GPT Image 1 and 1.5: OpenAI's First Autoregressive Models

This page covers GPT Image 1 and GPT Image 1.5, the first generation of OpenAI's autoregressive image models. Both are deprecated but remain available until their scheduled shutdown dates: October 23, 2026 for gpt-image-1 and December 1, 2026 for gpt-image-1.5. For the current-generation model, see ChatGPT Image 2. For the full family history and architecture overview, see the ChatGPT Image generation guide.

GPT Image 1 was the model that changed what OpenAI's image generation was capable of. Not because it looked better than DALL-E 3 on every output (it didn't, for photographic subjects where diffusion models excel) but because it was architecturally different in a way that mattered. It was the first OpenAI image model built on the same autoregressive transformer backbone as GPT-4o, treating image generation as a token prediction problem rather than a noise reversal problem. That structural choice closed the text rendering gap that diffusion models could never resolve. GPT Image 1.5, released nine months later, refined this foundation with faster generation and precision editing capabilities before GPT Image 2 became OpenAI's recommended successor for image-model migrations.

---

GPT Image 1: The Architectural Turning Point

Developer: OpenAI · API model IDs: gpt-image-1, gpt-image-1-mini · ChatGPT release: March 25, 2025 · API release: April 2025 · Scheduled shutdown: October 23, 2026 (gpt-image-1) and December 1, 2026 (gpt-image-1-mini)

Launch Context

GPT Image 1 launched inside ChatGPT on March 25, 2025, initially as the image generation capability of GPT-4o. It was not announced with the same fanfare as DALL-E 3's launch in 2023; OpenAI positioned it as an improvement to the existing ChatGPT image tooling rather than a standalone product launch. The cultural reaction, however, was immediate. Users rapidly discovered that the model could apply a thorough Studio Ghibli style transformation to portrait photos while preserving the subject's identity far more accurately than DALL-E 3 had managed. The resulting wave of Ghibli-style portraits briefly overwhelmed OpenAI's servers within the same afternoon.

API access followed approximately three weeks later, in April 2025, when gpt-image-1 became available via the OpenAI API for developers. A lighter variant, gpt-image-1-mini, was released alongside the main model as a faster, lower-cost option with reduced output quality: suited for latency-sensitive applications that could accept a quality trade-off.

Architecture: What Makes It Different from DALL-E 3

GPT Image 1 uses a unified transformer backbone that processes text and image tokens natively in the same representational space. This is fundamentally different from DALL-E 2 and DALL-E 3, which use separate systems (a CLIP-based semantic encoder for text and a diffusion model for image generation) that must be aligned through intermediate embeddings. GPT Image 1 was the first commercially deployed OpenAI image model to adopt this architecture, and it was built directly on the same GPT-4o backbone used for language and reasoning tasks.

The practical implication is that text appearing inside generated images is treated as a linguistic prediction problem: the model predicts which character tokens form the correct word sequence, then renders those tokens as image content. A diffusion model has no such concept; it predicts pixel intensities that statistically resemble characters. That structural difference explains why GPT Image 1 immediately produced more accurate text inside generated images than DALL-E 3, despite DALL-E 3's sophisticated recaptioning methodology. For the full architecture explanation covering diffusion, next-scale prediction, and the complete generational comparison, see the ChatGPT Image generation guide.

GPT Image 1 Capabilities

Text rendering inside images. The most consequential improvement over DALL-E 3. Character-level accuracy reached approximately 90–95% for Latin scripts, with meaningful improvement for common non-Latin scripts. For the first time, brand names, short headlines, and infographic copy could be generated with reasonable reliability without a post-processing correction step.

Native multimodal input. GPT Image 1 accepts both text and image inputs, enabling image-to-image transformation, style transfer from a reference image, and multi-turn editing within a conversation. You can provide a reference image and ask the model to "apply this style to a portrait of a coffee shop interior" as a single natural-language request.

Instruction following on complex prompts. Multi-part prompts specifying composition, subject attributes, spatial relationships, and style simultaneously are handled more reliably than DALL-E 3, where complex prompts often resulted in the model satisfying some conditions at the expense of others.

Integrated conversation editing. Within the GPT-4o chat context, generate an image, then request modifications in natural language. The model applies changes while preserving surrounding context. Early editing capability was less precise than GPT Image 1.5's later surgical editing, often resulting in larger-than-intended changes to unaffected image areas.

Style range. The model covers photorealistic, anime, illustration, oil painting, 3D-rendered, abstract, and mixed-media styles within a single endpoint. Style is specified in the prompt, not through a separate model parameter.

GPT Image 1 Specifications

Sources: GPT Image 1 model reference · GPT Image 1.5 model reference · OpenAI deprecation schedule

GPT Image 1 API Pricing

GPT Image 1 used a hybrid pricing model: flat per-image rates when calling the standard image generation endpoints, and token-based rates when using the Responses API for multi-turn contexts.

Per-image pricing (standard endpoints):

Token-based pricing (Responses API):

For multi-turn editing workflows using the Responses API, remember that input images consume tokens at the image input rate on each request. The cost covers not only the output generation but also the input images included in the conversation context.

---

GPT Image 1.5: Surgical Editing and Speed

Developer: OpenAI · API model ID: gpt-image-1.5 · Released: December 16, 2025 · Scheduled shutdown: December 1, 2026

What Changed

GPT Image 1.5 addressed the two most significant complaints about GPT Image 1 in production workflows: slow generation speed and imprecise editing that modified more of the image than intended.

Generation speed: 4× faster. GPT Image 1.5 reduced generation time to roughly one-quarter of GPT Image 1's latency at equivalent quality settings. For interactive workflows where users are iterating through prompt variations in real time, this difference materially changes the feedback loop. What previously took 20–40 seconds per generation dropped to 5–10 seconds.

Surgical region-aware editing. The most architecturally significant improvement in GPT Image 1.5 was its editing precision. GPT Image 1's editing could change the background when asked to adjust the foreground subject, or shift the lighting on a face when asked to change the background color. GPT Image 1.5 introduced region-aware editing that isolates changes to the specified area while preserving facial geometry, lighting, composition, and brand identity in unmodified regions. This "surgical" editing makes the model practical for brand asset workflows where a small change (swapping a background, adjusting a color, or correcting a label) cannot disturb the rest of the image.

Text rendering refinements. GPT Image 1.5 extended GPT Image 1's text accuracy gains with better handling of dense typography: infographics with multiple text layers, poster designs with headline and body copy, UI mockups with navigation and label text. Character-level accuracy stayed in the 90–95% range but degraded less frequently on complex multi-text compositions.

Stronger prompt adherence on multi-part instructions. Long prompts combining style, subject, composition, lighting, and text specifications were more completely satisfied by 1.5 than 1.0, with fewer cases of the model satisfying some conditions while dropping others.

Price reduction. API pricing for GPT Image 1.5 was approximately 20% lower than equivalent-quality GPT Image 1 calls. This made high-volume production workflows meaningfully more economical. Both models are deprecated, so treat historical pricing as migration context rather than a new-workflow recommendation.

GPT Image 1.5 Specifications

---

GPT Image 1 Mini: The Lighter Variant

gpt-image-1-mini launched alongside gpt-image-1 on March 25, 2025, as a faster, lower-cost variant targeting latency-sensitive applications. It used the same autoregressive architecture as the main model but with reduced model capacity, trading output quality for generation speed and lower API cost.

Its primary use case was applications where image generation needed to be near-real-time and visual quality was secondary: quick thumbnails, concept sketches, interactive generation interfaces. For any workflow where image quality or text rendering accuracy mattered, the main gpt-image-1 model was the correct choice. gpt-image-1-mini is deprecated and scheduled to shut down on December 1, 2026.

---

What GPT Image 1.x vs DALL-E 3 Actually Changed

The architectural shift from DALL-E 3 to GPT Image 1 is often described in terms of "better text rendering" without explaining what changed in practice for the workflows that depended on it.

Before GPT Image 1 (DALL-E 3 era): Generating a logo, infographic, social post with copy, or any image where readable text is required meant: generate with DALL-E 3, download the result, open Photoshop or Figma, manually correct every misspelled or garbled word, re-export. For a five-word headline, this was a minor correction. For a ten-panel infographic with 40 text elements, it was a substantial post-production task that negated much of the generation time saving.

With GPT Image 1 and 1.5: For most text-in-image use cases, the first-pass output is usable without a correction step. Brand names with standard Latin-script spelling are correct. Short headlines work. Multilingual copy in Japanese, Korean, and Chinese (which DALL-E 3 consistently garbled) became meaningful outputs rather than approximate decorations.

The practical result was that an entirely new category of workflows became viable without a human correction loop: social media graphics with copy, product packaging mockups, multilingual content for Asian markets, UI mockup screenshots with realistic interface text. This matters more for understanding what GPT Image 1.x represented than any individual benchmark number.

---

Migrating to GPT Image 2

Plan migration before the published shutdown dates: gpt-image-1 is scheduled to shut down on October 23, 2026, while gpt-image-1.5 is scheduled to shut down on December 1, 2026. Here is a parameter-level comparison with gpt-image-2:

Pricing model change — plan this carefully. GPT Image 1.x per-image flat pricing made cost forecasting straightforward. gpt-image-2's token-based pricing means editing workflows that include input images as context consume tokens at the image input rate ($8/M) before any generation. For a workflow that previously cost $0.042 per medium-quality 1024×1024 image under gpt-image-1, the equivalent gpt-image-2 cost will depend on how many input image tokens are included in each request. Model the token usage against your existing workflow before migrating production budget allocations.

API endpoint compatibility. The /v1/images/generations and /v1/images/edits endpoint structures are compatible between versions. The primary code change is updating the model parameter. If you use the Responses API for multi-turn context, the endpoint structure is unchanged.

For the full gpt-image-2 specification including Reasoning Mode, output specs, and rate limits, see ChatGPT Image 2: complete model reference.

---

Frequently Asked Questions

Can I still call the gpt-image-1 API endpoint?

Yes, while the model remains available. OpenAI lists gpt-image-1 for shutdown on October 23, 2026, and gpt-image-1.5 and gpt-image-1-mini for shutdown on December 1, 2026. Migrate and validate before those dates rather than waiting for an endpoint failure.

What is the difference between GPT Image 1 and GPT Image 1.5?

GPT Image 1.5 is a speed and editing precision update on the same autoregressive foundation. Generation speed increased approximately 4×. The editing system gained region-aware precision that preserved facial geometry, lighting, and composition in areas the user did not intend to change. Text rendering on dense multi-layer compositions improved. API pricing dropped approximately 20%.

What is gpt-image-1-mini and why does it exist?

gpt-image-1-mini was a reduced-capacity variant of GPT Image 1 designed for latency-sensitive applications. It used the same architecture but with lower capacity, trading output quality for faster generation and lower cost. It is deprecated and scheduled to shut down on December 1, 2026.

How does GPT Image 1.x pricing compare to GPT Image 2?

GPT Image 1.x used flat per-image pricing based on quality tier and resolution: $0.011–$0.250 per image depending on settings. GPT Image 2 uses token-based pricing: image input at $8/M tokens, image output at $30/M tokens, cached image input at $2/M tokens. For generation-only workflows without large input image context, the per-output cost of gpt-image-2 may be comparable to or lower than gpt-image-1 at medium quality. For editing workflows that include full input images in every request, model the token costs carefully before migrating budget assumptions.

Was the Ghibli viral trend driven by GPT Image 1?

Yes. The Studio Ghibli portrait conversion that went viral in March 2025 ran on gpt-image-1, which had just launched inside ChatGPT. The model's ability to apply strong artistic style transformations to uploaded portrait photos while preserving subject identity drove the viral moment. The same capability exists in gpt-image-2 with higher quality.

Why was GPT Image 1.5 released only nine months after GPT Image 1?

OpenAI's release pace for image models in 2025 was rapid relative to the DALL-E era. GPT Image 1.5 directly addressed two specific production pain points (generation speed and editing precision) that were reported by enterprise and API users in the months after GPT Image 1's launch. The nine-month cycle also reflects competitive pressure: Midjourney V7 and other model releases during that period raised the quality bar for the category.

---