GPT Image 1 vs 1.5 vs 2: Migration, Specs, and Pricing (2026)

OpenAI released three GPT Image generations in twelve months: gpt-image-1, gpt-image-1.5, and gpt-image-2. Each one rewrote the request shape, changed the pricing model, and left the previous generation on a deprecation timer. If your codebase still references gpt-image-1, the deadline that matters is October 23, 2026, when OpenAI shuts the route down. This guide covers what changed at each step, how the three generations compare on quality and price, and how to migrate cleanly.

TL;DR: gpt-image-1 is deprecated and shuts down October 23, 2026. gpt-image-1.5 (December 16, 2025) is a stable, conservative production target with the same prompt behaviour as 1.0 and better instruction following. gpt-image-2 (April 21, 2026) is the current best option: best text rendering, no warm yellow color cast, supports 3:1 / 1:3 ultrawide aspect ratios, native 2K output. On Unifically: GPT Image 2 prices flat at $0.03 / $0.05 / $0.06 per image for 1K / 2K / 4K. The legacy gpt-image-1 and gpt-image-1.5 routes are hidden in the pricing UI but still callable for in-flight workloads. Live rates: pricing page.

Three generations at a glance

Spec	gpt-image-1	gpt-image-1.5	gpt-image-2
Status	Deprecated. Shutdown October 23, 2026	Stable, hidden in Unifically pricing UI	Current. Recommended for new work
Released	Original GPT Image generation	December 16, 2025	April 21, 2026
Output resolution	1024×1024 baseline	1024×1024 baseline	Up to 2K, ultrawide aspect ratios
Aspect ratios	1:1, 2:3, 3:2	1:1, 2:3, 3:2	1:1, 2:3, 3:2, plus 3:1 / 1:3 ultrawide
Request shape	`quality` (low / medium / high)	`quality` (low / medium / high)	`resolution` (1k / 2k / 4k)
Text rendering	Workable, occasional artifacts	Better instruction following	Near-perfect; clean small text and UI elements
Color cast	Warm yellow tint visible	Warm yellow tint partially fixed	No yellow cast; neutral output
Reasoning	None	None	Adds reasoning step before generation
OpenAI direct pricing (1024×1024 ref)	low / med / high (legacy)	$0.009 / $0.034 / $0.133	$0.006 / $0.053 / $0.211
Unifically pricing	hidden, in-flight only	hidden, in-flight only	$0.03 / $0.05 / $0.06 (1K / 2K / 4K)

What changed between gpt-image-1 and gpt-image-1.5

gpt-image-1.5 released December 16, 2025 as a small upgrade. The headline changes:

Better instruction following. Prompts that asked for specific layouts, text, or compositional rules landed more reliably than on 1.0.
Same prompt behaviour overall. gpt-image-1.5 was positioned as a stability upgrade. Drop-in for most workloads, with a gentler learning curve than jumping straight to 2.0.
Same quality knob. Low / medium / high stayed as the request shape. No reformatting of the call payload was needed.
Slightly cheaper at low and medium quality ($0.009 / $0.034 vs the legacy 1.0 rates).

For teams that wanted gpt-image-1's look but needed slightly tighter prompt adherence, 1.5 was the safe upgrade. It is still callable on Unifically (hidden in the pricing UI, but the API works) for in-flight workloads that have not fully migrated.

What changed between gpt-image-1.5 and gpt-image-2

gpt-image-2 released April 21, 2026 and is the much bigger jump. Five things changed:

Request shape pivoted from quality to resolution. The old low / medium / high quality knob is replaced with explicit 1K / 2K / 4K output sizes. Pricing follows resolution, so drafts at 1K and finals at 4K are budgetable.
Text rendering is near-perfect. Small text, UI elements, manga panels, pixel art, packaging copy all render cleanly where 1.0 / 1.5 produced artifacts on dense type.
Warm yellow color cast eliminated. The 1.0 / 1.5 line had a recognisable warm tint on neutral scenes. 2.0 outputs are neutral.
Wider aspect ratio support. 3:1 and 1:3 ultrawide ratios join 1:1, 2:3, 3:2.
Reasoning step before generation. OpenAI added a planning pass, useful on prompts with complex spatial relationships, multi-element compositions, and dense scene descriptions.

The cost is that the old quality parameter does not exist on 2.0. Calls that hardcoded quality: 'high' need to be rewritten to use resolution: '4k' (or whichever variant matches the output target).

Pricing comparison

OpenAI direct rates are token-based on 2.0 and per-quality-variant on 1.5 / 1.0. Unifically prices flat per resolution variant for 2.0, with no token math. Numbers below were accurate at the time of writing; check the pricing page for current rates.

Model	Variant	OpenAI direct (1024×1024)	Unifically (per image)
gpt-image-1	Low quality	legacy rate	hidden / in-flight only
gpt-image-1	Medium quality	legacy rate	hidden / in-flight only
gpt-image-1	High quality	legacy rate	hidden / in-flight only
gpt-image-1.5	Low	$0.009	hidden / in-flight only
gpt-image-1.5	Medium	$0.034	hidden / in-flight only
gpt-image-1.5	High	$0.133	hidden / in-flight only
gpt-image-2	1K	$0.006 (low quality ref)	$0.03
gpt-image-2	2K	$0.053 (medium quality ref)	$0.05
gpt-image-2	4K	$0.211 (high quality ref)	$0.06

Direct rates from OpenAI's published pricing for gpt-image-1.5 and gpt-image-2 (1024×1024 reference points). Unifically's flat per-variant shape lines up with OpenAI's high-quality output rates while staying predictable for budgeting.

Migrating from gpt-image-1 or gpt-image-1.5

OpenAI's deprecations page lists gpt-image-1 shutdown for October 23, 2026. Calls to that route will start returning errors after that date. Two practical migration paths:

Option A: migrate to `gpt-image-2` (recommended for new work)

Best for: workloads that benefit from the latest text rendering, neutral colour, ultrawide aspect ratios, and the resolution-priced variant shape.

What changes in your code:

Switch model: 'openai/gpt-image-1' to model: 'openai/gpt-image-2'.
Replace quality: 'low' | 'medium' | 'high' with resolution: '1k' | '2k' | '4k'.
Optionally add ultrawide aspect ratios (3:1 / 1:3) where useful.
Keep image_urls[] as is (still up to 16 references at 100 MB each on 2.0).

Approximate quality mapping:

quality: 'low' to resolution: '1k'
quality: 'medium' to resolution: '2k'
quality: 'high' to resolution: '4k'

Option B: migrate to `gpt-image-1.5` (low-risk stopgap)

Best for: workloads where output stability matters more than the latest model behaviour. gpt-image-1.5 shares prompt behaviour with gpt-image-1, has better instruction following, and runs cheaper on the medium / high bands.

What changes in your code:

Switch model: 'openai/gpt-image-1' to model: 'openai/gpt-image-1.5'.
Keep quality parameter as is (low / medium / high still works).
Keep image_urls[] and prompt structure as is.

Note: gpt-image-1.5 is also slated for deprecation eventually. Treat Option B as a short-term stability bridge, not a long-term destination.

Code: migrating a `gpt-image-1` call to `gpt-image-2`

Before (`gpt-image-1`)

const start = await fetch(`${API}/v1/tasks`, {
  method: 'POST',
  headers,
  body: JSON.stringify({
    model: 'openai/gpt-image-1',
    input: {
      prompt: 'A photorealistic packaging mockup of a matte black bottle on marble',
      quality: 'high',
      aspect_ratio: '2:3',
      image_urls: ['https://example.com/brand-reference.jpg'],
    },
  }),
}).then((r) => r.json());

After (`gpt-image-2`)

const start = await fetch(`${API}/v1/tasks`, {
  method: 'POST',
  headers,
  body: JSON.stringify({
    model: 'openai/gpt-image-2',
    input: {
      prompt: 'A photorealistic packaging mockup of a matte black bottle on marble',
      resolution: '4k',
      aspect_ratio: '2:3',
      image_urls: ['https://example.com/brand-reference.jpg'],
    },
  }),
}).then((r) => r.json());

The only meaningful change is quality: 'high' to resolution: '4k'. Polling on /v1/tasks/{task_id} is identical across all three model generations.

When to use each model

Use `gpt-image-2` when

Starting a new workload. There is no good reason to start on a deprecated model.
Text rendering matters (UI mockups, packaging copy, posters, manga).
You need neutral output without the warm yellow cast.
You need ultrawide aspect ratios (3:1 or 1:3).
You want explicit resolution variants for budgeting (drafts at 1K, finals at 4K).

Use `gpt-image-1.5` when

You are mid-migration and need a low-risk stopgap.
Output stability matters more than the latest model behaviour.
Your codebase still passes quality and you cannot refactor immediately.
The cheaper medium-quality variant ($0.034 direct) hits the budget you need.

Avoid `gpt-image-1` because

It shuts down on October 23, 2026.
gpt-image-1.5 is a drop-in stability bridge with better instruction following.
gpt-image-2 is materially better on every quality dimension.

Migration pitfalls to watch for

Skipping the migration deadline. October 23, 2026 is the hard cutoff for gpt-image-1. Calls after that return errors. Do not leave it to the last week.
Hardcoding quality: 'high' and forgetting to update on 2.0. The quality parameter does not exist on gpt-image-2. Set resolution instead.
Migrating to 1.5 thinking it is safe forever. 1.5 is a stable bridge. OpenAI will eventually deprecate it too. Plan a 2.0 path even if you take the 1.5 stopgap.
Assuming Unifically's quality mapping carries over. Unifically prices gpt-image-2 flat per resolution variant ($0.03 / $0.05 / $0.06). The OpenAI-direct token-based pricing is not what you pay through Unifically.
Forgetting the warm cast. If your downstream pipeline applies a warm-tint LUT to compensate for gpt-image-1's yellow cast, that LUT will over-warm gpt-image-2 output. Strip it on migration.

Frequently asked questions

When does `gpt-image-1` shut down?

October 23, 2026, per OpenAI's deprecations page. Calls after that date will return errors. Migrate to gpt-image-2 (recommended) or gpt-image-1.5 (stable stopgap) before then.

What is the biggest difference between `gpt-image-1.5` and `gpt-image-2`?

Three things. The request shape switched from quality (low / medium / high) to resolution (1K / 2K / 4K). Text rendering went from "workable" to "near-perfect". The warm yellow color cast that defined the 1.0 / 1.5 line is gone on 2.0. Plus ultrawide aspect ratios (3:1 / 1:3) and a reasoning step before generation.

How much does GPT Image 2 cost on Unifically?

$0.03 per image at 1K, $0.05 at 2K, and $0.06 at 4K. There is no subscription. Billing is per generated image against the openai/gpt-image-2 price key. Approximate quality mapping from 1.5: low ≈ 1K, medium ≈ 2K, high ≈ 4K. Check the pricing page for current rates.

Can I still call `gpt-image-1` on Unifically?

For now, yes. The route is hidden in the live pricing UI but the API is still callable for in-flight workloads. After OpenAI's October 23, 2026 shutdown, calls will return errors. Migrate before then.

Should I jump to GPT Image 2 directly or migrate to 1.5 first?

Jump to 2.0 if your team can absorb the prompt-shape change (quality to resolution) and re-test outputs. The quality and pricing benefits are material. Use 1.5 as a stability stopgap only if you cannot refactor before the October 2026 deadline.

GPT Image 2 deep dive: full GPT Image 2 specs and code samples.
GPT Image 2 model page: live playground with 1K / 2K / 4K toggle.
Nano Banana 2: Google option with the same per-variant pricing shape.
Qwen Image vs Nano Banana: when Google's safety filter is too restrictive.
Flux.2: Black Forest Labs option with Flex / Pro / Max variants.