What Is GPT Image 2.5? Up to 50% Lower Latency

Up to 50% Lower Latency: A First Look at GPT Image 2.5

GPT Image 2.5 is OpenAI’s image-generation and image-editing system launched on September 8, 2026, available in ChatGPT as ChatGPT Images 2.5 and in the API as GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. It focuses on sharper detail, more natural lighting, stronger reference-subject preservation, and edits that remain consistent across multiple turns. OpenAI says generation latency can be up to 50% lower than Images 2.0, while official pricing and technical specifications remain limited.

Updated 2026-09-09: GPT Image 2.5 is now live on OpenRouter and Higgsfield (see the Update below).

Key Takeaways

• GPT Image 2.5 launched on September 8, 2026, after early reports had treated the name as unconfirmed.
• ChatGPT Images 2.5 is rolling out to all ChatGPT, ChatGPT Work, and Codex users on desktop, mobile, and web.
• The API has two variants: Flare, the default model for most applications, and Sunburst, the slower option for more precise creative work.
• The main upgrade is Multi-Turn Consistency: subjects, textures, layouts, and approved details are designed to survive repeated edits more reliably.
• New ChatGPT tools include Sketch, templates, image comments, and prompt sharing.
• Arena reporting placed Sunburst first and Flare second across its Text-to-Image, Image Edit, and Multi-Image Edit arenas, although the available reports do not provide complete scores and methodology.
• Exact official pricing, context limits, model architecture, weights, and license terms remain unconfirmed in the available material.

What Is GPT Image 2.5?

GPT Image 2.5 is OpenAI’s next image model generation after GPT Image 2. It handles image creation from prompts and image editing from references, with particular emphasis on preserving details while a user revises an image repeatedly.

OpenAI presents the release under two related names. ChatGPT uses ChatGPT Images 2.5 as the product experience. Developers receive two API models: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. The API names describe variants of the same release family rather than separate unrelated products.

The release followed a short period of uncertainty. Earlier reporting described possible anonymous image-model checkpoints and a likely ChatGPT upgrade, but the final name, variants, pricing, and release timing were not confirmed at that stage. The earlier signal analysis is documented in GPT-Image 2.5 Release Signals: Signal vs Noise. The September 8 announcement converted the model name and availability into confirmed release facts.

The practical problem GPT Image 2.5 addresses is edit drift. In many image workflows, a small change to one object can unintentionally alter a face, background, product label, lighting setup, or composition. OpenAI positions this release around targeted changes that leave more of the accepted image intact.

GPT Image 2.5 is therefore best understood as an image creation and revision system, not only a text-to-image generator.

GPT Image 2.5 at a Glance

Specification	Current information
Developer	OpenAI
Type	Image generation and image editing model family
Modality	Text-to-image and image-guided editing
Context window	Not yet confirmed
Pricing	Reported to match GPT Image 2 token rates; exact official usage and per-image costs not confirmed
Availability	ChatGPT, ChatGPT Work, Codex, API, and multiple third-party platforms
API variants	GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst
Release date	September 8, 2026
License	Not yet confirmed
Open weights	Not announced

How GPT Image 2.5 Works and What Makes It Different

OpenAI has not published enough architecture information to describe GPT Image 2.5 in terms of parameter count, training method, or model internals. The observable difference is its workflow behavior: generate, inspect, comment, and revise while retaining more of the previous image.

1. Multi-Turn Consistency

Multi-Turn Consistency is the central capability claim. GPT Image 2.5 is designed to keep subjects, textures, backgrounds, and other approved regions stable across repeated edits. This matters for product images, campaign assets, character variations, and any task where a first generation is only the beginning.

OpenAI describes the model as better at preserving subjects from reference photos and following editing instructions across multiple turns. A community test compared nine consecutive edits of the same robot image and reported strong preservation of the face, texture, and background, although that result is an individual test rather than a standardized benchmark. The nine-edit community comparison is useful evidence, but it should not be treated as a universal performance guarantee.

2. Flare and Sunburst

Flare is the default API variant. OpenAI recommends it for most applications that need a balance of quality, editing ability, and speed.

Sunburst is intended for detailed creative workflows that need additional precision and control across edits. Its trade-off is longer generation time. The available announcement does not provide model size or resolution limits, and benchmark reporting does not yet provide a complete numerical comparison between the two variants.

Flare is the default choice for most applications, while Sunburst is the precision-oriented choice for detailed creative work with longer generation times.

3. Faster Generation

OpenAI says GPT Image 2.5 reduces image-generation latency by up to 50% compared with Images 2.0. That figure describes a maximum claimed improvement, not a guaranteed result for every prompt, resolution, device, or API request.

Community reporting has also described the two models as producing higher-quality images in less than half the time, but the available reports do not provide a baseline, hardware details, latency figures, or test methodology. That finding is directional only.

OpenAI says GPT Image 2.5 can reduce generation latency by up to 50% versus Images 2.0.

4. Text and Layout Handling

Launch examples emphasize cleaner generated text, sharper details, complex layouts, and transparent backgrounds. Community testers particularly praised text rendering, though the evidence consists of demonstrations and individual judgments rather than a complete public benchmark.

Source: @LuminaBench

GPT Image 2.5 is also described as better at natural lighting, richer textures, recognizable subjects, and preserving likeness from reference photos. These improvements are relevant to posters, packaging, social graphics, product imagery, and branded compositions.

The model still requires inspection for exact spelling, small labels, logos, alignment, and complex visual constraints. A strong-looking image is not proof that every requested word or object is correct.

5. ChatGPT Tools

ChatGPT Images 2.5 adds several workflow features:

• Sketch lets users draw directly in ChatGPT as a visual reference. The documented invocation is @ Sketch.
• Templates provide starting points for formats such as posters and merchandise.
• Image comments allow users to point at a target region and request a localized change.
• Prompt sharing lets others reuse or remix the prompt associated with an image.

These tools matter because they reduce the distance between an idea and an actionable image instruction. A rough drawing can communicate placement more effectively than a long text description, while comments can identify a specific region without restating the entire scene.

What You Can Do With GPT Image 2.5

GPT Image 2.5 is suited to workflows where an image needs several controlled revisions rather than a single generation.

Product and campaign imagery: Create product scenes, packaging concepts, advertising layouts, and polished campaign variations. Sunburst is the more relevant option when small visual changes must preserve the rest of a composition.

Creator and social content: Flare is positioned for faster production of social graphics, thumbnails, promotional images, and visual experiments.

Reference-based image editing: Upload or provide a reference and ask for changes while preserving the main subject. This is useful for clothing changes, background swaps, lighting adjustments, and controlled composition edits.

Visual prototyping: Use Sketch to turn a rough interface, poster, product shape, or scene layout into a more complete image.

High-volume generation: Flare’s speed positioning makes it the more natural candidate for repeated generation, rapid exploration, and automated visual workflows. Actual throughput limits and rate limits are not yet confirmed.

Sequential creative work: The available demonstrations suggest that GPT Image 2.5 can support storyboarding and visual sequences through more consistent shots and better composition. WaveSpeedAI has also shown workflows involving 16 sequential frames from one generation, but these demonstrations do not establish a native video capability.

The model is therefore useful for iterative visual work, but production teams should test their own edit chains and target formats before relying on it for consistent output at scale.

How GPT Image 2.5 Compares

The reporting includes GPT Image 2 as the direct baseline. GPT Image 2 introduced 2K output, multiple aspect ratios, web research for current information, non-Latin language support, and Instant and Thinking intelligence levels according to launch coverage. GPT Image 2.5 shifts attention toward speed and editing stability.

Arena reporting placed GPT-Image-2.5 Sunburst at number one and Flare at number two across Text-to-Image, Image Edit, and Multi-Image Edit arenas. The reported gains over GPT Image 2 were +40 and +18 points in Text-to-Image for Sunburst and Flare, and +59 and +30 points in Single-Image Edit. The available reports do not include complete raw scores, sample sizes, or methodology.

Capability	GPT Image 2.5	GPT Image 2
Positioning	Faster generation and more consistent editing	Earlier OpenAI image-generation baseline
Reported latency	Up to 50% lower than Images 2.0, according to OpenAI	Baseline for that comparison
Editing	More precise localized edits across multiple turns	Supports image editing, with less confirmed detail about repeated-edit preservation
Reference subjects	Better preservation claimed for reference photos	Not yet confirmed in the reporting so far
API variants	Flare and Sunburst	GPT Image 2
Arena placement	Sunburst first and Flare second in reported Text-to-Image, Image Edit, and Multi-Image Edit arenas	Baseline used for comparison
Public standardized benchmark	Arena rankings and point gains are available, but complete methodology and raw scores are not	Not yet confirmed in the reporting so far

The strongest current case for GPT Image 2.5 is improved iterative editing, supported by OpenAI’s release claims and early Arena results. The full numerical quality advantage over GPT Image 2 remains difficult to assess because the available benchmark reporting is incomplete.

Availability: How to Access GPT Image 2.5

GPT Image 2.5 is available through the official ChatGPT product rollout and through OpenAI’s API. The official announcement says ChatGPT Images 2.5 is rolling out to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web. OpenAI Developers also direct builders to start with Flare for most applications and choose Sunburst for extra precision.

Access route	Current status	Best use
ChatGPT	Rolling out to all ChatGPT users	Prompting, editing, Sketch, templates, and comments
ChatGPT Work	Rolling out to all ChatGPT Work users	Team and work-oriented image workflows
Codex	Rolling out to all Codex users	Programmatic or development-assisted creative workflows
OpenAI API	Flare and Sunburst live	Application integration and automated generation
Replicate	GPT Image 2.5 live	Hosted image generation and editing
OpenRouter	GPT Image 2.5 live in two OpenAI models	Provider-routed API access
Higgsfield	GPT Image 2.5 live	Product imagery and advertising workflows
WaveSpeedAI	GPT Image 2.5 live	Storyboarding and sequential visual workflows
kie.ai API service	Offers GPT Image 2.5 via API	A hosted API route for developers
Open weights	Not announced	Not available as a confirmed self-hosting route

For the hosted API route, kie.ai offers GPT Image 2.5 via API. Developers can review the available image-model access page for GPT Image 2 and verify the current GPT Image 2.5 mapping, request format, and availability before integrating it.

The model is also live through third-party services including Replicate, OpenRouter, Higgsfield, and WaveSpeedAI. The official API documentation names Flare as the normal starting point and Sunburst as the precision-oriented alternative.

What We Don’t Know Yet

Several important engineering details remain unpublished or insufficiently verified:

• The official per-image or per-token price for GPT Image 2.5 is not stated in the available OpenAI release material. Independent launch reporting lists the same token rates as GPT Image 2—$8 per million image input tokens, $2 per million cached image input tokens, $30 per million image output tokens, and $5 per million text input tokens—but exact usage and per-image costs remain unconfirmed.
• Context-window limits, input-image limits, output-resolution limits, and supported aspect ratios are not confirmed here.
• OpenAI has not published parameter counts, architecture details, or downloadable weights.
• Arena reporting places Sunburst first and Flare second across three image arenas and gives point gains over GPT Image 2, but it does not provide complete raw scores, sample sizes, or methodology.
• The claimed 50% latency reduction lacks a disclosed test setup in the reports reviewed.
• Multi-turn consistency is promising but not settled. One nine-edit test was positive, while other community reports described degradation or visual artifacts after repeated generation.
• Pixel-art performance remains a reported weakness. One hands-on test found that GPT Image 2.5 could not produce grid-aligned pixel art as reliably as a specialist model.
• Noise and artifact behavior is not fully resolved. Community testing has reported both improved consistency and remaining noise or deformation issues in some use cases.

These gaps do not cancel the release. They define the information that engineering teams should validate before committing to a production workflow.

Frequently Asked Questions

What is GPT Image 2.5?

GPT Image 2.5 is OpenAI’s image-generation and image-editing system launched on September 8, 2026. It is available in ChatGPT as ChatGPT Images 2.5 and through the API as GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Its main focus is sharper output and more stable editing across multiple turns.

Is GPT Image 2.5 released?

Yes. GPT Image 2.5 was released on September 8, 2026. The release includes ChatGPT Images 2.5 and two API variants, Flare and Sunburst.

What is the difference between GPT Image 2.5 Flare and Sunburst?

GPT-Image-2.5 Flare is the default API model for most applications. GPT-Image-2.5 Sunburst is designed for more precise control in detailed creative work and takes longer to generate.

How much does GPT Image 2.5 cost?

The official release material does not state exact pricing. Independent launch reporting says the API token rates match GPT Image 2: $8 per million image input tokens ($2 cached), $30 per million image output tokens, and $5 per million text input tokens; exact usage and per-image costs remain unconfirmed.

Is GPT Image 2.5 available in ChatGPT?

Yes. ChatGPT Images 2.5 is rolling out to all ChatGPT, ChatGPT Work, and Codex users on desktop, mobile, and web. The rollout includes tools such as Sketch, templates, and image comments.

GPT Image 2.5 vs GPT Image 2: what changed?

GPT Image 2.5 emphasizes sharper detail, faster generation, better preservation of reference subjects, and more consistent multi-turn editing. OpenAI says latency can be up to 50% lower than Images 2.0, although the available evidence does not provide a complete standardized comparative benchmark.

What to Watch Next

The next useful signals are official pricing and rate limits, full raw results and methodology for the Arena comparisons, and independent tests of repeated edits, pixel art, text accuracy, and complex layouts. Those updates will determine whether GPT Image 2.5’s editing claims hold across production workloads rather than only demonstrations.

Update — 2026-09-09

Availability has expanded beyond OpenAI’s own products. OpenRouter reports that GPT Image 2.5 is live in two OpenAI models, while Higgsfield testing highlights sharper product detailing and improved text for advertising workflows.

OpenAI has also published a ChatGPT Images 2.5 System Card. Separately, an independent user report says the two API variants cost the same as GPT Image 2 and that repeated-edit consistency improved in practice; the report also notes that image metadata still identifies outputs as 2.0, so metadata cannot currently verify the model version.

Building similar image-generation and editing workflows? On kie.ai you can try GPT Image 2.5, Nano Banana 2.1, and Seedream 5.0 Pro.