GPT Image 2.6 vs 2.5: SVG, GIF, Editing and API Prices

What changes between GPT Image 2.5 and 2.6?

This compares documentation and prices, not a paid head-to-head benchmark. It does not establish a quality win, measured speedup or exclusive capability merely because 2.5 documentation is silent. Compare Flare with Flare and Sunburst with Sunburst; OpenAI positions Flare for interactive iteration and Sunburst for detailed visual work.

Complexity: planning and rendering are separate controls

complexity controls planning before rendering; quality controls rendering quality. Lower complexity can suit a focused edit; higher complexity can help dense text, layouts and object relationships, but may increase latency or unwanted changes. OpenAI lists no separate complexity surcharge; actual billed usage still determines total cost.

The official paper-cut football example holds quality=medium while changing complexity. Observe the trophy, uniforms and CHAMPIONS sign. It illustrates settings within 2.6, not a 2.6-versus-2.5 benchmark.

SVG and animated GIF: different deliverables

SVG: output_format: "svg" on the generation endpoint, with n=1, produces static vector paths, shapes and colors. The official Field & Flour bakery-logo example compares Flare and Sunburst. Scripts, external resources and embedded rasters are unsupported; SVG output is not available on the edit endpoint. Billing includes a reference image plus authoring-model tokens, so ordinary PNG estimates do not apply.

GIF: output_format: "gif" is generation-only. frame_count accepts 2–64, default 9; frame_duration_ms accepts 20–2000 milliseconds, rounded to 10-millisecond increments, default 200. The official guide shows a lantern animation. Per-frame dimensions and total pixel limits still apply. GIF does not support editing, streaming, partial images, custom compression or prompt_rewriter; this is not a video API or an audio feature.

Markup editing is also supported by 2.5

OpenAI explicitly supports markup-guided edits on supported 2.5 and 2.6 models. Its kitchen example uses a separate transparent PNG with a red circle around a vase: recolor it cobalt blue, preserve the shape, stems and kitchen, and remove the circle. Test changes outside the target area rather than calling this a 2.6-only upgrade. Official support does not guarantee that every third-party route accepts the same parameters.

API prices: equal token rates do not mean equal image costs

All four official model pages list these Standard rates in USD per million tokens:

Billed item	2.5 Flare / Sunburst	2.6 Flare / Sunburst
Text input	$5	$5
Cached text input	$1.25	$1.25
Image input	$8	$8
Cached image input	$2	$2
Image output	$30	$30

Equal token rates do not guarantee equal token use or per-image cost. Editing adds input-image usage; SVG has additional authoring costs. Use actual usage and the complete bill for a specified model, size and quality.

APIMaster's checked 2.5 minima are $0.0375 / $0.06 / $0.12 per image at 1K / 2K / 4K. For 100 images: $3.75 / $6 / $12, excluding retries and other charges. The 1K minimum and 2K/4K minima come from different routes; not every route supports every size or advanced field. These are not 2.6 quotes.

Compare cost per accepted image = total spend across all paid attempts / accepted images. A 20% cheaper attempt can still cost more if it needs twice as many retries.

Three tasks to test before switching

These are proposed acceptance tests, not completed model experiments:

1. Poster typography: request three coffee drinks, the exact text “Happy Coffee”, “Second cup half price” and “Nov 5–Nov 12”, logo top left and price bottom right. Check spelling, dates, counts and placement.
2. Targeted edit: change only a circled white bottle cap to dark blue, preserving label text, shape, light and background; remove the circle. Use the same original and markup layer for both versions.
3. New formats: request an editable SVG icon and a looping GIF of a pelican riding a bicycle. Inspect paths, wheel/pedal motion and loop continuity. Test the new workflow without treating unconfirmed 2.5 format support as a quality failure.

FAQ

Is GPT Image 2.6 only a rumor?

No. At the November 5 check, official model pages and the image guide included both 2.6 IDs. Account access and third-party availability still need separate checks.

Is 2.6 always better than 2.5?

This documentation comparison does not establish that. Measure spelling, composition, preserved details and paid retries on the same workload.

Can I just replace the model ID?

First confirm account access, endpoint, SDK and route support for complexity, SVG, GIF and markup. Keep a working 2.5 fallback.

Can I buy 2.6 on APIMaster now?

Neither exact 2.6 ID had a public route at this check. There is no current APIMaster 2.6 price or discount to promise.

Sources and API access

For static images today, compare 2.5 Flare / 2.5 Sunburst: 85% below the checked 1K market reference. register → create an API key. Use https://apimaster.ai/v1 with the exact supported model ID. Check the 2.6 model card before purchasing. See the separate GPT Image 2.6 buying guide.

• OpenAI: Image generation
• OpenAI: gpt-image-2.6-flare
• OpenAI: gpt-image-2.6-sunburst
• OpenAI: gpt-image-2.5-flare
• OpenAI: gpt-image-2.5-sunburst