GPT Image 2.6 vs Gemini, FLUX and Seedream Compared

OpenAI's GPT Image 2.6 shipped on November 4, 2026, and the question a marketing or product team asks the same day is whether to move to it. The honest answer depends on three things that are rarely laid out together: what each model costs in the units its vendor actually bills, where independent raters place it, and what it can do to an image after the first generation. This post does that for seven models as the vendor pages and two public leaderboards stood on November 5, 2026.

Two rules shape everything below. Prices are not converted between units. OpenAI and Google bill image output per million tokens, and only Google publishes what a 1K image costs in dollars, so only Google's figure appears as a per-image price. The others publish per-image or per-credit rates and are quoted as such. Second, a leaderboard position is a dated observation. LMArena's boards were marked as of November 2, 2026 when read, and both GPT Image 2.6 entries are flagged preliminary on a few thousand votes. Artificial Analysis did not list GPT Image 2.6 at all on November 5.

For what the new OpenAI models do and cost on their own terms, read our GPT Image 2.6 Flare and Sunburst guide. This post is about the field around them.

The seven are not alike in business model, and that explains most of the pricing differences before any quality judgement. OpenAI and Google are frontier labs that bill images as tokens inside a larger API. Black Forest Labs and ByteDance sell per-image inference, BFL with an open-weight line beside it. Midjourney is a subscription product. Ideogram and Recraft are design-tool companies with APIs priced in credits or in units of a tenth of a cent.

Release dates matter because the boards reward votes accumulated over time. GPT Image 2.6 is the newest, released November 4, 2026, per OpenAI's changelog. Google's changelog dates Nano Banana 2.1 to October 6, 2026. Midjourney's version page makes V8.2 the default from July 24, 2026. For FLUX 3 Image, Seedream 5.0 and Ideogram 4.5, the vendor pages we read state no release date; the months in the table are Artificial Analysis listings or, for Seedream, inferred from model ID suffixes, and are marked as such.

One example of why the vendor type matters: a team that generates 2,000 product variants a month pays Seedream 5.0 flash a flat $0.018 each and can forecast the bill to the cent. The same team on GPT Image 2.6 pays $30 per million image output tokens and cannot forecast until it has measured tokens per image at its own size and quality settings. Neither is wrong, but only one is predictable before the first run. The failure case is comparing a flat per-image rate with a token rate as if they were the same column.

Two independent boards are in public use. LMArena, now hosted at arena.ai, ranks models from pairwise human votes and publishes a score with an interval. Its text-to-image and single-image-edit boards were both labelled as of November 2, 2026 when we read them on November 5. Artificial Analysis runs its own Elo-style arena with FLUX.2 [dev] fixed at 1000 as the anchor; its page shows no as-of date.

On LMArena's text-to-image board, OpenAI holds the top five. GPT Image 2.5 Sunburst leads at 1427±5 on 54,913 votes. GPT-Image-2.6 Sunburst (Medium) is second at 1409±13 on 2,967 votes, then 2.5 Flare at 1399±5, 2.6 Flare (Medium) at 1384±13, and GPT Image 2 (Medium) at 1382±4 on 106,737 votes. Microsoft's MAI-Image-2.6 is sixth at 1331±5, Gemini Nano Banana 2.1 seventh at 1325±9, and Grok Imagine Image 2.0 eighth at 1317±7. The intervals on the two 2.6 entries are more than twice as wide as their neighbours, which is what a few thousand votes looks like; the gap between 2.6 Sunburst and 2.5 Sunburst is close to the width of those intervals.

On the single-image-edit board the order shifts. 2.5 Sunburst leads at 1517±5, 2.5 Flare is second at 1482±5, 2.6 Sunburst (Medium) third at 1476±10, MAI-Image-2.7-Preview fourth at 1469±7, and 2.6 Flare (Medium) fifth at 1462±10. Nano Banana 2.1 is eighth at 1427±6. Artificial Analysis tells a consistent story one generation back: 2.5 Sunburst (max) leads text-to-image at 1225 and editing at 1200, Nano Banana 2.1 is fourth on text-to-image at 1175, FLUX 3 Image is 13th at 1118 and Seedream 5.0 Pro 16th at 1101. GPT Image 2.6 is absent from both of its boards. The failure case is quoting the LMArena rank without the preliminary flag and the Medium quality label, both of which the board prints next to the score.

OpenAI's pricing page lists both GPT Image 2.6 models at $8 per million image input tokens, $2 cached, and $30 per million image output tokens, with text input at $5 and $1.25 cached. No per-image figure is published. Google's pricing page lists Nano Banana 2.1 at $1.50 per million input tokens and the same $30 per million for image output, $15 in batch, and then does what OpenAI does not: it states that a 1K image consumes 1,120 tokens and is equivalent to $0.0336, a 2K image 1,680 tokens and $0.0504, and a 4K image 3,780 tokens and $0.113. Those are Google's conversions, not ours.

The per-image vendors are easier to read and harder to verify in one case. ByteDance's ModelArk pricing page lists Seedream 5.0 pro at $0.045 per image up to 2.61 megapixels and $0.09 above, 5.0 flash at $0.018, 5.0 lite at $0.035 and Seedream 4.5 at $0.04, with the first input image free and $0.003 for each further input image on pro. Recraft's API pricing page lists V4.1 and V4.1 Utility at $0.035, V4.1 Flash at $0.007, and V4 and V3 at $0.04, with V3 edits at $0.04 raster or $0.08 vector. Ideogram's credit page lists 4.5 at 3, 6, 10 and 22 credits for Low, Medium, High and High with source images, and pay-as-you-go credits at one cent, so High works out near $0.10 if API credits are the same unit.

Black Forest Labs is the unverified cell. Its pricing documentation defines one credit as $0.01 and tiers FLUX 3 Image by resolution from 768-square to 4K, but the tier price cells did not render in the text we could fetch. The calculator on BFL's pricing page showed $0.048 per image for its default selection, which agrees with the $48 per thousand that Artificial Analysis lists. We print it as the default-tier rate pending a browser check, and that is how the table labels it. Midjourney's plan prices were not captured; the plans page we tried returned a 404. Our image API pricing comparison carries the wider market in the same unit discipline.

Generation quality gets the leaderboard attention. Editing is where production time goes, and the vendors have taken different routes. OpenAI's edit endpoint takes masks and, since November 4, a drawn markup layer that points at the region to change, with the first image as the subject and further images as references. Google documents Nano Banana 2.1 for conversational, multi-turn editing with character consistency across turns, taking text, image, video and PDF inputs.

ByteDance's Seedream 5.0 pro and flash accept up to ten reference images and can decompose a result into as many as sixteen PNG layers with alpha, a feature no other vendor here documents on the pages we read, and the one a design team working in layers will notice first. Black Forest Labs documents instruction edits with up to ten references and bounding boxes for FLUX 3 Image, with web-search grounding on by default. Midjourney's V8.2 added an Edit Model inside the product. Ideogram markets 4.5 around what it calls Precise Edit. Recraft's V3 line offers inpainting, outpainting and background replacement at $0.04 raster or $0.08 vector.

An example of why this matters more than rank: a packaging team with one approved render and twelve regional label variants does not need the best text-to-image score. It needs an edit path that holds the product and changes only the label. OpenAI's markup, Seedream's layers and Nano Banana's multi-turn consistency are three different answers to that brief. The failure case is choosing the text-to-image leader and discovering its edit workflow does not fit the team's tools.

OpenAI's guide allows custom dimensions for GPT Image 2.6 with edges in multiples of 16, aspect ratios from 1:3 to 3:1, no edge above 3,840 pixels and a total between 655,360 and 8,294,400 pixels, with anything above 2560 by 1440 marked experimental. Output is PNG, WebP with transparency, JPEG, and now SVG and animated GIF from the generations endpoint. Google states 1K, 2K and 4K output for Nano Banana 2.1 with 1K as the default; the page we read did not state file formats.

Seedream 5.0 pro and flash run 1K, 1.5K and 2K size levels with 2K as default and a ceiling of 4,624,220 pixels; 5.0 lite reaches 4,096 by 4,096. Output is PNG or JPEG, JPEG by default, with a transparent-background option. Black Forest Labs tiers FLUX 3 Image up to a 4K tier of roughly 16 megapixels and bills anything over about 4 megapixels at that tier. Midjourney's V8.1 and later render 2K without upscaling and allow aspect ratios to 14:1, or 4:1 in HD. Recraft returns WebP, PNG or editable SVG, which with OpenAI makes two vendors in this set that publish vector output.

The example that decides this section is a logo brief. Two models here return SVG: GPT Image 2.6 at a two-model token cost, and Recraft at a flat per-image rate with a separate vector price on its edit line. Everything else returns a raster that someone has to trace. The failure case is buying 4K because it is the largest number on the sheet when the deliverable is a 400-pixel ad unit that the per-image vendors render for a fraction of the price.

Four things in this post are weaker than they look in a table, and each is stated where it appears. First, the LMArena entries for GPT Image 2.6 are at Medium quality, preliminary, and dated two days before OpenAI's public release. We do not know why the board carries a pre-release date and do not guess. Second, Artificial Analysis lists its own prices per thousand images; we use that column only as a cross-check on BFL's calculator figure, never as a vendor price, because the vendor units differ.

Third, two models in this discussion are not purchasable in the ordinary way: MAI-Image-2.7-Preview is marked as having no API available on Artificial Analysis, and Midjourney has no public API on its own documentation. Fourth, the gaps are real: BFL's per-tier prices, Midjourney's plan prices, Ideogram's API-specific rate and release date, Recraft's resolution ceiling and Google's output formats were all missing from the pages we could read.

The routing below follows the documentation, not the leaderboard. A team already on OpenAI's API gains markup edits, GIFs and SVG by moving to 2.6 with no change to the rate card, and stays with the vendor holding LMArena's top five text-to-image places. A team that needs a predictable per-image bill at volume has ByteDance's flash tier at $0.018 and Recraft's Flash at $0.007 as published floors. A team inside Google's stack gets the only token-billed model whose vendor publishes the per-image equivalent, which makes its forecast as easy as the per-image vendors'.

Our Nano Banana 2.1 versus FLUX 3 Image comparison goes deeper on the two mid-table models for product content, and our FLUX 3 Klein note covers the open-weight route for teams that want to run edits locally. Teams deciding where an image pipeline should run, and how to measure it once it does, can take the question to our paid media practice.