GPT Image 2.5 vs Nano Banana 2.1: Same-Prompt Test and Price

We ran the same four prompts through GPT Image 2.5 Sunburst, GPT Image 2.5 Flare and Nano Banana 2.1 on October 26, 2026. Sunburst won the menu, Flare the product photo, and two tests were ties. A 2K image billed about $0.50 at OpenAI's rates with GPT at high quality, and about $0.056 at Google's for Nano Banana 2.1.

GPT Image 2.5 Flare made the best product photo, but drew an unrequested lighthouse on the menu and miscounted both jars. It has the same token prices as Sunburst and was not faster in our runs.

All three are on OmniaKey: GPT Image 2.5 Sunburst, GPT Image 2.5 Flare and Nano Banana 2.1.

How we tested

• Models: gpt-image-2.5-sunburst, gpt-image-2.5-flare and gemini-nano-banana-2.1 on OmniaKey. OpenAI also publishes dated snapshots, gpt-image-2.5-sunburst-2026-09-08 and gpt-image-2.5-flare-2026-09-08; we used the undated IDs.
• Settings: Sunburst and Flare at size 2048x2048 and quality high, Nano Banana 2.1 at image_size 2K, aspect_ratio 1:1 and its default thinking. Every image is 2048×2048.
• Date: October 26, 2026. The menu ran twice on Sunburst and Nano Banana 2.1 and once on Flare, and every other test once per model. With the edit source and three extra quality runs that makes 18 calls, and every one succeeded first time.
• The edit source was made with GPT Image 2, an older model, and all three models got the same file.
• Time runs from creating the task to seeing it finished, polling every 4 s through OmniaKey's asynchronous media API, so it includes queue time.
• We judged at full size: every word and price, objects counted in zoomed crops, whether each image follows every line of its prompt, and for the edit which items stayed the same.
• OmniaKey sells API access to all three models, so we ran the tests through it. Every price below is OpenAI's or Google's own list price, not ours.
• Sources: OpenAI's GPT Image 2.5 Sunburst and Flare model pages, its API pricing, image generation guide and changelog; Google's Gemini API pricing, image generation and changelog, each checked October 25, 2026.

One or two images per prompt per model is a small sample, so read each result as an example you can look at.

Four tests, same prompts

Test 1: Café menu with six prices (1:1)

This test checks every word and price on a printed menu card, and whether the prices land in one column.

Prompt:

A printed café menu card lying flat on a marble table, shot from directly above. The card is cream paper with dark green ink. At the top, a large heading reads "HARBOR CAFÉ". Below it, six items, each on its own line with the price aligned to the right: "Flat White — $4.50", "Oat Latte — $5.25", "Cardamom Bun — $3.75", "Avocado Toast — $9.00", "Shakshuka — $11.50", "Lemon Tart — $4.25". At the bottom, in small italic text: "Open daily 7am–3pm". Natural window light, soft shadows, a sprig of rosemary at one corner of the card.

• GPT Image 2.5 Sunburst (36.7 s and 34.3 s): every word and price right in both runs, including "CAFÉ" and the en dash in "7am–3pm". Both runs dropped the em dashes and set the prices in a right-aligned column, like a printed menu. The rosemary lies at the card's corner, partly off the card; run 1 added a coffee cup and a water glass at the edges, run 2 a coffee cup.
• GPT Image 2.5 Flare (42.9 s): every word and price right, em dashes dropped, prices right-aligned, with a thin rule between items and a border round the card. It added a small lighthouse-and-birds drawing above the heading, plus a coffee cup and a water glass.
• Nano Banana 2.1 (36.5 s and 32.8 s): every word and price right in both runs, and the rosemary lies on the card's corner as asked. It kept the em dashes, so the prices never line up: run 1 has leader lines of different lengths and run 2 no right alignment at all.

Verdict: GPT Image 2.5 Sunburst. All five menus spelled every word and price right, so text is a tie. Sunburst turned the prompt's "Flat White — $4.50" lines into a real menu with a right-aligned price column, twice, and added nothing to the card itself. Nano kept the dashes literally, so its prices never line up, and Flare drew a lighthouse nobody asked for.

Test 2: Counting objects on a shelf (1:1)

This test checks two exact counts in one frame, and whether the shelf stays as asked.

Prompt:

A studio photo of a wooden shelf against a light gray wall. On the shelf, from left to right: a clear glass jar holding exactly five red marbles, an empty clear glass jar, and a clear glass jar holding exactly three yellow lemons. A small white paper tag hangs from the middle jar with the number "2" written on it in black marker. A single blue feather lies on the shelf in front of the right jar. Soft, even lighting and no other objects.

• GPT Image 2.5 Sunburst (32.9 s): exactly 5 red marbles in the left jar and an empty middle jar with a "2" tag. The right jar shows 3 lemons in front plus a fourth partly hidden behind them, so 4. The feather lies in front, between the middle and right jars.
• GPT Image 2.5 Flare (37.7 s): 6 marbles in a pyramid, and 4 lemons on the right, 3 in front and one partly hidden. The middle jar is empty with the "2" tag.
• Nano Banana 2.1 (37.5 s): 7 marbles in the left jar and exactly 3 lemons on the right, stacked up the side of the jar against the glass. The middle jar is empty with the "2" tag, and the jars are taller, Mason-style.

Verdict: Tie, Sunburst and Nano Banana 2.1. No model got both jars right: Sunburst hid a fourth lemon, Nano put in seven marbles, and Flare missed both. If counts matter, check every image and try asking for items side by side ("three lemons in a row, none behind").

Test 3: Product photo of a white leather running shoe (1:1)

This test checks whether a studio brief survives: glossy reflections, neon out of focus, no text and no logos.

Prompt:

Product photo of a white leather running shoe on wet black asphalt at night, low angle, close-up. Pink and teal neon signs are out of focus in the background, and their colors reflect in the puddles and on the shoe's glossy surfaces. Small water droplets on the toe box. Sharp focus on the shoe, shallow depth of field, realistic commercial photography, no logos and no text.

• GPT Image 2.5 Sunburst (41.5 s): wet asphalt, pink and teal neon reflections, droplets on the toe, a glossy patent-like finish and no text. But the shoe is being worn: an ankle and black jogger trousers are in frame, so it is a lifestyle shot, not the product photo asked for.
• GPT Image 2.5 Flare (41.8 s): a standalone shoe on wet asphalt at night from a low angle, neon out of focus behind, colors in the puddles and on the glossy finish, droplets on the toe box, no text or logos. It is the closest match to every line of the brief.
• Nano Banana 2.1 (47.3 s): a standalone shoe in matte, perforated leather, the most like real leather of the three, with droplets on the toe box. The "no text" line slipped: the blurred pink neon behind is shaped like a word, the ground looks glittery rather than puddled, and the shoe is not glossy.

Verdict: GPT Image 2.5 Flare. It is the only one that delivered a plain product photo with glossy reflections and no text. Sunburst put the shoe on someone's foot, and Nano made the most leather-like shoe but added word-shaped neon.

Test 4: Room edit (1:1)

This test checks an edit that keeps the room: only the sofa changes and one plant arrives.

Prompt:

Change the sofa to dark green velvet and put a tall potted fiddle-leaf fig plant on the left side of the sofa. Keep everything else exactly the same: the walls, the window, the coffee table, the books, the rug, the light and the camera angle.

We made the source photo with GPT Image 2, an older model, from a prompt for a light gray three-seat fabric sofa against a white wall, a round oak coffee table with two books on it, a cream wool rug and a large window on the right with sheer curtains, in afternoon sun. All three models got the same file.

• GPT Image 2.5 Sunburst (43.1 s): a dark green velvet sofa and a tall fiddle-leaf fig in a basket on the left. The walls, both framed pictures, window, curtains, cushions, throw, oak table, both books, spine lettering, rug, light and camera angle all match the source. The bookshelf at the far left is gone, and the plant stands where it was.
• GPT Image 2.5 Flare (37.5 s): the same result as Sunburst, with the fig on the left, every listed item and the spine lettering kept, and the bookshelf gone.
• Nano Banana 2.1 (33.2 s): a dark green velvet sofa and a fig in a white pot on the left, with the bookshelf kept. But it shortened the sofa, moving its left end to the right to fit the plant along with the throw and cushions, and turned the spine lettering into different, unreadable letters. Table, rug, window and light stayed.

Verdict: Tie, Sunburst and Flare. Both kept every listed item, down to the spine lettering, and made room for the plant by removing the bookshelf. Nano kept the bookshelf but shortened the sofa and changed the spine lettering, and the prompt asked for the books to stay the same.

Results by test

Every menu spelled every word and price right, so layout decided that test: GPT treated the em dashes as separators and aligned the prices, and Nano kept them. Each model also missed one line of a brief: Sunburst made a lifestyle shot instead of a product photo, Flare drew on the menu, and Nano put word-shaped neon into a "no text" photo.

Price and speed

OpenAI prices GPT Image 2.5 by the token and publishes no per-image price for it, so the figures below are the tokens our images billed times OpenAI's rates.

OpenAI charges $30 per million image output tokens, so a high-quality image is 16,640 × $30 per million = $0.4992, plus 79–160 text input tokens at $5 per million, at most $0.0008.

At phone size the three look alike; at full size the low-quality menu is slightly softer. That is one prompt, so treat it as a hint.

Sunburst and Flare billed the same tokens at the same quality: 16,640 at high, 4,224 at medium. The edit at high added 1,659 input image tokens at $8 per million and 56 text tokens, about $0.51.

Google's list rates, checked October 25, 2026, are $30 per million image output tokens, $7.50 per million text and thinking output tokens and $1.50 per million input tokens.

Our Nano edit used 258 thinking tokens and a 1,120-token input image, about $0.054. At high quality GPT Image 2.5 costs about 9 times as much per 2K image, at medium about 2.3 times, and at low it costs less: $0.033 against $0.056.

Nano costs more than Google's listed image price because thinking is on by default at medium, billed as text output at $7.50 per million. Our Qwen Image 3 vs Nano Banana 2.1 test measured the same effect.

Both vendors charge half the standard rate for Batch jobs.

Low and medium quality did not make GPT faster: the low and medium menus took 37.3 s on Sunburst and 37.8 s on Flare medium. For OmniaKey's current rates, see the Sunburst, Flare and Nano Banana 2.1 model pages.

Sunburst or Flare?

Pick Sunburst for text, layouts and counts, and Flare for product photos; per token they cost the same. OpenAI's own advice is "Choose Sunburst for workflows where editing precision matters most, and Flare for fast, high-quality everyday image generation." Its model pages label Sunburst "Highest" at "Medium" speed and Flare "Higher" at "Very fast", both at the same token rates.

In our runs the two billed identical tokens and gave near-identical edits, and Flare made the best product photo. Flare also drew on the menu, miscounted both jars and had the slowest median at 41.8 s. OpenAI's docs list two more quality levels, xhigh and max, which we did not test, and what changed from the previous generation is in GPT Image 2.5 vs GPT Image 2.

Pros and cons

GPT Image 2.5 Sunburst

• Pros: every word right in both menu runs with prices aligned, the right marble count, an edit that kept every listed item, and the fastest median at 35.5 s.
• Cons: put the shoe on a person's foot, hid a fourth lemon, added a coffee cup and a water glass to the menu scene, and costs about $0.50 per 2K image at high quality.

GPT Image 2.5 Flare

• Pros: the best product photo, and an edit as good as Sunburst's.
• Cons: drew an unrequested lighthouse on the menu, miscounted both jars, and had the slowest median at 41.8 s despite OpenAI's "Very fast" label.

Nano Banana 2.1

• Pros: every word right, exactly three lemons, the rosemary where asked, the bookshelf kept, and about $0.056 per 2K image. Google's docs give it 4K and up to 14 reference images, which Nano Banana 2.1 vs Pro measures next to Pro.
• Cons: kept the em dashes so prices never line up, seven marbles instead of five, word-shaped neon in a "no text" photo, a shortened sofa and changed spine lettering in the edit, and thinking tokens on every image.

Which one for which job

Run the same test through one API

This sends the menu prompt from test 1 to Sunburst at 2048×2048 and high quality, as we ran it:

curl https://api.omniakey.com/v1/media/tasks \
  -H "Authorization: Bearer $OMNIAKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: harbor-cafe-menu-001" \
  -d '{
    "model": "gpt-image-2.5-sunburst",
    "type": "image",
    "operation": "generate",
    "input": {"prompt": "A printed café menu card lying flat on a marble table, shot from directly above. The card is cream paper with dark green ink. At the top, a large heading reads \"HARBOR CAFÉ\". Below it, six items, each on its own line with the price aligned to the right: \"Flat White — $4.50\", \"Oat Latte — $5.25\", \"Cardamom Bun — $3.75\", \"Avocado Toast — $9.00\", \"Shakshuka — $11.50\", \"Lemon Tart — $4.25\". At the bottom, in small italic text: \"Open daily 7am–3pm\". Natural window light, soft shadows, a sprig of rosemary at one corner of the card."},
    "parameters": {"size": "2048x2048", "quality": "high"}
  }'

1. The reply is 202 with a task id.
2. Fetch https://api.omniakey.com/v1/media/tasks/{id} every few seconds until status is succeeded.
3. Download the image from assets[0].url.

Switch model to gpt-image-2.5-flare for Flare. On OmniaKey quality takes low, medium, high or auto, and medium cut our output tokens from 16,640 to 4,224.

For Nano Banana 2.1 set model to gemini-nano-banana-2.1 and parameters to {"aspect_ratio": "1:1", "image_size": "2K"}, because OmniaKey rejects size on that model. To edit, set operation to edit and send the source as input.image with {"data": "<base64>", "mime_type": "image/jpeg"}, keeping the whole request at most 1 MiB. GPT Image models return PNG and Gemini models usually return JPEG.

Start from the GPT Image 2.5 Sunburst, GPT Image 2.5 Flare and Nano Banana 2.1 model pages, and the image API docs for every parameter.

Common questions

Is GPT Image 2.5 better than Nano Banana 2.1?

By test: Sunburst won the menu, Flare the product photo, and counting and the room edit were ties. Price goes the other way, about $0.50 against about $0.056 per 2K image at high quality. Pick the GPT models for layouts and edits you must keep, Nano when the bill matters more.

What is the difference between GPT Image 2.5 Sunburst and Flare?

OpenAI positions Sunburst for editing precision and Flare for fast, everyday generation, labelling Sunburst "Highest" at "Medium" speed against Flare's "Higher" and "Very fast". Both use the same token rates and billed the same tokens in our runs. Flare gave the best product photo, Sunburst the clean menu layout.

How many tokens does one GPT Image 2.5 image use?

At 2048×2048 ours billed 1,088 output tokens at low quality, 4,224 at medium and 16,640 at high, and the edit added 1,659 input image tokens. OpenAI's calculator does not estimate 2.5 token use, so measure a first batch and multiply by $30 per million output tokens.

Why does Nano Banana 2.1 cost more than its listed image price?

Because thinking is on by default at medium and billed as text output at $7.50 per million. Our four generations used a median of 730 thinking tokens, about $0.0055 on top of the $0.0504 image.

Is gpt-image-2 deprecated?

No. OpenAI's deprecations page has no row for gpt-image-2, and the image guide still documents it. The older gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest leave the API on December 1, 2026, with Sunburst or Flare as the replacement, and gpt-image-1 shut down on October 23, 2026.

Update log

• October 26, 2026: first version, four same-prompt tests run the same day.