GPT Image 2.5 Flare vs Sunburst

The short answer

Same bill, close quality, and a speed gap that grows with the tier

Latency by quality tier

Flare vs Sunburst speed is not one number

Community benchmarks disagree about how fast Flare is — 17 s, 22 s, 35 s — because they ran different quality tiers. We ran the same product-shot prompt at every tier on both models. The ratio climbs steadily from 1.18× at low to 2.26× at max.

• 01Low or medium tier: use Sunburst. It costs the same and adds 2–6 seconds.
• 02High and above in something a person is waiting on: Flare. At max the gap is over a minute.
• 03Benchmark with an explicit tier, never auto — with auto each model picks its own effort level, so you would be timing two different jobs.

Same prompt, both models

What the images actually show

Three tests from round one, chosen because each one changes a decision. Flare is on the left, Sunburst on the right, at the same tier and size; timings are per image.

Flare24 s

gpt-image-2.5-flare

Sunburst47 s

gpt-image-2.5-sunburst

Sprite sheet: both models make it, neither makes it sliceable

The most-shared GPT Image 2.5 trick on X is the 4×4 combat sprite sheet. Both models delivered four clean rows — idle, walk, attack, hit — on a keyable magenta background. Neither obeyed “uniform cell size, even spacing”, which is exactly why so many people are stuck at “how do I turn this into a GIF”: a fixed-grid slicer cuts through frames. Sunburst kept armour, cape and belt buckle identical across all 16 frames; Flare’s sword changed colour and the character changed size.

Try this comparison

Both models · high · 1:1 · 24 credits per pair

Sprite sheet playbook

Flare20 s

gpt-image-2.5-flare

Sunburst44 s

gpt-image-2.5-sunburst

Photorealistic portrait: the slower model made the famous mistake

We wrote a cafe portrait that deliberately bakes in the six “tells” Reddit crowd-sourced from the most-upvoted realism thread of the week. Sunburst won on skin, hair and readable poster text — and mounted the EXIT sign beside the door instead of above it, the single most-upvoted tell in that thread. Flare put it exactly where it belongs.

Try this comparison

Both models · high · 1:1 · 24 credits per pair

Photorealistic playbook

Flare25 s

gpt-image-2.5-flare

Sunburst40 s

gpt-image-2.5-sunburst

Exclusion lists: ten “do not add” lines emptied the room

Two exclusions gave us a cluttered, golden-hour bookshop with plants, a rug and framed prints. Ten exclusions gave us bare floors and more shelving. The model did not get dumber — half of the ten lines banned the very things that made the first version feel warm. Exclude what would ruin the image, not everything that might appear.

Try this comparison

Both models · high · 1:1 · 24 credits per pair

Prompt rules

How to choose

Pick by tier and by job, not by reputation

Playbooks

The GPT Image 2.5 tricks people are actually sharing — tested on both models

We collect what the community is doing with GPT Image 2.5, run it on Flare and Sunburst ourselves, and write down what worked, what broke, and which model to use.

Want to check any of this on your own prompt? The studio runs it on Flare and Sunburst side by side and shows you the seconds and tokens. Studio →

Sprite sheets & pixel art

The 4×4 combat sheet, the transparent-background trap, why the grid never lines up, and how to get to a GIF anyway.

Read the playbook

Photorealistic people

The six-point tell checklist from Reddit’s biggest realism thread, scored on both models, plus the reference-image method.

Read the playbook

Prompt rules

Seven rules four independent sources agree on, the one rule all four share, and the claims we tested instead of repeating.

Read the playbook

Decide

Speed, cost, the two models, and the move from gpt-image-1

The rest of what we measured, one page per question: how long each tier takes, what an image costs at list price, what each model won and lost, and what changes when gpt-image-1 shuts down.

• Is Sunburst more expensive than Flare?

  No. OpenAI prices both at $5 per million text-input tokens, $8 per million image-input tokens and $30 per million image-output tokens. In our runs the two models also returned identical output-token counts at every tier — 196 at low up to 7,024 at max for 1024×1024 — so the same request costs the same on either model.

• How much faster is Flare than Sunburst?

  It depends on the quality tier. In our text-to-image runs Sunburst took 1.18× as long at low (13 s vs 11 s), 1.60× at medium, 1.75× at high, 2.08× at xhigh and 2.26× at max (113 s vs 50 s). Complex prompts widen the gap further: Flare barely slowed down on a busy interior, Sunburst added about ten seconds.

• Which model should I use by default?

  If low or medium quality is enough for you, Sunburst — it costs the same and the wait is a few seconds longer. If you run high or above inside something a user is waiting on, start with Flare and only move to Sunburst for the shots where consistency, text or fine detail fail.

• What are the exact model IDs?

  gpt-image-2.5-flare and gpt-image-2.5-sunburst. Both support generation, editing and transparent backgrounds through the Images API, and both accept the same prompt unchanged.

• Which quality tiers does GPT Image 2.5 have?

  auto, low, medium, high, xhigh and max. The xhigh and max tiers are new in 2.5 — earlier GPT Image models stopped at high. Max cost us about 36× the output tokens of low for the same 1024×1024 prompt, so treat it as a deliberate choice, not a default.

• What image sizes are allowed?

  Any width×height where both sides are multiples of 16, the aspect ratio is at most 3:1, no side exceeds 3,840 px and the total is between 655,360 and 8,294,400 pixels. Presets include 1024×1024, 1536×1024, 1024×1536, 2048×2048, 2048×1152 and 3840×2160. Bigger sizes cost more tokens on both models equally.

• Is Sunburst better at everything?

  No. In our photorealistic test Sunburst produced better skin, hair and legible text, but placed the EXIT sign beside the doorway instead of above it — the most-upvoted tell in Reddit’s realism thread — while Flare placed it correctly. Treat “quality-optimised” as a tendency, and run a checklist on whichever model you use.

• Do I need different prompts for Flare and Sunburst?

  No. Both models take the identical prompt, size and quality parameters, which makes an A/B trivial: send the same request twice with a different model string. Set the quality tier explicitly rather than auto so both jobs do the same amount of work.

• What about gpt-image-1 and gpt-image-1.5?

  Both are being retired: gpt-image-1 on 2026-10-23 and gpt-image-1.5 on 2026-12-01. When you migrate, drop the input_fidelity parameter — it is not accepted by GPT Image 2 or 2.5 — and pick a 2.5 tier explicitly rather than relying on the old default.

• How were these numbers measured?

  Through the Images API on 2026-09-12, 1024×1024, one run per cell, from a single region in a single time window, via an API gateway. Seconds are wall-clock from request to finished image. Costs are recomputed from output tokens at OpenAI’s list price. With n=1 expect run-to-run variation; the scripts and prompts are published so anyone can rerun them, and round two will report medians.