An in-depth review of ChatGPT Images 2.5, officially introduced by OpenAI on September 8, 2026. Discover 50% reduced generation latency vs Images 2.0, interactive @Sketch tools, native transparent PNGs, and official token-based USD pricing.
3-Line Summary
- OpenAI officially unveiled ChatGPT Images 2.5 on September 8, 2026, delivering major performance and UX upgrades following the previous Images 2.0 generation.
- It slashes generation latency by up to 50%, introducing interactive @Sketch canvas composition, image comment pins, native transparent PNGs, and Pixel Lock.
- The API operates under standard token-based pricing ($30.00 per 1M output tokens), offering two specialized endpoints:
gpt-image-2.5-flare(high speed) andgpt-image-2.5-sunburst(high precision).
Anyone who creates AI visuals regularly has struggled with two recurring obstacles: fine-grained localized edits and lengthy generation wait times.
While GPT Image 2 (Images 2.0) resolved typography warping and anatomical distortion through reasoning passes, deep pre-generation calculations added noticeable seconds to every request. Furthermore, describing intricate compositions through words alone often resulted in misaligned layouts.
OpenAI's ChatGPT Images 2.5, announced on September 8, 2026, tackles workflow ergonomics and responsiveness head-on. Offered in two API tiers—the agile Flare and high-fidelity Sunburst—let us explore how this generation transforms production pipelines.
What Has Changed Compared to Images 2.0? 3 Core Upgrades
[Images 2.0 vs ChatGPT Images 2.5 Comparison]
1. Generation Latency & Responsiveness
- Images 2.0 ➔ Deep visual reasoning adds deliberate wait times (~5~7s)
- Images 2.5 ➔ Up to 50% faster latency via optimized pipeline (Flare engine)
2. Composition Guidance & Surgical Editing
- Images 2.0 ➔ Pure text prompt descriptions (spatial ambiguity)
- Images 2.5 ➔ Draw rough guides with @Sketch or pin precise visual comment pins
3. Asset Output Formats
- Images 2.0 ➔ Standard opaque scene backgrounds
- Images 2.5 ➔ Native transparent 32-bit PNGs and Pixel Lock continuity
1. Up to 50% Faster Generation Latency
The most immediate operational upgrade is speed. Modernized rendering optimizations reduce inference latency by up to 50%, enabling high-frequency creative iteration without bottlenecking workflows.
2. Interactive @Sketch Canvas and Visual Comment Pins
Instead of writing convoluted paragraphs to specify object locations, users can trigger @Sketch in ChatGPT to draw quick contours and shapes. Additionally, dropping visual comment pins directly on image coordinates enables surgical, multi-turn feedback just like collaborating with a human designer.
3. Native Transparent Backgrounds (PNG) and Pixel Lock
App icons, web buttons, and merchandise stickers can now be output directly as 32-bit PNG files with alpha channel transparency, eliminating third-party background cutout tools. Furthermore, Pixel Lock freezes facial structures and scenery while swapping only targeted items like clothing or props.
Official API Pricing: Token-Based Architecture
With Images 2.5, OpenAI has retired legacy per-image flat pricing (DALL-E 3 was completely removed from the API on May 12, 2026), replacing it with standardized token-based metering.
1. Official Token Rates (USD)
| Billing Category | Standard Price (per 1M Tokens) | Cached Prompt Price (per 1M Tokens) |
|---|---|---|
| Text Input | $5.00 | $1.25 |
| Image Input (Reference Photos) | $8.00 | $2.00 |
| Image Output | $30.00 | N/A |
- Token Consumption: Total output tokens are governed by your selected resolution (1K, 2K, 4K) and quality parameter (
low,medium,high,xhigh,max). - Batch API: Asynchronous Batch API jobs (24h turnaround) receive an automatic 50% discount on all token tiers.
2. Estimated Real Cost per Image by Resolution & Quality Tier
Looking strictly at the raw token rate ($30.00 / 1M output tokens), it can be difficult to gauge the actual out-of-pocket cost for generating a single image.
Images 2.5 meters tokens based on your chosen quality tier (low, medium, high, xhigh, max). Here is how that translates to realistic per-image costs:
| Quality Tier | Target Resolution & Use Case | Average Output Tokens per Image | Estimated Standard Cost per Image | Cost per Image via Batch API (50% Off) |
|---|---|---|---|---|
low | ~512px (Quick drafts, small icons) | ~150 - 200 tokens | ~$0.005 | ~$0.0025 |
medium | ~1024px 1K (Web banners, social thumbnails) | ~800 - 1,000 tokens | ~$0.025 - $0.030 | ~$0.0125 - $0.015 |
high | ~2048px 2K (High-res displays, marketing assets) | ~2,000 - 2,500 tokens | ~$0.060 - $0.075 | ~$0.030 - $0.037 |
max / xhigh | Up to 3840px 4K (Print mockups, studio hero assets) | ~4,000 - 5,000 tokens | ~$0.120 - $0.150 | ~$0.060 - $0.075 |
- Granular Budget Control: Unlike legacy models that locked users into rigid $0.040 or $0.080 fees, creators can now generate rapid draft concepts for just $0.005 each—saving up to 85% on exploratory iterations—while reserving higher token budgets only for finalized 4K master files.
3. Flare vs. Sunburst Endpoints and Model IDs
Calling gpt-image-2.5 directly returns an invalid model error. Developers must specify either -flare or -sunburst. Both models share identical token pricing, but feature distinct execution backends:
gpt-image-2.5-flare: Reduces inference latency by 50%, optimized for rapid ideation and high-throughput pipelines.gpt-image-2.5-sunburst: Delivers maximum photographic fidelity, refined surface specularities, and consistent multi-turn inpainting control.
Practical Selection Criteria
- When to choose ChatGPT Images 2.5:
- Storyboards, illustrated books, or serial campaigns requiring character consistency across frames (Pixel Lock).
- UI assets, merchandise stickers, and product catalogs demanding clean transparent backgrounds.
- Marketing teams producing hundreds of daily variations on tight turnaround times (
Flareengine). - Compositions where drawing rough lines is faster than writing long prompts (
@Sketch).
- When alternative models make more sense:
- Real-time video frame streaming and voice interaction are better suited to GPT-6 Astra or Google Project Astra.
- High-resolution typography-only signage can be generated economically on GPT Image 2.
Frequently Asked Questions (FAQ)
Q1. How do I use the @Sketch feature?
Type @Sketch at the start of your prompt in ChatGPT to open the interactive canvas tool. Sketch rough outlines or shapes with your mouse or stylus, add brief descriptive text, and the model synthesizes the image adhering to your spatial blueprint.
Q2. Which model should I choose: Flare or Sunburst?
For standard web graphics, blog thumbnails, and social posts, Flare is twice as fast and highly efficient. Choose Sunburst only when producing commercial print assets or intricate photorealistic textures.
Q3. Is there an extra charge for generating transparent PNGs?
No. Transparent background generation is a native format option available at standard pricing with zero surcharges.