GPT-5 Image: Text Reasoning and Image Output in One Call

Image generation APIs usually work like a vending machine: send a prompt, get a picture. GPT-5 Image is built for a different shape of request, where the model reads text, images and files, reasons about what is wanted, and returns text and images together. That is useful in some workflows and an expensive detour in others.

A pairing behind one model ID

GPT-5 Image is an OpenRouter listing, openai/gpt-5-image, available since October 2025. Its catalogue description says it combines OpenAI's GPT-5 with image generation. It is not a model OpenAI documents under that name; OpenAI's own API exposes GPT-5 as a text model and offers image generation through separate GPT Image models, either through the Images API or as an image_generation tool in the Responses API.

The listing does not name which image model sits behind it. Its sibling, GPT-5 Image Mini, explicitly pairs GPT-5 Mini with GPT Image 1 Mini, which makes gpt-image-1 the obvious candidate for the full version, but that is an inference rather than something the listing states.

The capability profile from the catalogue:

• Input: text, images and files.
• Output: text and images.
• Context: 400,000 tokens, matching GPT-5.
• Features: reasoning, JSON mode and structured outputs. Tool calling is not listed.

Two ways to call it

OpenRouter documents two routes. Through the chat completions endpoint you get the conversational behaviour: send messages, optionally with images attached, and receive text plus generated images in the response, with streaming supported. Through OpenRouter's images endpoint you get a more conventional image API, with text-to-image and image-to-image generation using up to sixteen reference images, and parameters for aspect_ratio, quality, background, output_compression and n from one to ten.

The chat route is the reason to pick this model at all. It lets the language model interpret a loose brief, look at reference material, and decide what to draw, then explain what it produced. If you only need the images route, a dedicated image model gives you the same kind of output without paying for the reasoning layer.

Pricing you cannot read off a token table

The OpenRouter catalogue lists GPT-5 Image at $10 per million input tokens and $10 per million output tokens at the time of writing. Treat those figures with care. Text, image input and image output are different kinds of tokens, and a generated image consumes far more tokens than a paragraph of text, so the headline rate tells you little about the cost of one picture.

For a sense of scale, OpenAI's own per-image prices for gpt-image-1 range from about $0.011 for a low-quality square image to $0.167 for a high-quality square and $0.25 for high-quality portrait or landscape. OpenRouter's combined model will not match those numbers exactly, but they show how much quality settings move the bill.

The practical approach is to measure. OpenRouter returns a usage.cost field in US dollars on each response, so generate fifty representative images at each quality setting you plan to use and record the actual cost. That figure, not the per-million rate, belongs in your budget.

Where the combination earns its cost

The reasoning layer helps when the brief needs interpretation:

• Iterating on a design in conversation, where each turn refers back to earlier images and the model needs to keep track of what changed.
• Diagrams and explanatory images that must reflect a document the model has read, such as turning a spec into an illustrated overview.
• Mixed deliverables where you want a written explanation and an accompanying image from one request, with structured output describing what was produced.
• Image editing driven by intent, where a user says what they want changed in plain language and the model works out how to apply it to a reference image.

When a dedicated image model is better

For bulk generation from well-specified prompts, such as product shots from a template or variations of an existing asset, the language model adds cost and latency without improving the image. Call an image model directly.

The generation also matters. OpenAI's documentation now describes gpt-image-1 as a previous image model and lists GPT-Image-2.5 variants as current, including one positioned as its most capable for generation and editing. If image quality is the priority rather than the conversational workflow, compare against newer options such as GPT-5.4 Image 2 before committing. For a broader view of where image models fit, multimodal models explained covers how images become tokens on the way in.

The base model is ageing too. GPT-5's knowledge cutoff is 30 September 2024, so prompts that depend on recent products, interfaces or events need that context supplied in the request.

How to evaluate it for your use case

1. Write ten briefs that reflect real requests, including at least three that need the model to interpret a document or a reference image.
2. Run each through GPT-5 Image on the chat route and through a dedicated image model with a hand-written prompt.
3. Score the outputs blind on whether they match the brief, not just whether they look good.
4. Record usage.cost and latency for every run.

If GPT-5 Image wins mainly on the briefs that need interpretation, use it for those and route the rest to a cheaper image model. If it does not win anywhere, the reasoning layer is not doing work you need.