. It's not a single provider suddenly making everything twice as fast – it's competitive pressure between several major providers driving down costs and speeding things up industry-wide. OpenAI is improving its flagship model's quality and latency; Microsoft is positioning a separate model explicitly tuned for speed and cost against it. For your company, that means: the AI image generation market as a whole is moving fast – but the concrete speed gain depends on which model you actually use.

What an AI image really costs – compared with traditional photography

API pricing for GPT-Image-2.5 has stayed unchanged from its predecessor, GPT-Image-2: $30 per million output image tokens, $8 per million input image tokens, $2 per million cached image input tokens, and $5 per million text input tokens. Depending on quality level and resolution, third-party calculations put the resulting cost at roughly $0.03 to $0.08 per generated image – sometimes noticeably less at lower quality tiers.

Microsoft's MAI-Image-2.6-Flash is deliberately positioned to be cheaper here: $1.75 per million text input tokens, $2.50 per million image input tokens, and $19 per million image output tokens – compared with $5, $8 and $38 for its larger sibling model, MAI-Image-2.6. That's noticeably cheaper than GPT-Image-2 or 2.5, and shows just how much pricing pressure is now at work in the market.

Set these figures against traditional product photography in Germany, and the leverage becomes tangible: a simple background removal costs from around €6.50, studio photos run €50 to €200, and freelancer sessions €30 to €100 per image. Booking a photographer for a day means day rates of €70 to €250 net per hour, or flat day rates of €500 to €2,000. Against that stand AI-powered subscription models that promise essentially unlimited images for €10 to €50 a month – sometimes advertised by providers as an 80 to 90 percent saving over traditional photography. Those percentages come from the providers' own marketing material and should be read as a rough indicator, not an independently verified study – but the general direction of the cost advantage is plausible given the per-image API pricing.

Important for your budget planning: the real value of the new model generation lies less in a further reduction of the per-image price – after all, GPT-Image-2.5's API pricing is unchanged – and more in iteration speed. Up to 50 percent lower latency means, in practice: more image variants for a seasonal campaign, more A/B-test creatives for social posts, and faster reaction to current offers – in the same amount of time that used to be enough for fewer variants.

A concrete example makes this tangible. A trades business that wants to prepare ten reference photos of completed projects for social media ahead of the summer season would traditionally pay several hundred euros quickly for a photographer session plus post-processing – or skip it altogether for lack of budget and time. Through an AI-powered workflow, a handful of phone photos can be turned into several polished variants in different formats for Instagram, Facebook and the company website in a short amount of time, at a fraction of the traditional cost. A retailer promoting new offers every week can vary product images with different backgrounds, seasons or occasions without booking a new photoshoot every time. And a coach or consultant who regularly publishes social posts can generate accompanying visuals in the time it used to take just to search stock photo databases. In all three cases, the decisive lever isn't the cost per image, but the ability to produce content regularly and in multiple variants at all – something that would simply be uneconomical for most smaller businesses with traditional photography alone.

That said, not every use case is equally suited to purely AI-generated images. For background removal, lifestyle scenes and social media variants, current models are considered production-ready. For premium products or your central hero image in the shop, several German industry sources still recommend a hybrid approach: a real reference photo as the starting point, subsequently refined, cut out or placed into different contexts using AI tools like @Sketch or targeted image comments. This distinction – when a purely AI-generated image is enough, and when you need a real photo as a base – is itself a budget-relevant decision best made once, up front, for your most important image categories, rather than re-deciding it for every single image.

The labeling requirement: what you need to watch out for since August

Anyone reflexively switching to cheaper, faster AI images shouldn't overlook one point: since August 2, 2026, the EU-wide AI labeling requirement under Article 50 of the AI Act has been in force. AI-generated or substantially manipulated images must be clearly identifiable as such at first perception. This mainly affects "deployers" – i.e., companies that use AI systems professionally, for example to create product photos, social posts or offer visuals.

The obligation doesn't apply where there is human editorial review with accountability – in practice, that means not every AI-assisted image is automatically subject to labeling, but the line should be carefully checked on a case-by-case basis. Ignoring the labeling requirement risks steep fines: up to €15 million or 3 percent of global annual revenue.

A second, often-overlooked point concerns copyright: purely AI-generated images are largely not protected by copyright in Germany. So if you have an exclusive campaign visual created by AI, you should know that this image generally doesn't enjoy the same legal protection as a traditionally photographed image – a point especially relevant for trades businesses and retailers wanting to advertise with unique visuals.

Both points – the labeling requirement and copyright – aren't arguments against using AI images. They're arguments for not treating their use as a casual "let's just try it in the chat" exercise, but as part of a process where labeling, sign-off and documentation are built in from the start.

In practice, this means for most mid-market businesses: clarify internally who is responsible for checking whether a given image is subject to the labeling requirement – ideally before the first campaign with AI images goes live, not after. If you regularly publish AI-generated product photos or offer visuals, build a standard labeling step into your workflow rather than reconsidering it for every single post. Precisely because the line between images that require labeling and those that don't can shift depending on the degree of human editorial control, it's worth settling this once with a clear internal policy rather than making case-by-case decisions under time pressure right before a campaign launch.

What this specifically means for your marketing budget in the mid-market

Let's sum up the points relevant to your budget. First: API pricing for GPT-Image-2.5 has stayed stable, while competition – Microsoft's MAI-Image-2.6-Flash chief among them – is adding further price pressure across the market. That suggests image generation is likely to get cheaper rather than more expensive for the foreseeable future, regardless of which specific provider you choose. Second: the real productivity gain from the new model generation lies less in the price per image than in how fast you can iterate – more variants, faster adjustments, shorter reaction times to current occasions.

Third: the choice between models like Flare and Sunburst – or more generally between fast, cheap models for mass variants and higher-quality, slower models for hero images – is a decision that should be made per use case. A social media post that's outdated in 48 hours doesn't need Sunburst-level quality. A cover image for your online shop that influences purchase decisions for months justifies the higher quality tier – and in some cases still a hybrid approach with a real reference photo as the starting point.

This is exactly where the difference lies between a company that experiments in the ChatGPT window and one that treats image generation as part of a workflow. An agent or workflow tool can automatically pick the right model for the use case, build the labeling requirement directly into the approval process, and feed the generated images into your shop, your CRM or your social media planning without a manual intermediate step. The value doesn't come from the model alone, but from clean integration into existing systems – compliance included.

If you're looking for an entry point into structured AI use, ChatGPT Business, or ChatGPT Work as it's been called since April 2026, is an obvious starting point: around $25 per user per month, roughly €29 to €34 including VAT in Germany, designed for teams of five users or more, with an admin console and exclusion of your own data from model training. For smaller businesses, this can be a sensible first step toward introducing image generation into the team in a structured way – before it grows into an automated workflow.

Conclusion: it's not the model that decides, but the process behind it

ChatGPT Images 2.5 is a solid, noticeable update – more natural images, more reliable reference fidelity, new tools like @Sketch, Poster and Merch templates, and targeted image comments. The circulating figure of "twice as fast and 72 percent more efficient," though, doesn't belong to this update – it belongs to Microsoft's separate model, MAI-Image-2.6-Flash. OpenAI's own claim – up to 50 percent lower latency than its predecessor – is less spectacular, but it's solid.

For your marketing budget, what matters in the end isn't the precise percentage claimed by a single provider, but the overall direction: AI-generated product photos, social posts and offer visuals are becoming faster, cheaper and more versatile across the industry. At the same time, regulatory requirements – labeling obligations, copyright questions – are growing alongside them. Whoever brings both together into a clean, automated workflow will actually extract a budget advantage from this model leap. Whoever only takes away the most spectacular number from the headline risks missing the part of the story that actually matters.