GPT Image
GPT Image is OpenAI's image generation model family, accessed through ChatGPT or the API. It leads on in-image text rendering and multi-constraint instruction following, but has a documented style ceiling, strict content moderation, token-based API billing that resists simple per-image cost estimates, and a free tier too small for production work.
What is GPT Image?
GPT Image is OpenAI's current image generation model family, replacing DALL-E 3 after OpenAI removed DALL-E 2 and DALL-E 3 from the API on May 12, 2026. It is not a single model but a tiered lineup: gpt-image-1-mini for lightweight, cost-sensitive work; gpt-image-1.5 as a mid-tier; gpt-image-2 and its dated snapshot; and the gpt-image-2.5 pair, Flare (speed-oriented) and Sunburst (precision-oriented), released September 8, 2026. gpt-image-1 was shut down on October 23, 2026.
The model is exposed through two distinct channels. Inside ChatGPT, image generation is a tool feature whose usage draws from the plan's tool allowance; users do not see a per-image charge. Through the OpenAI API, every request is billed by token—text input, image input, and image output are priced separately, and the output token count varies with resolution, quality, and aspect ratio. These are not interchangeable billing models, and the same underlying model can feel “free” inside ChatGPT while costing measurable money per API call.
GPT Image's official positioning is not “most beautiful image.” Its documented differentiator is the ability to follow compositional instructions and render text accurately—use cases such as a product shot with only the background changed, or a poster with specified text blocks and a reserved CTA area. That is a narrower and more operational claim than “best image quality,” and it should be read that way.
How does GPT Image work?
A user submits a text prompt and, for editing or reference-based generation, up to 16 input images. The model processes the prompt and optional visual context, then returns generated or edited images. In ChatGPT, this happens conversationally: the user can ask for a change in natural language and iterate without re-specifying the entire scene. Through the API, the caller selects a model, size, quality level, and output format, and receives the image and a usage record that itemizes token consumption by category.
For image editing, GPT Image supports mask-based inpainting and a markup parameter for transparent-PNG-guided edits on 2.5 and 2.6 models. The Responses API can also route image generation as a tool call from a language model, but in that configuration the outer language model's token usage is billed separately and is not reflected in the image model's own pricing table.
Quality tiers (low through max, plus auto) trade output tokens against visual fidelity. Batch processing halves image output token cost from $30/1M to $15/1M, at the cost of latency. The practical workflow is therefore not “generate one image and stop” but “choose a quality tier, estimate token consumption, and iterate”—which is why OpenAI provides an output-token calculator for 2.5, even though that calculator excludes input-side costs.
Use Cases
- Producing marketing and social assets where text must render legibly inside the image, rather than being added in post
- Following multi-constraint layout prompts such as “keep the product unchanged, replace only the background” or “create a poster with three text blocks and reserved CTA space”
- Iterating on visual concepts conversationally inside ChatGPT, where the absence of per-image billing makes rapid iteration cheaper than API-driven iteration
- Batch-generating product or editorial images at scale through the API with Batch processing at half the output token rate
- Editing existing images with reference inputs and masks, within the model's 16-image and 50MB-per-file constraints
Pricing
Pricing model: Freemium
| Plan | Price | Billing | Included usage | Notes |
|---|---|---|---|---|
| ChatGPT Free (in-app) | Free | — | Image generation inside ChatGPT is available but tightly limited. Community-reported limits are roughly 2–3 images per 24-hour rolling window, though OpenAI does not publish an exact figure and limits vary by load and system conditions. Access is to the standard model tier only. | — |
| ChatGPT Plus / Pro / Business (in-app) | Varies by ChatGPT plan | — | Higher in-app image generation limits than Free, but still described by OpenAI as subject to limits that vary by plan, model, feature, and system conditions. Exact quotas are not published. | — |
| API — GPT Image 2.5 (Flare / Sunburst) | Token-based | — | Text input $5.00/1M tokens; cached text input $1.25/1M. Image input $8.00/1M tokens; cached image input $2.00/1M. Image output $30.00/1M tokens under Standard processing, $15.00/1M under Batch. Text output is not billed because these models output images. Token rates match GPT Image 2, but the GPT Image 2 calculator does not estimate GPT Image 2.5 usage. | — |
| API — GPT Image 1 Mini / 1.5 | Token-based | — | Lower-cost tiers for lighter or higher-volume workloads. Per-image rates through providers such as fal.ai start around $0.005 for 1 Mini. These models carry narrower capability sets and are positioned below the 2.5 Flare/Sunburst pair. | — |
Common Limitations
- Free-tier quota is effectively a demo, not a workflow · The in-app Free tier generates roughly 2–3 images per 24 hours according to community reports. OpenAI does not publish an exact quota, and the limit is described as varying by feature, plan, and system conditions. This is sufficient to test the model, not to produce a deliverable. Anyone using GPT Image for recurring work will need a paid ChatGPT plan or API access, and the paid in-app quotas are also unpublished, so capacity cannot be planned from official documentation alone.
- Style ceiling: strong at instruction-following, weak at stylization · GPT Image's documented strength is text rendering and multi-constraint instruction following—for example, keeping a product unchanged while replacing only the background, or laying out a poster with three text blocks and a CTA zone. Its corresponding weakness is style control. Independent evaluations note that attempts to push toward anime, watercolor, or other strongly stylized aesthetics tend to drift back toward photorealism. Cross-image character consistency is also limited. Teams whose primary need is a distinctive artistic look will find GPT Image underperforming relative to Midjourney, while teams producing text-heavy marketing assets or layout-constrained graphics are its intended fit.
- Content moderation is active and cannot be fully disabled · OpenAI applies safety classifiers to both prompts and generated outputs. A content_policy_violation error blocks the request when either the prompt or the generated image trips the filter. The API exposes a moderation parameter with low and auto settings, but this adjusts strictness rather than removing filtering; prohibited categories such as CSAM and non-consensual intimate imagery remain blocked at all settings. Creators working with horror, medical, or edge-case editorial content should expect intermittent blocks and build retry or prompt-rephrasing logic into their pipelines rather than assuming deterministic success.
- Token-based billing makes per-image cost non-obvious · The API prices text input, image input, and image output separately, and image output tokens vary with resolution, quality, and aspect ratio. There is no single per-image list price. A Standard 1024×1024 request, an HD 1536×1024 request, and an edited image with a reference input will all cost different amounts. The official GPT Image 2.5 output token calculator estimates only output cost and excludes text input, image input, and streaming intermediate images, so it does not produce a complete request bill. Teams budgeting at volume should instrument their own token accounting rather than relying on third-party per-image estimates.
- Input constraints and model fragmentation · GPT Image models accept up to 16 input images per edit request, each a PNG, WebP, or JPG under 50MB, with a maximum prompt length of 32,000 characters. Arbitrary resolutions are supported on 2.5 and 2 models as WIDTHxHEIGHT strings, but both edges must be divisible by 16, the aspect ratio must fall between 1:3 and 3:1, and resolutions above 2560×1440 are marked experimental. The maximum supported resolution is 3840×2160. Separately, the model lineup is fragmented: gpt-image-1 was shut down on October 23, 2026, and the current API spans 1-mini, 1.5, 2, 2-2026-04-21, 2.5-flare, 2.5-sunburst, and chatgpt-image-latest, each with different capability and pricing profiles. Building against a dated snapshot without a migration plan is a real maintenance risk.
- DALL-E 2 and DALL-E 3 are retired from the API · OpenAI removed DALL-E 2 and DALL-E 3 from the API on May 12, 2026, and directs developers to GPT Image models. Existing integrations pinned to dall-e-2 or dall-e-3 will fail. The GPT Image models are not parameter-compatible drop-in replacements: prompt length limits, supported sizes, and moderation behavior differ. Migration requires testing, not just a model name swap.
Key Features
- In-image text rendering with low error rates on Chinese and Latin scripts
- Multi-constraint instruction following for layout and compositional prompts
- Conversational image editing and iteration inside ChatGPT
- Reference-image editing with up to 16 input images per request
- Mask-based inpainting and transparent-PNG markup guidance on 2.5 and 2.6 models
GPT Image vs. Midjourney, Stable Diffusion, and FLUX
This comparison uses official pricing, model documentation, and independently published evaluation data available as of November 2026. It focuses on documented capability boundaries and billing structures rather than declaring a universal winner. The image model market has no single best option: GPT Image leads on text rendering and instruction following, Midjourney on aesthetic style, Stable Diffusion on open customization, and FLUX on photorealistic portraiture and open-weight availability.
| Tool | Positioning | Key differences | Best fit | Pricing note |
|---|---|---|---|---|
| GPT Image GPT-Image-2.5 Flare model page (pricing and capabilities) Create image API reference (constraints and parameters) ChatGPT free tier FAQ (image generation limits) | OpenAI's image generation model family, positioned around in-image text rendering, multi-constraint instruction following, and conversational editing inside ChatGPT. | Strongest documented performance on text rendering and layout-constrained prompts. Weakest on stylization: attempts at anime, watercolor, or other strong aesthetics drift toward photorealism, and cross-image character consistency is limited. API billing is token-based, not per-image. Free tier is a small demo allowance (~2–3 images/24h by community report). Content moderation cannot be disabled, only relaxed from auto to low. DALL-E 2 and 3 are retired from the API as of May 12, 2026. | Marketing assets, posters, infographics, and social graphics where text must render legibly inside the image and the prompt contains layout or multi-constraint instructions. | ChatGPT in-app: included in plan allowances, Free tier heavily limited. API: text input $5/1M, image input $8/1M, image output $30/1M Standard or $15/1M Batch. No per-image list price; per-image figures are token-derived estimates. |
| Midjourney Midjourney plans (official pricing) 2026 image model evaluation (independent, documents style vs text-rendering split) | An art-first image generation service, subscription-based and accessed primarily through its own interface and Discord, with a strong reputation for aesthetic quality and style control. | Midjourney's documented advantage is aesthetic quality and stylistic distinctiveness, the inverse of GPT Image's profile. It offers style references and a large community style ecosystem. It is subscription-only with no API, which is a structural difference from GPT Image's API-first access. Midjourney V8 rewrote the codebase for roughly 5x faster generation and added native 2K output, but its text rendering and multi-constraint instruction following remain behind GPT Image. It does not expose token-based billing or per-call API integration. | Artists and designers whose primary requirement is a distinctive visual style—concept art, editorial illustration, stylized branding—rather than text-inside-image precision or programmatic API integration. | Subscription only. Basic $10/mo, Standard $30/mo, Pro $60/mo, Mega $120/mo. No API. No per-image billing. |
| Stable Diffusion Stability AI official site 2026 image model evaluation (documents SD4 control/cost profile) | An open-weights image generation model family, positioned around local deployment, fine-tuning, and maximal user control through ControlNet, LoRA, and inpainting. | Stable Diffusion's core differentiator is controllability and zero marginal cost at self-hosted scale. Models can run locally on consumer GPUs, fine-tuned on custom datasets, and composed with ControlNet for pose, depth, and composition constraints. It is not a managed API product in the same sense as GPT Image. Its documented weaknesses are text rendering and complex-prompt consistency. GPT Image and Stable Diffusion occupy opposite ends of the control-versus-convenience spectrum: GPT Image is managed, instruction-driven, and token-billed; Stable Diffusion is self-managed, weight-level customizable, and infrastructure-billed. | Teams that need local deployment, full control over model weights, repeatable character or composition constraints via ControlNet/LoRA, or near-zero marginal cost per image at high volume. | Model weights are free to self-host. API access through Stability AI or third-party providers is per-image or per-compute; self-hosted cost is GPU infrastructure. No token-based billing comparable to GPT Image. |
| FLUX Black Forest Labs official site TokenCalculator image model pricing comparison (documents FLUX per-image rates) | A Black Forest Labs image model family, positioned around photorealism, fast generation, and a mix of hosted API and open-weight variants. | FLUX's documented strengths are photorealistic portraiture, skin texture, and generation speed—independent evaluations place it ahead of GPT Image on photorealism, while GPT Image leads on text rendering and layout instruction following. FLUX offers open-weight variants (Dev, Schnell) alongside hosted Pro and Ultra tiers, giving it a deployment flexibility that GPT Image's managed-only API does not match. FLUX per-image pricing through hosted providers is typically lower than GPT Image's token-derived per-image cost for comparable output. | Teams producing photorealistic imagery—product photography, editorial portraits, realistic scenes—that want either hosted access at a lower per-image cost or open-weight self-hosting for customization. | FLUX 1.1 Pro around $0.04/image; FLUX 1.1 Pro Ultra around $0.06/image through hosted providers. Open-weight variants (FLUX.1 Dev, FLUX.1 Schnell) are free to self-host under their respective licenses. |
Comparison checked against official product information on 2026-11-03. Features and pricing may change.
Frequently Asked Questions
Is GPT Image free to use?
Inside ChatGPT, image generation is available on the Free tier but tightly limited—community reports place it at roughly 2–3 images per 24 hours, and OpenAI does not publish an exact quota. That is enough to test the model, not enough for recurring production work. The API has no free image-generation tier under standard billing; third-party gateways such as Vercel AI Gateway may offer trial credits, but those are gateway promotions, not OpenAI free access.
How much does GPT Image cost per image through the API?
There is no single per-image list price. The API bills by token: text input $5/1M, image input $8/1M, and image output $30/1M under Standard processing or $15/1M under Batch. A Standard 1024×1024 generation is commonly estimated at around $0.02–0.04, HD 1536×1024 around $0.07, and higher resolutions or quality tiers cost more, but these are token-derived estimates, not fixed prices. The official 2.5 output-token calculator covers only output cost and excludes input-side tokens, so it does not produce a complete request bill.
Is GPT Image better than Midjourney for artistic work?
No, and OpenAI's own positioning does not claim that. GPT Image's documented strength is instruction following and text rendering, not aesthetic style. Independent evaluations note that GPT Image attempts at anime, watercolor, or other strongly stylized aesthetics tend to drift toward photorealism, and cross-image character consistency is limited. Midjourney is the stronger choice when the primary requirement is a distinctive artistic look; GPT Image is the stronger choice when the prompt contains layout constraints or must render legible text inside the image.
Can I control how strict GPT Image's content filter is?
Partially. The API exposes a moderation parameter with low and auto settings. auto is the default; low is less restrictive. However, this adjusts strictness rather than disabling filtering, and prohibited categories such as CSAM and non-consensual intimate imagery remain blocked at all settings. Requests can still fail with a content_policy_violation error when either the prompt or the generated output trips the classifier. Pipelines should treat moderation blocks as expected failure modes, not anomalies.
What happened to DALL-E 3?
OpenAI removed DALL-E 2 and DALL-E 3 from the API on May 12, 2026, and directs developers to GPT Image models. Integrations pinned to dall-e-2 or dall-e-3 will fail. GPT Image is not a parameter-compatible drop-in replacement: prompt length limits differ (32,000 characters for GPT Image models versus 4,000 for DALL-E 3), supported sizes differ, and moderation behavior differs. Migration requires testing, not just changing the model name.
Alternative Tools
-
Ideogram Ideogram is an AI image generation platform built around text rendering accuracy, with a free weekly credit tier, priority-credit subscriptions from Plus to Team, and per-image API pricing across three quality tiers. -
Nano Banana Nano Banana is Google's Gemini image-generation and editing experience for creating visuals, transforming reference images, combining photos, and placing text into designs. -
FLUX FLUX is Black Forest Labs' family of image and video generation models—available via a pay-per-use API, open-weight releases for local deployment, and a free browser Playground. It is the most widely adopted open image model family in 2026, with commercial licensing split by revenue and deployment mode. -
Midjourney Midjourney is a paid, subscription-based AI image and video generation platform metered by Fast GPU hours, with a web Editor for inpainting and outpainting, a legacy Discord bot, and Stealth mode on higher tiers. -
Adobe Firefly Adobe Firefly is a generative AI platform built around commercially safe, licensed training data and deep integration with Adobe’s creative applications, offering image, video, audio, and partner-model generation through a credit-based subscription system. -
Canva Canva is a template-driven visual design platform whose AI layer—Canva AI 2.0—is a paid-plan feature for conversational editing inside the design editor, metered by a pooled monthly allowance that Free users can access only in a limited, earlier-generation form.