OpenAI just shipped ChatGPT Images 2.5, a flagship image model that generates up to 50% faster than its predecessor and ships with a sketch-to-image tool, days after Google and Midjourney pushed their own upgrades.
OpenAI announced ChatGPT Images 2.5 on September 8, 2026. The company says the new model cuts generation time by as much as 50% compared with Images 2.0, while producing sharper detail, richer texture, and more natural lighting. It's a real product ship, not a teaser: the model is already rolling out across every ChatGPT tier, from Free up through Work and Codex, on web, desktop, and mobile.
The bigger fix might be under the hood. Images 2.5 is built to hold onto a subject across multiple rounds of editing. Old versions drifted. Ask it to change a jacket's color, then adjust the background, then swap the pose, and earlier versions would often quietly reshape a face or drop a logo along the way. OpenAI says 2.5 follows multi-turn instructions more reliably and keeps the subject consistent from edit to edit. That matters most to anyone using the tool for product shots or brand assets, not one-off art.
Four new features ship alongside it. Sketch lets you draw a rough shape and pair it with a prompt. Poster and Merch templates format an image for a specific output. Image comments let you annotate a generation and ask for a targeted fix instead of rewriting the whole prompt. Shared prompts let a team reuse a setup that already worked instead of reverse-engineering it from a finished image.
Two models, split by price #
Developers don't get one model. They get two. GPT-Image-2.5 Flare is the fast, high-throughput default, positioned by OpenAI as the right call for most everyday API generation. GPT-Image-2.5 Sunburst trades speed for precision, aimed at edits where consistency can't slip: product photography, packaging, anything that has to match a reference exactly across multiple passes. Both accept text and image inputs, and both plug into the Images API or the image-generation tool inside the Responses API.
OpenAI Chief Scientist Jakub Pachocki Warns No AI Lab Is Ready to Scale Safely OpenAI chief scientist Jakub Pachocki published an essay saying no AI lab, including OpenAI, has solved alignment well enough to keep scaling at full speed, just three days after OpenAI shipped GPT-6 Astra, its first model to cross a Critical cybersecurity threshold. Sam Altman called the essay 'an important post' even as Astra keeps rolling out... - how to safely scale autonomous AI systems - AI lab safety concerns for frontier models
That's a pricing lever as much as a technical one. OpenAI's prior model, gpt-image-2, billed at $8 per million image input tokens and $30 per million image output tokens, which worked out to roughly $0.005 for a low-quality image and up to $0.165 for a high-quality square render. Splitting Flare from Sunburst gives developers a cheap default and a deliberate, pricier upgrade path, rather than one model priced for the hardest case every time.
Google and Midjourney aren't standing still #
None of this happens in a vacuum. Google's Nano Banana Pro, built on Gemini, already scores 9.50 against Midjourney v7's 8.62 on CuriousRefuge's 29-prompt benchmark, and it renders text inside images with 94% accuracy against Midjourney's 10%, a gap that matters to anyone generating signage, labels, or infographics. Nano Banana Pro is priced at $0.134 per image at 1K to 2K resolution and $0.24 at 4K, and it ships with an official REST API, something Midjourney still doesn't offer. Midjourney users are stuck on paid subscriptions and unofficial third-party wrappers.
Midjourney still wins on aesthetic taste and color, according to multiple side-by-side comparisons of v7 against Nano Banana Pro. It's just slower to move. A four-image batch takes 30 to 60 seconds through Discord or the web app, versus roughly four minutes for Nano Banana Pro to produce 50 images. Speed and API access are exactly where OpenAI is choosing to compete.
Here's the real test. A 50% latency cut is a number OpenAI controls the framing of, and it hasn't yet been independently benchmarked against Nano Banana Pro or Midjourney v7 head to head. What is checkable right now is availability: Images 2.5 is live today, across every ChatGPT tier, with an API split into a cheap default and a precision option. Google ships fast too. Midjourney doesn't ship an API at all.
That leaves the real fight where it's been for a year, not who has the sharpest sample image on launch day, but who developers actually build production pipelines on. OpenAI just made its case with two models, a lower price floor, and a tool that remembers what your last edit looked like.
Also read: Businesses Don't Have an AI Problem, They Have a Context Problem • Ledger Hires New Security Chief as AI-Driven Crypto Hacks Hit $1.4 Billion • Instacart Launches Clementine, an AI Assistant That Builds Your Cart for You
Oracle Plans More Layoffs in September to Pay for Its AI Spending Spree Oracle managers have reportedly been told to submit employee lists for a new round of layoffs, with some teams facing double digit percentage cuts, timed before the company's fiscal second quarter begins. The cuts follow a fiscal 2026 that saw Oracle shed 21,000 jobs even as it spent $55.7 billion on AI infrastructure and posted negative $23.7... - oracle layoffs september 2024 ai spending costs - how oracle funding artificial intelligence technology initiatives
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.