ChatGPT Images 2.5 Is Here (And It’s Actually Amazing)

From the creator

Link to our newsletter: https://bitbiased.ai/ OpenAI just replaced ChatGPT’s image engine with ChatGPT Images 2.5 — claiming a 50% reduction in generation time, while outside testers are reporting real-world speeds two to four times faster than the previous model. But speed isn’t the biggest upgrade. OpenAI quietly shipped two separate models under the Images 2.5 name — Flare and Sunburst — and the bigger change targets one of the most frustrating problems with AI image editing: asking for one small edit and watching everything else in the image subtly change with it. ChatGPT Images 2.5 went live on September 8, 2026, replacing ChatGPT Images 2.0 across free, Plus, Work, and Codex. Flare is designed for fast everyday generation and iteration, while Sunburst prioritizes detail, texture, and higher-quality multi-turn edits. Through the API, developers can access them directly as gpt-image-2.5-flare and gpt-image-2.5-sunburst. OpenAI also expanded image quality and resolution. New “xhigh” and “max” quality settings sit above low, medium, and high, with maximum output resolution reaching 3840×3840 and supported aspect ratios ranging from 1:3 to 3:1. The release also adds Sketch mode, templates for posters and product shots, targeted click-and-comment editing, prompt sharing, improved transparent backgrounds, and tighter provenance and deepfake safeguards. The real test, though, is editing consistency. GPT-Image-2 could generate impressive images, but dense prompts often lost instructions, objects landed in the wrong places, and repeated edits could gradually alter faces, lighting, poses, or unrelated parts of the scene. Images 2.5 is specifically designed to reduce that behavior. OpenAI says the new system should change the requested element while preserving the rest of the image, with multi-turn edits building on previous changes instead of progressively degrading the original. Reference-image handling is also supposed to be stronger, helping preserve recognizable faces, poses, expressions, and pets while changing styles or environments. OpenAI and early reviewers also point to sharper textures, more natural lighting, improved text rendering, better structured layouts and infographics, and proper alpha-channel transparency. There are still limitations. Dense text can misspell or blur, hands and fine anatomy remain unreliable, exact object placement isn’t guaranteed, and identity can still drift during long editing chains. There’s another important catch: inside the regular ChatGPT interface, users can’t explicitly select Flare or Sunburst the way API developers can. Then there’s prompting — because Images 2.5 rewards specificity far more than simply writing longer prompts. We break down a practical structure built around subject, action, environment, composition, lighting, style, and constraints. For editing, that final category becomes especially important. Instead of redescribing an entire image, tell the model exactly what should change and explicitly protect what shouldn’t: “Change only the coffee mug to blue ceramic. Keep the person, the lighting, and everything else exactly the same.” We also walk through prompts for product photography, infographics, reference images, sketches, transparent-background assets, and how API users can iterate cheaply with Flare before moving to Sunburst at high or max quality for the final result. So is ChatGPT Images 2.5 actually a major upgrade, or mostly a faster image model with better branding? The early evidence points to meaningful improvements — particularly where AI image generators have historically struggled most: controlled editing and maintaining consistency across multiple changes. The question is how well those improvements hold up once you push the model beyond clean demos and into real workflows. CHAPTERS 00:00 OpenAI Just Replaced ChatGPT’s Image Engine 01:29 What Actually Shipped 03:50 Where The Old Model Kept Breaking 05:15 What Actually Changed Under The Hood 07:54 Where It Still Slips 08:53 How To Actually Prompt It 12:32 Real Upgrade, Or Just A Rebrand With Better Marketing #openai #chatgpt #chatgptimages #aiimagegenerator #ai

Choose to Build with AI
Matched to ChatGPT Images

AI Maker Residence at KOKO

The third AI workshop taught by our legendary teacher, Nick Sarafa. In one of the last events we did an asset manager raised an additional £25M on their fund within a space of 9 months. This is a full-day hands-on training workshop for purposeful co-creation with AI using Claude Code. Imagine having access to hundreds of billions of dollars of computing power and knowing exactly how to make it work for you through the power of super intelligence. One person did.

◆ Fri 09 Oct 2026 ◆ KOKO Cafe, London ◆ With Nick Sarafa
AI Maker Residence at KOKO
Live event
AI Maker Residence at KOKO
Fri 09 Oct 2026
Find a job · 1 live role

Jobs in AI.
Apply now.

Apply once › get screened › meet the company

See all roles

More like this

Running one yourself?

List your AI event,
wherever it is.

A meetup, a workshop, a hackathon, a conference. Any city, or online. Tell us about it and it lands in front of people already learning this stuff.