Nano Banana vs GPT Image vs Recraft: Which Should You Use?

"Nano Banana" isn't a model. It's a family of at least four, and mixing them up is how people quietly pay double for the same picture. Google's naming runs Gemini 2.5 Flash Image, Gemini 3 Pro Image, Gemini 3.1 Flash Image and a Lite tier, all wearing the same fruit. OpenAI shipped GPT Image 2 in April 2026. Recraft, meanwhile, kept doing the one thing none of the others do at all.

So, Nano Banana vs GPT Image vs Recraft. The arena boards say GPT Image 2 wins. The arena boards are answering a different question from the one you have.

Yours is a job question. Ours too. Here's the answer we'd give a colleague across a desk.

The 20-second verdict

Prices are first-party and dated 26 August 2026. They churn, so re-check before you build a budget on them.

The single line worth keeping: rank tells you which model wins an average prompt, and you don't have average prompts.

"Nano Banana" is four models, and the names cost you money

The name started as a joke. Google DeepMind put Gemini 2.5 Flash Image into LMArena under an anonymous codename in August 2025, it broke engagement records, and Google leaned in rather than fighting it. Now the fruit covers a whole ladder: Nano Banana is the original 2.5 Flash Image, Nano Banana Pro is Gemini 3 Pro Image from November 2025 (up to 4K, up to 14 reference images, reliable re-creation of up to five people), Nano Banana 2 is Gemini 3.1 Flash Image from 26 February 2026, and Nano Banana 2 Lite arrived in June 2026 as the cheap fast floor. Google's Imagen family is being retired into this series.

OpenAI's line is shorter and better behaved. GPT Image 1 replaced DALL·E 3 in 2025, GPT Image 1.5 followed in December, and GPT Image 2 (snapshot gpt-image-2-2026-04-21) landed around 21 April 2026 as a quality-first, instruction-obsessed model. Mind the calendar: gpt-image-1 shuts down on 23 October 2026, and gpt-image-1-mini and 1.5 on 1 December 2026. DALL·E 2 and 3 left the API on 12 May 2026.

Recraft is the London outlier. V2 in 2024 was its first model trained from scratch, V3 shipped as red_panda that October, V4 in February 2026, and V4.1 on 14 May 2026 in eight variants including the Vector tiers.

Three families, roughly a dozen live model IDs. Write down which one you're actually calling before you compare prices with anyone.

Photorealism depends which kind of real you mean

On blind-vote arenas, GPT Image 2 (high) leads. Artificial Analysis had it at an Elo of 1,371 across 14,330 votes on 26 August 2026, with Nano Banana 2 at 1,321 and Nano Banana Pro at 1,298. Recraft's V4.1 variants sit lower, between 1,188 and 1,219.

Hands-on tests complicate that neatly. Daria Cupareanu of AI Blew My Mind ran 10 identical prompts through Nano Banana 2 and GPT Image 2 in May 2026 and found GPT Image 2 more saturated, more agency-ready, and dramatically better at holding a real face from a reference selfie, while Nano Banana 2 came out cleaner, flatter, occasionally over-smoothed. Recraft's own seven-prompt test (vendor-authored, so read it with one eye closed) put Recraft V4 Pro ahead on cinematic, editorial atmosphere and Nano Banana Pro ahead on literal, brief-faithful specs. A separate production note preferred Recraft for close-up skin, calling Nano Banana Pro's faces generic.

Think of it like three photographers with different kit. One owns the lighting rig, one owns the retouching suite, one owns the art direction.

Our read: Nano Banana Pro for lit product hero shots, GPT Image 2 for photographic polish and likeness from a reference, Recraft when the image needs to look designed rather than shot.

GPT Image reads the brief; Nano Banana takes the photo

This is GPT Image 2's clearest lane, and it isn't close. Give it a grid, ordered panels, a UI-like layout or exact object placement and it behaves like a layout-aware design assistant. Give the same brief to Nano Banana 2 and you get a fast visual photographer: lovely image, half your constraints politely ignored.

Cupareanu's test found the same thing on dense briefs. Luxury campaign boards, physics atlas pages, technical infographics: GPT Image 2 packed more of the correct elements into the frame and respected the intent behind them.

Nano Banana has a counter. In complex multi-subject spatial scenes it holds compositional logic well, and raising its thinking level improves adherence at the cost of 3 to 5 seconds. Recraft sits at the other end: it rewards precision and takes creative liberties when you under-specify, which is a gift for hero imagery and a problem for a strict logo spec.

Long prompt with rules in it? GPT Image 2. Short prompt with a mood in it? Either of the others.

Text inside images is solved, until about 200 characters

Mangled lettering used to be the giveaway of an AI image. At the top of the market that's mostly over. GPT Image 2 is marketed at roughly 99% character-level accuracy across Latin and CJK scripts, and Nest Content's 2026 testing put it first on arena Elo for overall image quality while naming Ideogram the specialist on text specifically. MindStudio's independent testing puts Ideogram V3 at about 90 to 95% text accuracy, with FLUX.2 and GPT Image 2 just behind.

Those percentages are mostly vendor or non-standardised claims, so treat them as directional. The rigorous number is different and more useful: the STRICT benchmark (Zhang et al., EMNLP 2025) found text rendering collapses past roughly 200 characters across every model tested, and the practical sweet spot sits under 25.

Which is a length limit, not a model problem. A poster headline fits. A paragraph of body copy doesn't, and no amount of model-swapping fixes that.

For headline copy, packaging and UI mockups, GPT Image 2 first. For a wordmark you'll need to edit later, Recraft, because you get real letterforms as paths. For anything genuinely text-critical, Ideogram is worth naming even though it sits outside this three-way fight.

For editing and consistency, chain two models instead of picking one

The editing boards disagree with each other, which tells you something. On Artificial Analysis, Microsoft's MAI-Image-2.6-Preview leads image editing at Elo 1,284 with GPT Image 2 (high) at 1,256. On Lumenfall's board, Nano Banana Pro leads at 1,249, narrowly over Nano Banana 2, with GPT Image 2 at 1,238. Different scoring systems, different vote pools, no clean answer.

Consistency splits the same way. Cupareanu found GPT Image 2 far better at preserving one real face across styles from a single reference, where Nano Banana 2 lost the likeness. But Nano Banana Pro's documented strength is the opposite direction: up to 14 reference images and reliable re-creation of up to five people, which is what you want holding a product or a cast across a campaign. GPT Image 2 accepts up to 16 reference images per edit request.

The practitioners who do this daily stopped choosing. invideo documented a jewellery campaign built by chaining: GPT Image 2 for the base composition, then Nano Banana Pro to lock the exact product across every shot.

Build with the one that follows instructions. Lock with the one that remembers faces.

Recraft is the only one that hands you a real SVG

Here's the catch with almost every AI logo: it's a PNG in a costume. Ask GPT Image 2 or Nano Banana for a vector logo and you get raster pixels, no matter how the prompt is phrased.

Recraft natively generates production-ready SVG with real, editable Bézier paths, the kind you can open in Figma or Illustrator and actually pull a node on. V4.1 Vector produces one in about 13 seconds, and the Pro Vector tier targets high-aesthetic logos, icons and line illustration. Recraft also does brand-style lock, style reference and mixing, transparent backgrounds and print-ready DPI and CMYK output.

It has limits, and they're worth saying out loud. Recraft does best on simpler geometric and semi-geometric forms; complex photoreal vector is still an emerging area, and a serious brand mark still benefits from a human refining it afterwards.

That's the whole verdict for the design job. If the asset has to scale, be recoloured, or survive a client's revision round, Recraft. If it just has to look good at one size, anything.

What they cost and how fast they run, as of 26 August 2026

All three meter by tokens or credits now, which makes headline comparisons useless and per-image maths necessary.

Google publishes the cleanest numbers. Nano Banana 2 image output runs $60 per million tokens, and a 1K image eats 1,120 tokens, so about $0.067, rising to $0.101 at 2K and $0.151 at 4K. Nano Banana Pro is $120 per million output tokens, so roughly $0.134 at 1K to 2K and $0.24 at 4K. Nano Banana 2 Lite is $30 per million, about $0.0336 per 1K image.

OpenAI meters GPT Image 2 at $8 per million image input, $2 cached, $30 image output and $5 text input, which works out to roughly $0.006 at low quality, $0.053 at medium and $0.211 at high for a 1024×1024 image. Edit requests bill reference images at high fidelity and add a 2 to 3 times multiplier, so an editing-heavy workflow costs more than the sticker price suggests.

Recraft prices in units, at $0.001 each. V4.1 raster is 35 units ($0.035), V4.1 Vector 80 units ($0.08), V4.1 Pro raster 210 units ($0.21) and Pro Vector 300 units ($0.30). Subscriptions start at a free tier of daily credits and a Basic plan at $12 a month for 1,000 credits, which is 1.2 cents a credit.

Batch APIs cut Google and OpenAI costs by around half if your work isn't interactive.

Speed roughly tracks price. GPT Image 2 feels fastest at around 3 seconds, though its thinking mode adds 15 to 30 seconds when it plans before drawing. Nano Banana 2 returns 1K images in about 4 seconds; Lite is faster still, at a 3.8-second median in one scientific-figure benchmark from K-Dense where GPT Image 2 took roughly 49 seconds on the same test. Nano Banana Pro sits at 10 to 15 seconds. Recraft runs about 6 seconds raster, 13 vector.

One gap we have: we pulled OpenAI's per-image figures via fal.ai, an authorised host, because OpenAI's own pricing page renders the numbers in a dynamic calculator. The token rates match across sources, but check your own dashboard before you forecast a year of spend on them.

The licence footnotes that bite six months later

Nobody reads these until a client asks, so read them now.

OpenAI assigns output ownership to you in its terms, subject to policy and third-party rights. Other users can get similar output from similar prompts, which matters if you thought your hero image was yours alone.

Google also says you own your generated images, with a permanent asterisk: every one carries an invisible SynthID watermark that can't be disabled through the API, and free and Pro tier images add a visible Gemini sparkle. Ultra subscribers and AI Studio lose the visible mark, not the invisible one.

Recraft grants full ownership and commercial rights on paid plans. On the free plan, the images are owned by Recraft, public in the community gallery, and not licensed for commercial use, which rules out listing them on stock platforms.

None of the three offers a copyright-registration guarantee, and IP indemnification is thin across the board. If a client needs indemnity in writing, this whole comparison is the wrong shortlist.

Nano Banana vs GPT Image vs Recraft, by who you are

Etsy or Shopify seller. Nano Banana 2 for volume product shots at $0.067 a go, Recraft V4.1 Vector when you need a shop logo or sticker artwork that prints. Skip GPT Image 2 high quality unless a hero shot earns the $0.211.

Indie founder building a landing page. GPT Image 2 for anything with a layout or a headline in it, Recraft for the icon set, Nano Banana 2 for the photographic filler. Two models minimum, and you'll use all three within a month.

Small-agency designer. Recraft for deliverables the client will edit, Nano Banana Pro for campaign consistency across shots, GPT Image 2 for the pitch boards. Chain them the way invideo did.

The obvious objection: learn one tool deeply and stop faffing about. Fair, and for a while it worked. It stopped working when the failure modes diverged, because a single-model habit now costs you either the vector file, the readable headline or the face that stays the same.

There's a bigger objection too. A fourth model might beat all three at your specific job: MAI-Image-2.6 leads the editing arena, Ideogram usually wins on text, FLUX.2 is the strongest open option, and Reve, Seedream 5.0 Pro and Qwen-Image-3.0-Pro all rank near the top on text-to-image. We've scoped this to the three families we run daily. That's a choice, not a claim that nothing else is good.

And treat the boards themselves with some suspicion. Elo measures aggregate human preference on generic prompts, scores from different platforms aren't directly comparable, and the Llama 4 LMArena episode showed what happens when a vendor plays to the test. Recraft ranks mid-table on raw text-to-image Elo while being the only one of the three that can hand you an editable file. The arenas simply don't measure that.

Nano Banana vs GPT Image vs Recraft: quick answers

Which is best for logos and SVG files?
Recraft, and it isn't a close call. Recraft V4.1 Vector is the only model of the three that outputs production-ready SVG with real editable Bézier paths, generated in about 13 seconds and openable in Figma or Illustrator. GPT Image 2 and the Nano Banana models return raster pixels regardless of how you word the prompt.

Is GPT Image 2 better than Nano Banana 2?
On aggregate blind votes, yes: Artificial Analysis had GPT Image 2 (high) at Elo 1,371 against Nano Banana 2's 1,321 on 26 August 2026. For a specific job, often no. Nano Banana 2 is roughly three times cheaper at high quality, returns 1K images in about 4 seconds, and holds lighting on product shots well.

Which of the three is cheapest per image?
Recraft V4.1 raster at $0.035 beats Nano Banana 2 at $0.067, and GPT Image 2 undercuts both at $0.006 if you accept low quality. At high quality GPT Image 2 costs $0.211, roughly six times Recraft's raster price. Nano Banana 2 Lite sits lowest overall at about $0.0336. Prices dated 26 August 2026.

Can I use these images commercially?
Usually yes, with conditions. OpenAI assigns output ownership to you, and Google says you own Gemini images while stamping every one with an invisible SynthID watermark. Recraft grants commercial rights on paid plans only: free-plan images belong to Recraft and are public in its gallery. None of the three offers meaningful IP indemnification.

Which model handles text inside an image best?
GPT Image 2 for headline copy, posters and UI mockups, with Ideogram V3 better still if text is the whole point. The real constraint isn't the model: the STRICT benchmark in 2025 found rendering collapses past roughly 200 characters in every model tested, with reliability best under 25.

Do I need more than one of these?
If your work spans photos, layouts and logos, yes. Each family fails in a different place, so a single-model workflow loses either the vector file, the readable text or the consistent face. Most people we watch settle on two, then reach for the third within a month.

You don't have to pick a favourite fruit

Back to the naming mess we opened on. Four bananas, three families, a dozen model IDs, and a leaderboard that ranks them on prompts nobody actually writes. The reason that's confusing is that everyone keeps asking which model is best, when the only answerable version of the question is which model is best at this.

Photoreal and editing goes to Nano Banana. Long fussy prompts and readable text go to GPT Image. Anything that has to scale or be edited goes to Recraft. That's the map, and it'll survive the next few releases even as the Elo numbers shuffle.

Pick your next three jobs and route them deliberately. In JammyJar all three sit behind one prompt box and one bill, so switching model mid-project is a click rather than a new subscription, and a saved theme keeps the style steady while you move between them.

Start with the job you've been avoiding because the last model got it wrong.