Nano Banana 2.1 vs Pro vs GPT Image 2.5/2.6: 9-Prompt Test, Cost
Same 9 prompts, one draw each at 1K: Pro, GPT Image 2.5 and 2.6 matched the requested wording; Nano Banana 2.1 slipped on text but costs a quarter of Pro.
On this page

For exact wording inside an image, run Nano Banana Pro, GPT Image 2.5 or GPT Image 2.6 first; pick 2.6 when a mockup or infographic must not gain data or slogans you didn't write. For cheap, fast drafts, infographics, product shots and simple edits, run Nano Banana 2.1. That split comes from a same-prompt test on October 8, 2026: eight text-to-image prompts plus one shared edit, one draw per model, 1K output, no rerolls. GPT Image 2.6 Sunburst ran the same nine prompts with the same settings on November 20, 2026. Pro and both GPT versions reproduced the requested wording and prices on a poster, a Japanese shop sign and a Chinese menu, apart from one Chinese character that all four models drew in its Japanese form. Nano Banana 2.1 got the words right but restyled a subline, added a stray word to the sign and wrote a price as 三十八元 instead of 38元. On the infographic, the counting prompt, the product photo, the portrait and the recolor edit, all four were usable.
Price and wait time point the other way. At Google's list price a 1K image costs $0.0336 on Nano Banana 2.1 and $0.134 on Pro, so 2.1 is about a quarter of the price before thinking tokens. GPT Image 2.6 costs the same as 2.5 at quality high, about $0.054 per 1024×1024 image at OpenAI's token price, but at low and medium quality it uses three to four times as many image tokens as 2.5, so cheap drafts are where 2.6 costs more. In the tests, 2.1 and Pro returned images in 17–33 seconds, GPT Image 2.5 took 58–76 seconds, and 2.6 took 34–45 seconds on its own run with three calls at a time.
Two limits apply to everything below. Each model drew each prompt once, so the results show what a model did on one try, not how often it gets it right. And both GPT versions (the Sunburst variant) ran through LaoZhang API's VIP per-call routes, gpt-image-2.5-sunburst-vip and gpt-image-2.6-sunburst-vip at $0.03 per call, not through OpenAI's API directly. Their output and timing may differ from an OpenAI-direct call.
Disclosure: laozhang.ai is our own service. Its prices are platform pricing, not an independent comparison.
Which to run first: Pro for exact text, 2.1 for cost, GPT for polish
| Job | Run first | What the test showed | Watch for |
|---|---|---|---|
| Posters, signs, menus where every character must match | Nano Banana Pro or GPT Image 2.6 Sunburst | Both matched the requested wording and prices on all three text prompts; Pro in 18–23 s, 2.6 in 34–45 s | $0.134 per 1K image for Pro at Google's price; check CJK character forms yourself |
| UI mockups and layouts with fixed copy | GPT Image 2.6 Sunburst | All weather-app labels and numbers right, no invented readings; no tagline on the infographic | It still adds scenery, such as a skyline photo in the app or a lantern beside the menu |
| Ad visuals and key art where the look matters most | GPT Image 2.5 or 2.6 Sunburst | Requested wording and prices; 2.5 drew the strongest poster layout | About a minute per image for 2.5 on the VIP route; 2.5 added a tagline and extra background signs nobody asked for |
| Infographics and explainers | Nano Banana 2.1 | Five correct steps in the right order with clear labels | It invented extra fields in a weather-app mockup; Google lists blurry small text at 1K as a known limit |
| Product shots and portraits | Nano Banana 2.1 on price, GPT Image 2.5 for a styled look | All four realistic | GPT Image 2.5's grading is more dramatic, which may not suit a plain catalog; 2.6 went plainer |
| Recolor one object, keep the scene | Any of the four | All kept the composition, earbuds, stone and light | 2.1 and Pro also turned the dark inner lid white; both GPT versions kept it dark |
| High-volume drafts | Nano Banana 2.1 | 17–33 s per image, 9 of 9 calls succeeded first try | On the Gemini API, thinking tokens are billed on top of the image; GPT Image 2.6 at low or medium quality costs three to four times as much as 2.5 |
If you need the right words in one draw and can pay more per image, Pro is the safest first choice from this test: it matched the requested wording and returned in about 20 seconds. GPT Image 2.6 Sunburst matched it on wording, invented no data fields or tagline, added fewer unrequested elements than 2.5 and costs well under half of Pro's list price at quality high, so it is the better pick when the copy must stay exactly as written and a 40-second wait is fine. If you plan to proofread and reroll anyway, 2.1's lower price buys you two to three attempts for the cost of one Pro image.
How the 9-prompt test was run (and what one draw can't tell you)
All 27 calls went through LaoZhang API on October 8, 2026:
Disclosure: laozhang.ai is our own service. Its prices are platform pricing, not an independent comparison.
- Nano Banana 2.1 (
gemini-nano-banana-2.1) and Nano Banana Pro (gemini-3-pro-image) through the Gemini-nativegenerateContentendpoint, aspect ratio 1:1, image size 1K, default thinking. - GPT Image 2.5 Sunburst through
/v1/images/generationsand/v1/images/editswith the modelgpt-image-2.5-sunburst-vip, size 1024×1024, qualityhigh.
The eight prompts covered an English poster, a five-step U.S. civics infographic, an e-commerce product photo, a Japanese ramen sign, a Chinese tea menu, a counting and left/right test, a portrait, and a weather-app mockup with exact labels. The edit asked all three models to turn the same black earbud case white (the photo 2.1 produced for the product prompt) and keep everything else unchanged.
Why the VIP route for GPT: on October 8, 2026, LaoZhang API's official-forwarding models (gpt-image-2.5-sunburst, gpt-image-2.5-flare) required a key in the Sora2Official or GPTImage2 Sora2 Enterprise group, and a key in the default group got a 503 "no available channels" error, which was not charged. Those groups are now named OpenAI Official and OpenAI Official Enterprise, and since November 6, 2026, the official Flare and Sunburst names called from the default group run on the VIP route at $0.03 per image. The VIP per-call models work from the default group, according to LaoZhang's GPT Image 2.5 route documentation. GPT Image 2.5 Flare was not tested; for how it differs from Sunburst, see GPT Image 2.5 Sunburst vs Flare.
GPT Image 2.6 was added on November 20, 2026, after its November 4 release. The same eight prompts, copied verbatim from the October script, and the same edit on the same input photo went to gpt-image-2.6-sunburst-vip on LaoZhang API, at 1024×1024 and quality high, one draw each with no rerolls. All nine calls returned on the first try. They ran three at a time instead of one after another, so their times describe that run, not a speed difference between the models. Each call reported 1,756 output image tokens, the same count as 2.5 at quality high, and the nine calls cost $0.27 at $0.03 each. GPT Image 2.6 Flare was not tested.
For the poster, the Japanese sign, the Chinese menu and the weather mockup, the results below compare the rendered text with the prompt character by character. The other results describe the images side by side.
What the test does not cover: speed or output when calling Google or OpenAI directly, 2K and 4K quality, GPT quality levels other than high, GPT Image 2.6's GIF and SVG output, Batch, multi-reference character consistency, search grounding, and repeatability. For how Nano Banana 2.1 holds a character across reference images, see Nano Banana 2.1 Character Consistency: Outfits Hold, Faces Drift. Nine prompts with one draw each is a spot check, not a benchmark. A model that slipped once here may get the same prompt right on the next draw, and the reverse is just as possible.
Which renders exact text best? Pro and GPT Image 2.5, with one shared miss
On the three prompts where specific wording was the point, Pro, GPT Image 2.5 and GPT Image 2.6 reproduced the requested words and prices; Nano Banana 2.1 made a different slip on each. The Nano Banana 2.1, Pro and GPT Image 2.5 results are from October 8, 2026; the GPT Image 2.6 results are from the November 20 run with the same prompts.
| Prompt | Requested text | Nano Banana 2.1 | Nano Banana Pro | GPT Image 2.5 (VIP) | GPT Image 2.6 (VIP) |
|---|---|---|---|---|---|
| Concert poster | MIDNIGHT TRAIN / Live at Hall 7 - Nov 14, 2026 - Doors 19:30 | Words right, but the subline set in all caps with em dashes | Exact, hyphens kept | Exact; train-track motif, strongest layout | Exact, hyphens kept; flat Swiss layout, nothing added |
| Ramen sign, Tokyo street | 本日のおすすめ ラーメン 980円, plus OPEN | Requested words present, plus a stray meaningless word (roughly エソープ) | Exact sign text and OPEN | Exact sign text, vertical OPEN, extra background signs (麺, やきとり) | Exact sign text and an OPEN plate; added ラーメン shop curtains and part of a road marking |
| Tea menu, brush calligraphy | 今日茶单 / 龙井 38元 / 普洱 42元 / 茉莉花茶 28元 | 38元 written out as 三十八元; title uses 単 | Prices exact; title uses 単 | Prices exact; title uses 単 | Prices exact; title uses 単; added a 茶 lantern and a 茶馆 seal |
| Weather app mockup | Lisbon 23°C Sunny; Mon–Fri with 23, 21, 19, 22, 24 | All labels and numbers right; added air quality, humidity and wind speed | All right; added feels-like, UV, high/low temperatures, hourly, air quality; the Fri card spills past the screen edge | All right, few extras | All right; added °C units, a status bar and a Lisbon skyline photo, but no extra readings |

The menu exposed a shared weakness. The prompt asked for 今日茶单 in simplified Chinese, and all three models drew the last character as 単, the Japanese form of the same character, in October; GPT Image 2.6 made the same substitution in November, so the newer GPT version does not fix it. A Chinese reader will notice it at a glance. If your text is in simplified Chinese, check variant forms in every output, whichever model you use.
The weather mockup shows a different kind of error: all four got the requested labels and numbers right, but Pro and 2.1 filled the screen with fields that were never requested, in small text. In a UI mockup, invented readings such as a UV index or an air-quality value are content you would have to remove or verify. GPT Image 2.5 stayed close to what was asked, and 2.6 added only decoration, units and a status bar, with no readings that weren't in the prompt.
Google's own model card lists blurry small text at 1K as a known Nano Banana 2.1 limit, though the slips here were changes to the wording rather than blur. 2.1 also thought hardest on these prompts, reporting 1,209–1,704 thinking tokens on the poster, sign, menu and infographic, against 122–252 for Pro across all nine prompts. The extra thinking did not make its text more literal in these draws.
Nano Banana 2.1 vs Pro: a quarter of the price, looser with wording
Nano Banana 2.1 is the better default and Pro is the better choice for literal text. That is the short answer to "which is better", and it rests on three kinds of evidence that point in slightly different directions.
Price. At Google's list price, 2.1 costs $0.0336 per 1K image against Pro's $0.134, about a quarter. At 4K it is $0.113 against $0.24, a little under half. Google's pricing page lists a Batch rate for 2.1, $15 per 1M image tokens or half the standard rate; as of November 20, 2026, Pro's section no longer shows a Batch rate. On LaoZhang API's per-call pricing, 2.1 is $0.045 at any size and Pro is $0.09.
Disclosure: laozhang.ai is our own service. Its prices are platform pricing, not an independent comparison.
Google's own evaluations. Google's model card scores 2.1 (with thinking) above Pro on overall preference (1050 vs 935), infographic factuality (0.521 vs 0.265), general editing (1026 vs 939), multi-character consistency (1106 vs 1011) and mask editing (1049 vs 927). These are Google's internal tests, not an independent check.
This test. Both drew an accurate five-step "How a bill becomes law in the U.S." infographic. 2.1 followed the familiar sequence (introduced, committee, floor vote, other chamber, president), while Pro merged the House and Senate votes into one step and added a conference-committee step. Both handled counting (three red cups left, two blue right) correctly. Pro was the one that kept the exact wording on the poster, sign and menu.
Capabilities also differ, which can settle the choice before quality does:
| Nano Banana 2.1 | Nano Banana Pro | GPT Image 2.5 | |
|---|---|---|---|
| Output sizes | 1K, 2K, 4K (no 0.5K) | 1K, 2K, 4K | Custom sizes up to 3840×2160 or 2880×2880 |
| Aspect ratios | 14, including 1:8 and 8:1 | Not listed here | Custom, within 1:3 to 3:1 on LaoZhang's routes |
| Reference images | Up to 14 (10 objects, 4 characters) | 6 objects, 5 characters | Up to 16 on official forwarding (LaoZhang doc) |
| Search grounding | Google Web and Image Search | Google Search | Not listed |
Pick Pro when the job is a print-ready sign, label or menu that you cannot proofread and regenerate cheaply. Pick 2.1 for everything else, and budget for a proofread and an occasional reroll. For Pro's full cost picture, including plan prices, see How Much Is Nano Banana Pro?.
Nano Banana 2.1 vs GPT Image 2.5: literal text and layout vs speed
GPT Image 2.5 (often searched as "GPT 2.5") was the more literal and the more polished of the two in this test; Nano Banana 2.1 returned images two to three times faster.
GPT Image 2.5 reproduced the requested wording and prices on every text prompt, and its weather mockup carried fewer invented extras than either Gemini model's. Its poster had the strongest composition of all nine outputs: a red sun over converging train tracks under the exact title and date line. Its product shot looked the most like a premium ad, and its portrait had the most dramatic color grading.
It also added things nobody asked for. The infographic gained a subtitle and the tagline "A LAW FOR A STRONGER TOMORROW" at the bottom, and the Tokyo street gained extra shop signs. Those are plausible additions, but in a deliverable with fixed copy they are something to remove.
On time, 2.1 returned in 17–33 seconds per image and GPT Image 2.5 in 58–76 seconds. The GPT timings come from the VIP route through LaoZhang's gateway, so they do not tell you how fast OpenAI's own endpoint is.
Arena's blind text-to-image votes, updated October 6, 2026, ranked GPT Image 2.5 Sunburst first (1425 ± 7, preliminary) and Nano Banana 2.1 fifth (1328 ± 9, preliminary). Artificial Analysis's preference leaderboard as of November 20, 2026 shows the same order, with 2.5 Sunburst second and 2.1 eighth. That matches the polish gap seen here, but these boards compare single images by preference and say nothing about cost, speed or edits.
Pick GPT Image 2.5 for final creative where layout quality matters and a minute per image is fine. Pick 2.1 when you need many images quickly or want Google Search grounding. If you are moving from GPT Image 2, GPT Image 2.5 vs GPT Image 2 covers what changed.
GPT Image 2.6 vs Nano Banana 2.1: copy that stays as written vs lower cost
Between these two, GPT Image 2.6 Sunburst is the one to run when the wording in an image has to match your copy exactly and no invented data or slogans may appear, and Nano Banana 2.1 is the one to run for volume, quick drafts and anything you will proofread anyway. GPT Image 2.6 Sunburst ran the nine prompts on November 20, 2026, and Nano Banana 2.1 ran them on October 8, both at 1K with one draw per prompt.
On text, 2.6 did what 2.1 didn't. It kept the poster's subline exactly as typed with plain hyphens, wrote the ramen sign without a stray word, and wrote 38元 as digits on the tea menu, where 2.1 restyled the subline, added a meaningless word and spelled the price out as 三十八元. Both drew the menu title with the Japanese 単. On the weather-app mockup, 2.1 added air quality, humidity and wind speed nobody asked for, while 2.6 added none. 2.6's five-step "How a bill becomes law" infographic was accurate, merged the House and Senate votes into one step as Pro did, included the veto override, and, unlike 2.5, carried no tagline. Where 2.6 still goes beyond the prompt is scenery: shop curtains on the Tokyo street, a lantern and a seal beside the tea menu, a skyline photo in the weather app.
On the recolor edit, 2.6 turned the case shell white and left the inner lid dark, as 2.5 did, where 2.1 turned the lid white too. Outside the case, 2.6 drifted slightly further from the source photo than 2.1 did and a little darker, though not by enough to see (the measurements are in the edit section below). That darkening is consistent with a three-round edit test on another page, where 2.6 darkened untouched areas a little more each round; see GPT Image 2.6 vs 2.5: Same Price at High, 4× the Cost at Low for that test and the rest of what changed between the two GPT versions.
Cost depends on the quality setting. At quality high, a 1024×1024 image on 2.6 costs about $0.054 at OpenAI's token price, against $0.043–$0.046 for a text-heavy 1K image on 2.1 including thinking tokens, so 2.1 is a little cheaper. At low quality 2.6 costs about $0.026, which undercuts 2.1's $0.0336 image price, but low quality was not part of this test, so you would be trading an untested quality level for the saving. Per call on LaoZhang API the order flips: the 2.6 VIP route is $0.03 and 2.1 is $0.045. The full per-quality and per-size table is in GPT Image 2.6 Pricing: Sunburst & Flare, $0.02–$0.21 per Image.
Disclosure: laozhang.ai is our own service. Its prices are platform pricing, not an independent comparison.
Features can decide it before quality does. OpenAI released GPT Image 2.6 on November 4, 2026 with GIF output, SVG output and a complexity parameter, according to its Introducing Images 2.6 announcement. On LaoZhang's VIP route, 2.6 can return GIFs but not SVG, and quality stops at high; xhigh and max return a 400 error that is not charged. Nano Banana 2.1 has Google Web and Image Search grounding, up to 14 reference images and 4K output. On Artificial Analysis's preference leaderboard as of November 20, 2026, GPT Image 2.6 Sunburst ranks first and Nano Banana 2.1 eighth, though the GPT score was measured at quality max, a level above the VIP route's ceiling. To call 2.6 from code, see GPT Image 2.6 API: Model IDs, Python and curl Code, Error Fixes.
Product photos, portraits and a recolor edit: all three usable
The non-text prompts separated the models far less than the text prompts did, and GPT Image 2.6 Sunburst, added on November 20, 2026, was usable on all of them too.
- Product photo (open matte black earbud case on pale stone, no text, no logo): all four produced clean, realistic studio shots with no stray text. GPT Image 2.5 used a larger, more premium-looking case with a soft-focus stone block behind it; 2.6 went for a plainer top-down catalog look.
- Portrait (elderly fisherman mending a net at golden hour): all four were natural and realistic. GPT Image 2.5 pushed the warm grading and backlight furthest; 2.6 used warm golden-hour light with boats in the harbor behind.
- Counting and position (exactly three red cups left, two blue right, nothing else): all four were correct.
For the edit, each model received the same black earbud case photo with the instruction to make the case glossy white and keep the earbuds, angle, stone surface, light and shadow unchanged. All four recolored the case and preserved the scene. The one difference was the inside of the lid, which was dark in the source: Nano Banana 2.1 and Pro turned it white along with the shell, while GPT Image 2.5 and 2.6 left it dark. Which result is "right" depends on whether you meant the whole case or only its outer shell, so spell that out in an edit prompt.
How well each model left the rest of the photo alone can be measured. Comparing each edited image with the source in grayscale, outside a 620×620 box around the case, the average difference was 1.6 levels out of 255 for Pro, 1.7 for Nano Banana 2.1, 2.5 for GPT Image 2.5 and 3.5 for GPT Image 2.6, and no model changed more than 0.02% of those pixels by more than 40 levels. None of that is visible. The only pattern is that 2.6 shifted the background slightly darker (by 1.4 levels on average), which matters only if you chain several edits on the same image.
The GPT Image 2.5 VIP route does not support the mask parameter, according to LaoZhang's GPT Image 2.5 documentation. If your edit workflow depends on a mask, use the official-forwarding route or one of the Gemini models. Google scores 2.1 above Pro on mask editing but also lists partial instruction following in mask edits as a known 2.1 limit.
Which is cheapest? 2.1 at list price, GPT Image 2.5 per call
The answer depends on how you pay. At the providers' list prices, Nano Banana 2.1 is cheapest per 1K image. On LaoZhang API's per-call pricing, the GPT Image 2.5 and 2.6 VIP routes are cheapest.
Disclosure: laozhang.ai is our own service. Its prices are platform pricing, not an independent comparison.
| Model | Google or OpenAI list price | LaoZhang API, per call |
|---|---|---|
| Nano Banana 2.1 | $0.0336 (1K), $0.0504 (2K), $0.113 (4K) per image, plus input and $7.50 per 1M thinking/text tokens | $0.045, any size |
| Nano Banana Pro | $0.134 (1K/2K), $0.24 (4K) per image, plus input and $12 per 1M thinking/text tokens | $0.09 |
| GPT Image 2.5 Sunburst/Flare | $30 per 1M image output tokens; OpenAI's calculator estimates $0.0059 (low) to $0.2107 (max) per 1024×1024 image | VIP routes $0.03; official forwarding billed at OpenAI's token rates |
| GPT Image 2.6 Sunburst/Flare | Same $30 per 1M image output tokens; about $0.0247 (low), $0.0435 (medium), $0.0527 (high) and $0.2107 (max) per 1024×1024 image | VIP routes $0.03, quality up to high |
The Google prices are as of November 20, 2026 and unchanged since October 8; the GPT Image 2.6 per-image figures are OpenAI's token counts as of November 5, 2026. Sources: the Gemini API pricing page, the OpenAI API pricing page, LaoZhang's GPT Image 2.5 documentation and LaoZhang's GPT Image 2.6 route documentation. Google lists a Batch rate at half price for Nano Banana 2.1's image output; Pro's section no longer shows a Batch rate as of November 20, 2026.
The per-image figure is not the whole bill on the Gemini API, because thinking tokens are charged on top. Using the token counts the test calls reported for a 1K image:
- Nano Banana 2.1, text-heavy prompt: $0.0336 + 1,209 to 1,704 thinking tokens × $7.50 per 1M ($0.0091 to $0.0128) ≈ $0.043–$0.046.
- Nano Banana Pro: $0.134 + 122 to 252 thinking tokens × $12 per 1M ($0.0015 to $0.0030) ≈ $0.136–$0.137.
- GPT Image 2.5 at quality high, 1024×1024: each call reported 1,756 image output tokens × $30 per 1M ≈ $0.053, plus about $0.001 for roughly 200 text input tokens at $5 per 1M ≈ $0.054.
- GPT Image 2.6 at quality high, 1024×1024: every call in the November 20 run also reported 1,756 image output tokens, so the cost is the same ≈ $0.054.
So at list price, a 1K image costs roughly $0.04 on 2.1, $0.05 on GPT Image 2.5 or 2.6 at high quality, and $0.14 on Pro. Input-token charges for these short prompts are a fraction of a cent and are left out.
The two GPT versions part ways below quality high. At low, 2.6 uses 824 image tokens per 1024×1024 image against 196 for 2.5, so an image costs about $0.026 instead of $0.007 once the roughly $0.001 of text input is added. At medium it is 1,449 tokens against 439, about $0.044 instead of $0.014. If you generate cheap drafts at low or medium, staying on 2.5 for that work costs a third to a quarter as much; GPT Image 2.6 vs 2.5 covers what else differs.
Per-call pricing changes the order. LaoZhang's flat $0.045 for 2.1 is slightly above Google's 1K cost with thinking but well below its $0.113 at 4K. Pro at $0.09 is under its $0.134 list price at every size. The GPT VIP routes at $0.03 undercut OpenAI's token rate at quality high, with the caveats above: they are not OpenAI-direct, and the 2.5 VIP route has no mask. Running the October test (27 successful calls) cost $1.485 at LaoZhang's list prices, 9 × ($0.045 + $0.09 + $0.03), and the November GPT Image 2.6 run added 9 × $0.03 = $0.27.

More detail on each price: Nano Banana 2.1 API Pricing and How Much Is Nano Banana Pro?.
Speed and failures: about 20 seconds on Gemini, about a minute on GPT VIP
Wall time per image, measured from the test machine through LaoZhang's gateway:
| Model | Time per image | First-try success |
|---|---|---|
| Nano Banana 2.1 | 17–33 s | 9 of 9 |
| Nano Banana Pro | 18–23 s | 9 of 9 |
| GPT Image 2.5 Sunburst (VIP route) | 58–76 s | 9 of 11 attempts; 2 returned 503 "Upstream model service error" and succeeded on retry |
| GPT Image 2.6 Sunburst (VIP route), November 20, three calls at a time | 34–45 s | 9 of 9 |
Nano Banana 2.1's slowest call (33 seconds) was the Chinese menu, one of the prompts where it spent the most thinking tokens. Pro stayed in a tight band. If you call a GPT VIP route from code, add a retry on 503 and a timeout well above 76 seconds; LaoZhang's GPT Image 2.6 documentation gives 25–55 seconds as the usual range for its VIP routes.
These are single samples that include gateway and network time, so treat them as a rough ranking rather than a latency guarantee. The GPT Image 2.6 row in particular comes from a different day and from parallel calls, so it shows that the 2.6 route answered in well under a minute on November 20, not that 2.6 is faster than 2.5.
Arena and Google's own scores: what they add to a 9-prompt test
The current ranking of all four comes from Artificial Analysis: as of November 20, 2026, its text-to-image leaderboard puts GPT Image 2.6 Sunburst first, GPT Image 2.5 Sunburst second, Nano Banana 2.1 eighth and Nano Banana Pro eighteenth. Larger sources like this sit alongside the nine-prompt spot check, each measured by someone else:
- Artificial Analysis text-to-image leaderboard (AA-Image-T2I v2.1, Elo from pairwise human preference votes, as of November 20, 2026): GPT Image 2.6 Sunburst #1 at 1261 ± 14, GPT Image 2.5 Sunburst #2 at 1230 ± 10, GPT Image 2.6 Flare #3 at 1227 ± 12, Nano Banana 2.1 #8 at 1177 ± 12 from 6,820 samples, Nano Banana Pro #18 at 1119 ± 10. The GPT entries were measured at quality
max, not thehighused in this test. It ranks preference only and does not check text accuracy, cost or edits. - Arena text-to-image (blind preference votes, as of October 6, 2026, before GPT Image 2.6 existed): GPT Image 2.5 Sunburst #1 at 1425 ± 7 and Flare #2 at 1398 ± 7, Nano Banana 2.1 #5 at 1328 ± 9 from 5,312 votes, Nano Banana Pro 2K #16 at 1248 ± 3. The top three scores were marked preliminary at the time. Arena ranks how much voters like one image over another; it does not check text accuracy, cost or edits.
- Google's model card for Nano Banana 2.1: 2.1 ahead of Pro on every reported axis, from overall preference to infographic factuality and mask editing, with known limits of blurry small text at 1K, partial instruction following in mask edits and left/right confusion. Google ran these evaluations on its own models.
All three agree with the broad picture from the nine prompts: the GPT models lead on look, 2.6 Sunburst now edges 2.5 Sunburst on preference, and 2.1 has overtaken Pro on general quality. None captures the specific failure seen here, which is 2.1 changing exact wording. That is why Pro and the GPT models still earn a place for text-critical work.
Which version you get in the Gemini app
If you are not calling an API, the choice is partly made for you. As of November 20, 2026, Google's Gemini app page lists Nano Banana 2 as free, at up to 2K, and says a Google AI plan unlocks Nano Banana 2.1. Google's help page describes the image model by the Gemini model you pick instead: Flash-Lite uses Nano Banana Lite, Flash or Pro uses Nano Banana 2, and the help page does not mention 2.1 at all. The two pages don't line up, so check which model name the app shows on your result.
To retry a sign or menu whose text came out wrong with Nano Banana Pro, set the Gemini model to Pro and choose More > Redo with Pro on the image (one heading on the help page calls the redo model "Nano Banana Pro 2", while its text says Nano Banana Pro). Once you reach your daily quota of Nano Banana 2 images, the help page says you can't redo any more with Pro. Since November 13, 2026, Google AI Plus includes only the Flash-Lite and Flash models, not Pro. Because Redo with Pro needs the Gemini model set to Pro, Plus subscribers can no longer use it; Google doesn't say this outright, but it follows from the help page and the plan change. Google AI Pro and Ultra still include the Pro model.
In ChatGPT, OpenAI made Images 2.6 available across all subscription tiers when it launched on November 4, 2026, without publishing per-plan image limits. For an app-side comparison of editing, see ChatGPT vs Gemini Image Editing: Same-Photo Test of 5 Edits.
On the API, Google has deprecated Nano Banana 2 and points migrations to 2.1, which costs half as much at 1K; as of November 20, 2026, there is still no shutdown date. Its prices remain in Nano Banana 2 API Pricing for anyone still migrating.
FAQ: Nano Banana 2.1, Pro and GPT Image 2.5
Is Nano Banana 2.1 better than Nano Banana Pro?
For most work, yes: Google's own evaluations rate it higher, Artificial Analysis's preference leaderboard ranks it eighth against Pro's eighteenth as of November 20, 2026, and it costs about a quarter of Pro per 1K image at Google's price. Pro was more literal with exact text in a same-prompt test on October 8, 2026, so it remains the safer first pick for posters, signs and menus.
What's the best version of Nano Banana right now?
As of November 20, 2026, Nano Banana 2.1 is the model Google tells API users to migrate to from the deprecated Nano Banana 2, and the version the Gemini app unlocks with a Google AI plan. Use Nano Banana Pro when exact wording matters more than price: it kept the requested wording on all three text prompts in the same-prompt test.
Is GPT Image 2.5 better than Nano Banana 2.1 for text in images?
In a one-draw-per-prompt test at 1K, GPT Image 2.5 Sunburst rendered the requested wording and prices exactly, while Nano Banana 2.1 restyled a subline, added a stray word to a Japanese sign and spelled out a price in Chinese numerals. Neither got the simplified Chinese character 单 right. GPT Image 2.5 was also two to three times slower on the route tested. GPT Image 2.6 Sunburst, run on the same prompts on November 20, 2026, matched 2.5's exact wording and made the same 单 slip.
Which is cheapest: Nano Banana 2.1, Pro or GPT Image 2.5?
At list prices as of November 20, 2026, Nano Banana 2.1 (about $0.04 per 1K image including thinking tokens), then GPT Image 2.5 or 2.6 at quality high (about $0.054), then Pro (about $0.137). At low or medium quality GPT Image 2.6 costs three to four times as much as 2.5. On LaoZhang API's per-call pricing the order is the GPT Image 2.5 or 2.6 VIP route at $0.03, 2.1 at $0.045 and Pro at $0.09.
Which model is best for editing a product photo?
All four recolored an earbud case and kept the rest of the scene in the tests on October 8 and November 20, 2026; only the GPT models left the dark inner lid dark. Nano Banana 2.1 is the cheapest of the four at list price for routine edits. If you need mask-based edits, avoid the GPT Image 2.5 VIP route, which does not accept a mask.
Is GPT Image 2.6 better than Nano Banana Pro?
For text inside an image they tied on the same prompts: GPT Image 2.6 Sunburst on November 20, 2026 and Nano Banana Pro on October 8 both reproduced the requested wording and prices, and both drew 单 as 単. On a weather-app mockup, 2.6 added no readings that weren't requested, while Pro added several. On Artificial Analysis's preference leaderboard as of November 20, 2026, 2.6 Sunburst ranks first and Pro eighteenth. Pro returned images faster on its run (18–23 seconds against 34–45), but at quality high 2.6 costs about $0.054 per 1024×1024 image at OpenAI's price against about $0.137 for Pro at Google's.
Sources3
External pages this guide links to, in the order they appear. Last updated Nov 21, 2026.
Sources3
External pages this guide links to, in the order they appear. Last updated Nov 21, 2026.





