GPT Image 2.6 is the more flexible upgrade for teams that need faster image creation, tighter editing control, or GIF and SVG workflows. Released on November 4, 2026, it retains the two-model lineup: Flare for responsive creation and Sunburst for detailed work. Its clearest additions are independent planning control and new output workflows; standard token rates remain the same as 2.5.
If you already use GPT Image 2.5, start with the corresponding 2.6 variant and your existing prompts. Upgrade when it improves the work you deliver: faster previews, fewer corrections, or an output format you previously had to build separately. Keep 2.5 in the comparison for text-heavy assets, where the Arena results below still favor its current entries.
Key Takeaways
Compare Flare against Flare and Sunburst against Sunburst before drawing conclusions about either generation.
OpenAI reports faster generation in 2.6, improved layouts in Flare, and stronger edit consistency in Sunburst at medium or high quality.
Reviewed for clarity, source attribution and current API terminology.
GPT Image 2.6 separates rendering quality from planning effort through quality and complexity.
Both generations have a published image-output rate of $30 per million billable tokens, but individual images can consume different amounts.
Complexity, GIF, and SVG are the headline additions. Markup-guided API editing is available to supported models in both generations.
GPT Image 2.6 vs GPT Image 2.5 at a glance
Both generations ship as two variants: Flare (optimized for speed) and Sunburst (optimized for detail and edit consistency). GPT Image 2.6 launched November 4, 2026.
Quick comparison table
Aspect
GPT Image 2.5
GPT Image 2.6
Winner / Notes
Speed
Baseline
~50% faster officially; tests often show 1.5–2× (or more) especially on Flare. Flare can hit ~6s at low quality.
2.6
Image quality
Clean, sharp, polished
Richer texture, more natural detail (especially skin, people, complex scenes)
2.6 (noticeable on portraits & dense scenes)
Text rendering
Highly accurate
Highly accurate + often better styled/integrated type
Tie / slight edge 2.6
Product fidelity
Excellent (labels, materials, counts)
Excellent (same accuracy level)
Tie
Reference images
Strong subject preservation
Stronger fidelity + fuller scenes
2.6
Multi-turn editing
Good, but more drift over steps
Better consistency (esp. Sunburst at medium/high); less unwanted change
2.6
Cost (token rates)
$30 / 1M image output tokens
Identical rates
Same list price
Actual cost
Lower tokens at low/medium
Similar tokens/cost at high+; 3–4× more tokens (higher cost) at low/medium
Switch to 2.6 for most new work, especially high-quality generation, portraits, complex scenes, reference-based work, and multi-step editing. You get better detail + consistency at the same (or lower) effective cost and significantly higher speed.
Keep 2.5 for high-volume low/medium-quality assets where token cost matters, or if you rely on Batch API (not supported on 2.6 in early reports).
Prompts that worked on 2.5 generally transfer directly. Use Flare for fast iteration and Sunburst for final/precise results.
Overall, 2.6 is the clear upgrade for quality-sensitive and iterative workflows, with the main trade-off being higher cost on the cheaper quality tiers.
What is GPT-Image-2.6
OpenAI released Images 2.6 (available as GPT-Image-2.6 Flare for speed and Sunburst for detail/precision) on November 4, 2026, as the successor to the 2.5 generation. Official claims center on ~50% faster generation overall (Flare as low as ~6 seconds at low quality), greater consistency across edits (especially Sunburst at medium/high quality), plus new capabilities like GIF output, SVG export, a complexity parameter, and improved markup-guided edits. Token pricing rates remain the same as 2.5 ($30 per 1M image output tokens), though real-world token usage (and thus cost) is similar at high quality and higher at low/medium on 2.6.
GPT Image 2.5 remains useful for established static-image workflows. It already supports generation, editing, reference images, transparency, and multiple quality levels. To isolate the generational change, compare 2.5 Flare with 2.6 Flare, or 2.5 Sunburst with 2.6 Sunburst.
API Identifiers and Output Dimensions
Both generations support image creation and editing. In API configuration, use the explicit variant identifier rather than the family name: gpt-image-2.5-flare, gpt-image-2.5-sunburst, gpt-image-2.6-flare, or gpt-image-2.6-sunburst. Match the variant to the workflow before comparing output quality.
Both generations accept dimensions in multiples of 16 and aspect ratios from 1:3 to 3:1. Start with a common size such as 1024×1024, 1536×1024, or 1024×1536. For larger custom outputs, the documented limits need a closer look:
Dimension rule
GPT Image 2.5
GPT Image 2.6
Width and height
Multiples of 16
Multiples of 16
Aspect ratio
1:3 to 3:1
1:3 to 3:1
Maximum edge
3840 pixels
Not specified in the cited 2.6 prompting guide
Total pixel range
655,360–8,294,400
Not specified in the cited 2.6 prompting guide
Experimental-size threshold
Above 2560×1440
The same threshold is not confirmed for 2.6 in the cited guide
Sources: image-generation guide, Images API reference, and 2.6 prompting guide, checked November 6, 2026. An unspecified ceiling does not mean unlimited size. Confirm larger 2.6 dimensions with the endpoint you use before making them a product requirement.
For a landscape asset, 1536×864 is a practical custom-size example in both guides. Under the stated 2.5 limits, 3840×3840 is too large: each edge fits, but the total pixel count does not. GIF frame dimensions and animation pixel budgets are checked separately.
GPT Image 2.6 Flare vs Sunburst
Flare is the smaller, speed-focused 2.6 model; Sunburst is the base model for more demanding detail and editing. Choose Flare when the user is waiting for the next draft. Choose Sunburst when a small error in a product, face, or label would send the image back for correction. Their publishedtoken rates are identical, so the choice turns on response time and usable output rather than a lower listed rate.
Decision
GPT Image 2.6 Flare
GPT Image 2.6 Sunburst
Primary role
Fast iteration and interactive generation
Detailed compositions and precise editing
Release improvements
Better layouts for posters and product visuals
Stronger edit consistency at medium or high quality
A useful first task
Explore several campaign directions or preview a background change
Refine packaging, a recurring character, or a complex final composition
What to compare
Time to the next usable draft
Preservation of the details you need to publish
GPT Image 2.6 vs GPT Image 2.5: Arena Results
In Arena’s Text Rendering category, the displayed November 2, 2026 snapshot places 2.5 ahead of the corresponding 2.6 entries. All four scores are preliminary. The 2.6 entries specify Medium quality; the 2.5 names do not specify an equivalent setting.
Model
Text Rendering rank / score
Votes
GPT Image 2.5 Sunburst
1 / 1462 ± 8
16,868
GPT Image 2.6 Sunburst (Medium)
2 / 1433 ± 23
871
GPT Image 2.5 Flare
3 / 1432 ± 8
16,539
GPT Image 2.6 Flare (Medium)
5 / 1403 ± 22
860
The displayed score gap is 29 points for each matching variant: 1462 versus 1433 for Sunburst, and 1432 versus 1403 for Flare. Those are leaderboard points, not percentage differences in text accuracy. The 2.6 entries have far fewer votes and explicitly specify Medium quality, so this snapshot is a reason to retain 2.5 in a text-heavy evaluation, not evidence that it wins every image task. The snapshot also predates the November 4 public launch.
GPT Image 2.6 vs GPT Image 2.5: Image Quality Comparison
Independent side-by-side tests (identical prompts, matched sizes/quality where possible) show 2.6 generally producing richer texture, more natural detail (especially in people and complex scenes), and better multi-step fidelity, while text accuracy stays excellent on both.
Text Rendering
Both versions render text with high accuracy—spelling, numbers, and multi-language strings (including non-Latin scripts in tested cases) are reliable on posters, menus, infographics, product labels, and dense layouts. 2.6 often produces more styled or refined typography and integrates text into richer surrounding scenes without sacrificing legibility.
In one poster test, both spelled every word correctly, but 2.6 packed in more environmental detail (castle, tiles, flowers) and used finer subtitle styling; 2.5 sometimes invented unrequested text elements. Infographics and exact-count prompts (e.g., specific numbers of objects) performed comparably, with 2.6 adding richer icons or more specific supporting text while staying on-brief. Dense festival-poster and bilingual newspaper tests similarly showed clean, correct text on both, with 2.6 often delivering fuller page layouts and better secondary details.
For product labels with fine print and figures (e.g., “2% Retinol,” “Net weight 250 g”), both versions read correctly at full size across quality settings. Bottom line: text is a strength of the family; 2.6 edges ahead mainly in stylistic integration and scene richness rather than raw accuracy.
Product Photography
Product accuracy (shape, color, materials, labels, fine details like button counts or mesh texture) holds up similarly well on both. Tests with serum bottles, coffee bags, running shoes, and cardigans found labels fully readable and products largely intact. High-quality generations on 2.6 were roughly twice as fast (e.g., ~11.5 s vs ~23.4 s averages in one set) at essentially the same cost.
Differences tend to be subtle: 2.6 sometimes avoided unrequested additions (e.g., water droplets on a shoe) and produced slightly cleaner results, while both could invent background elements (chalkboard text) if not constrained. Precise/local edits preserved product fidelity comparably (pixel-difference scores on the product region were close), though 2.6 precise edits were slower and used more tokens in the tested runs. At default/low/medium quality, 2.6 consumed substantially more output tokens (often 3–4×), raising cost for a modest quality lift.
For catalog-style or e-commerce work that prioritizes fidelity + speed at high quality, 2.6 is a safe and faster drop-in. Lower-quality bulk work may still favor 2.5 on cost.
Reference Image Editing
Both support multiple reference images (up to ~8–10 depending on the interface). 2.6 generally produces fuller, more complete scenes while better preserving identity, materials, and requested elements from the references.
In person + garment combination tests (face/hair details from one photo + specific sweater lettering/pattern from another into a new environment), both kept the subject recognizable and text correct, but 2.6 delivered stronger overall compositions and environmental integration. Product-in-scene and person-in-new-scene edits similarly showed 2.6 placing items more naturally (e.g., better lighting interaction) while maintaining the rest of the frame. Official positioning and early tests emphasize improved reference fidelity and subject preservation relative to prior generations.
Multi-turn Editing Consistency
This is one of the clearer areas of improvement claimed for 2.6 (particularly Sunburst at medium/high quality): better retention of unedited details across successive edits. Quantitative pixel-drift measurements on product labels across three sequential edits (wall color change → pouch color change → add croissant, with instructions to hold label, drawing, cup, counter, and lighting) showed 2.6 Sunburst drifting the least overall and handling the final addition most cleanly. 2.6 Flare also performed strongly; 2.5 variants and older models showed progressively more shift in composition, size, or details.
Qualitative multi-instruction and living-room-style edit tests similarly found 2.6 following more of the requested changes while leaving more of the original intact (though neither is perfect—unwanted alterations can still occur). Consistency is not absolute, and results vary by prompt complexity and quality setting, but 2.6 reduces cumulative drift compared with 2.5 in controlled multi-step workflows.
GPT Image 2.6 vs 2.5: Speed and Performance
Official Speed Claims and Test Conditions
OpenAI’s launch announcement reports 50% faster image generation and says 2.6 Flare can produce an image in as little as six seconds at low quality. This makes Flare a useful starting point for interfaces where people repeatedly request another variation.
The six-second figure is a best case. The time you experience depends on the prompt, size, and quality setting.
Performance measure
GPT Image 2.5
GPT Image 2.6
Interpretation
Official speed improvement
Its launch described reduced latency relative to the preceding generation
Its launch describes 50% faster generation
The claims use different generational baselines
Fastest stated generation time
Use your existing workflow’s measured response time as the baseline.
As little as 6 seconds for Flare at low quality
A best-case figure, not average latency
Editing Precision and Complexity
Independent complexity control is a new 2.6 capability. The documented 2.5 workflow offers quality controls without this separate planning setting. In 2.6, quality determines rendering detail, while complexity changes how much planning goes into arranging the scene.
Both 2.6 variants offer low, medium, and high complexity, defaulting to high when you omit the field. Use low for a focused change, such as replacing a background; consider medium or high when a scene has several objects and layout constraints. Keep quality fixed while comparing complexity so the effect is easier to see.
GIF and SVG expand 2.6 beyond the documented 2.5 static-image workflow. A GIF can turn a mascot into a looping reaction or show a short product movement. SVG is useful for graphics you want to scale or continue editing.
GIF output uses output_format=gif and supports 2–16 frames, with 9 as the default. Set the motion in the prompt, then review the transition from the final frame back to the first.
SVG output uses the generations endpoint with output_format=svg and n=1. The API creates a reference image and then produces static vector content. SVG output is not available on the image edit endpoint. For a logo or icon, inspect the resulting paths and lettering before handing it to a designer.
GPT Image 2.6 vs GPT Image 2.5: Cost
Official API Pricing
GPT Image 2.6 and 2.5 have the same standard token rates for both Flare and Sunburst. Text prompts and reference images are charged differently, so an edit with several input images has a different cost breakdown from a text-only request.
OpenAI standard rates per 1M tokens
GPT Image 2.5 Flare / Sunburst
GPT Image 2.6 Flare / Sunburst
Text input
$5.00
$5.00
Cached text input
$1.25
$1.25
Image input
$8.00
$8.00
Cached image input
$2.00
$2.00
Image output
$30.00
$30.00
Rates checked November 6, 2026. Cached rates apply only to eligible usage. In particular, OpenAI’s guide limits 2.5 cached-input pricing to the Responses API image tool, excluding direct Images API calls.
Cost per accepted image = Total workflow spending ÷ Number of accepted images.
A model can become more economical by reducing failed attempts or revisions, even when its token rate stays the same. Conversely, a faster model is not necessarily cheaper if it consumes more tokens or produces fewer acceptable results.
Where Can You Access GPT Image 2.6 and GPT Image 2.5?
Official Provider Access
OpenAI provides 2.6 Flare and 2.6 Sunburst for image generation and editing. Use the Image API’s generations endpoint for a new image and its edits endpoint when you supply an image to change.
To begin, create a CometAPI key and configure your client with base URL https://api.cometapi.com/v1. Select the model identifier and send a generation request with your prompt. The model page includes setup examples. CometAPI currently marks SVG support as undergoing adaptation and verification, so use a supported route for SVG until that integration is available.
Should You Choose GPT Image 2.6 or GPT Image 2.5?
Best Model by Use Case
For most new generation and iterative editing work, GPT Image 2.6 (especially Sunburst for finals/precision and Flare for drafts) is the stronger choice—richer natural detail, comparable or better text and product fidelity, improved reference handling, and meaningfully better multi-turn consistency, often at higher speed and similar high-quality cost. Keep 2.5 in mind for cost-sensitive low/medium-quality volume work or if an existing pipeline is tightly tuned to its behavior. Real results still depend on prompting, quality setting, and specific use case; side-by-side testing on your own images remains the best validation.
Workflow
First option to evaluate
Why it is worth testing
Main acceptance criterion
High-volume social graphics
2.6 Flare
Faster iteration may improve throughput
Usable output rate and response time
Product imagery and campaign assets
2.6 Sunburst
Precision matters when small changes affect brand accuracy
Product shape, labels, and materials
Established static-image service
Keep 2.5 as the baseline and test 2.6
Existing output behavior may already meet customer expectations
Regression results
Focused background or sign edits
2.6 with low complexity initially
Planning effort can be adjusted independently
Preservation of untouched regions
Short looping animations
2.6 GIF workflow
Animated output is explicitly documented
Frame consistency and loop continuity
Design assets requiring further editing
2.6 SVG workflow
The output can support additional design work
Structure and practical editability
Migration Guide: GPT Image 2.5 to GPT Image 2.6
Keep your existing integration available while introducing a separately configured 2.6 route. Migrate one workflow at a time so that request changes and output differences remain easy to diagnose.
Replace gpt-image-2.5-flare with gpt-image-2.6-flare, or the corresponding Sunburst identifier, only after confirming access through your provider.
Run your saved prompts and reference images through 2.6 with the same size, quality, and format. Start with the default complexity; include at least one exact-text image and one repeated-edit sequence.
Tune complexity after the baseline comparison: low for focused edits, higher levels for more involved layouts. Change quality separately, and remove legacy request fields your new endpoint does not accept.
Check the returned image, dimensions, transparency, and billing. Add GIF or SVG as a separate workflow; confirm provider support before enabling it for users.
Route a small portion of traffic to the validated configuration. Track acceptance, latency, charges, and errors, and retain a rollback path to the working 2.5 configuration.
Conclusion
GPT Image 2.6 is a worthwhile upgrade candidate when you need more responsive creation, stronger control over iterative edits, or GIF and SVG output. Start with Flare for quick iteration and Sunburst for assets whose details must survive repeated changes.
If 2.5 already delivers the static images you need, keep it until 2.6 demonstrates a useful improvement on those same tasks. With identical standard token rates, the deciding factors are the quality of the finished asset, the time needed to reach it, and the total cost of the attempts along the way.
