How to Use GPT Image 2: Text Rendering and Thinking Mode

OpenAI announced ChatGPT Images 2.0 on April 21, 2026, powered by the new gpt-image-2 model, and began rolling it out the next day^[1]. The headline change is not resolution. It is that generated text inside images now works.

Not "mostly works, with a handful of charming typos." Text at roughly 99% accuracy across Latin, Japanese, Korean, Hindi, Bengali, Arabic, and a dozen other scripts^[1]. For two years the standard workflow was to generate the image, then composite real text on top in Figma. That step is largely gone.

Learning how to use GPT Image 2 well is mostly about two things: writing briefs precise enough to exploit the text rendering, and knowing when to turn Thinking Mode off.

TL;DR

• Text rendering at ~99% accuracy, including non-Latin scripts at near-native quality^[1].
• Thinking Mode lets the model browse the web for references, plan multiple variants, and self-verify its output before returning^[1].
• 2K native resolution with 16:9 support, and a December 2025 knowledge cutoff^[1].
• It replaces DALL-E 3, which retired May 12, 2026, and the interim GPT Image 1.5^[1].
• A completely new architecture, single-pass rather than the prior two-stage pipeline, giving sub-three-second typical generation^[1].
• Not the photorealism leader. Nano Banana 2 stays ahead on pure photographic realism and 4K output; GPT Image 2 leads on text fidelity and prompt adherence^[1].

What GPT Image 2 is

GPT Image 2 is OpenAI's flagship image model, shipping on three surfaces: inside ChatGPT as Images 2.0, inside the Codex coding agent for inline asset generation, and as the gpt-image-2 API^[1].

The architecture is a clean break from the DALL-E line. OpenAI has declined to confirm whether it is diffusion-based, autoregressive, or hybrid, saying only that it is a completely new architecture moving from a two-stage inference pipeline to a single-pass generator. The practical consequences are faster generation and much tighter coupling between prompt semantics and rendered output^[1].

The text rendering breakthrough

The specific improvements are worth separating, because they unlock different workflows^[1]:

• Latin-script spelling at ~99%, including dense menu text, packaging labels, nutrition panels, long editorial headlines, and UI copy.
• Non-Latin scripts at near-native accuracy for major world languages. Previous models treated these as decoration: visually plausible, semantically meaningless.
• Small text and iconography in dense compositions, including form fields, badge labels, and legal fine print, rendered readably.
• Stylistic consistency between text and design language, so a Brooklyn café sign and a Tokyo subway poster read as different designers rather than the same generic model.

The bookstore example is the one to study. Each title is rendered in a distinct Indic script, and each is coherent text a reader of that language can identify^[1]. That is the capability that opens localized content pipelines: regional ad creative, multilingual packaging, Indic-language e-commerce thumbnails, at a speed that was previously impossible.

Thinking Mode

Thinking Mode is the feature that changes how you interact with the model. With standard prompting it does what any image model does: read prompt, generate pixels. With Thinking Mode on, it also^[1]:

Browses the web for reference material when the prompt names real brands, products, or time-sensitive content.

Plans multiple output variants before committing to pixels, which is what makes one prompt produce an ad campaign in three sizes or a comic across five panels.

Self-verifies before returning, checking that rendered text is spelled correctly and that layout hierarchy survived.

The demonstration OpenAI leads with is a single prompt asking for a poster about merch currently on its own website, which produced an accurate composition because the model went and looked. No reference upload, no manual product list^[1].

For briefs containing phrases like "in the style of their current hero page," this removes the manual reference-gathering step. For bulk generation of simple visuals, turn it off, because it costs more and takes longer.

Commercial-grade output

The advertisement above is a fully synthetic single-prompt generation. It carries a plausible product photograph, a coherent brand wordmark in two scripts, secondary typographic detail, a legible announcement with location and social handle, and stylistic consistency between photograph, type, texture, and palette^[1].

Two years ago this was the classic generative-image failure case. The model could produce the photograph, but any real text in the composition devolved into garbled glyphs.

The implication is worth naming directly: for most small-business marketing the bottleneck is no longer visual production. It is the brief.

Deeper structural understanding

OpenAI's positioning line is "image generation with a point of view," meaning the model has structural understanding of what it renders rather than statistical coverage of a corpus. In practice^[1]:

• Instruction following improves on compositional requests: left-aligned headline, product at right, brand mark bottom-right.
• World knowledge is stronger. The model knows which camera a product shot should look like it was taken on, which paper stock a vintage cover should emulate.
• Character consistency holds across multi-panel sequences.
• Iconographic literacy improved, so periodic tables, anatomical diagrams, and architectural plans render correctly rather than as plausible-looking nonsense.

This is the axis where it most clearly differs from purely photographic models. On a photorealistic still life, Nano Banana 2 is competitive or ahead. On a composition requiring the model to know something and render it correctly, GPT Image 2 leads^[1].

Where it does not fit

The honest limitations^[1]:

• Not a replacement for designers on identity-critical brand work. First-draft quality is excellent; a creative director should still direct.
• Not the photorealism leader. Nano Banana 2 and specialist photo-real models retain edges on hyper-realistic still life and portraiture.
• Thinking Mode is slower and pricier. Disable it for bulk generation of simple visuals.
• Content policy applies. Real public figures, copyrighted characters, and certain commercial references are restricted.

How to use GPT Image 2 on reAPI

reAPI exposes the model on an async task endpoint. Submit returns a task_id; poll until ready.

curl https://reapi.ai/api/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2",
    "prompt": "Editorial poster for a Kyoto coffee shop opening, hand-lettered headline in cream serif, warm cinematic light, legible address line at the bottom",
    "size": "16:9",
    "resolution": "2k"
  }'

Three things to know before writing the integration.

Resolution is a priced dimension. The gateway exposes 1K, 2K, and 4K, each with its own rate, so a resolution choice is a cost choice. Current rates are on reapi.ai/models/gpt-image-2, which is the canonical source rather than any figure quoted in an article.

Media inputs are public URLs only. No base64, no data: payloads, on any model on the platform.

Prompt precision pays here more than on older models. Briefs that worked with DALL-E 3, vague and visually focused, underperform relative to what this model does with explicit layout, typographic, and reference instructions. The ceiling is higher and the floor requires more precise input^[1].

Full request and response shapes are in the reapi.ai/docs/gpt-image-2 reference.

FAQ

What is GPT Image 2?

OpenAI's flagship image generation model, announced April 21, 2026, replacing DALL-E 3 and GPT Image 1.5. It ships in ChatGPT as Images 2.0, in Codex, and as the gpt-image-2 API^[1].

How accurate is GPT Image 2 at rendering text?

Roughly 99% on Latin script, with near-native accuracy on major non-Latin scripts including Japanese, Korean, Hindi, Bengali, and Arabic^[1].

What is Thinking Mode?

A mode in which the model browses the web for references, plans multiple variants before generating, and self-verifies rendered text and layout before returning output. It costs more and takes longer, so disable it for bulk simple generation^[1].

What resolution does GPT Image 2 support?

2K native with 16:9 support alongside existing aspect ratios^[1]. On reAPI, 1K, 2K, and 4K are exposed as separately priced options.

Is GPT Image 2 better than Nano Banana 2?

They lead on different axes. Nano Banana 2 is ahead on pure photographic realism and 4K output. GPT Image 2 leads on text fidelity, prompt adherence, and integrated Thinking Mode^[1].

What happened to DALL-E 3?

It retired on May 12, 2026. GPT Image 2 replaces it and the interim GPT Image 1.5^[1].

Does GPT Image 2 accept base64 image input on reAPI?

No. Every model on the platform takes public http(s) URLs only for media inputs.

How should I write prompts for GPT Image 2?

More specifically than for DALL-E 3. Give explicit layout, typographic, and reference instructions rather than a vague visual description, because the model now acts on structural detail that older models ignored^[1].

Updating the brief, not just the model string

The useful summary of this release is that it moves the bottleneck. When generated text was unusable, the constraint was the model. Now that a café advertisement with readable Japanese secondary type comes out of a single prompt, the constraint is how precisely you can describe what you want.

That has a practical consequence for anyone migrating a pipeline. Prompt templates written for the previous generation are tuned for a model that ignored structural instruction, and they will underuse this one. Rewrite them with explicit layout and typography, keep Thinking Mode for briefs that reference the real world and off for volume, and pick the resolution deliberately since it is a priced dimension. That is how to use GPT Image 2 without paying for capability your briefs never ask for.

References

1. OpenAI. GPT Image 2 — model card, capabilities, and API reference. Retrieved July 2026 from platform.openai.com/docs/models
2. OpenAI. Pricing — per-image and per-token rates by model. Retrieved July 2026 from platform.openai.com/docs/pricing

Further reading

• reAPI. How to use Nano Banana 2 Lite. reapi.ai/blog/how-to-use-nano-banana-2-lite
• reAPI. GPT Image 2 endpoint reference. reapi.ai/docs/gpt-image-2
• reAPI. Model catalog. reapi.ai/models