GPT-5.4 Image 2: One Endpoint That Writes and Draws

Most image models are a single function: prompt in, picture out. GPT-5.4 Image 2 is listed differently. OpenRouter describes it as GPT-5.4 combined with the image generation of GPT Image 2, so one model id can answer in text, produce images, or do both in the same exchange. That is convenient, and it also makes the model easy to misunderstand, because it behaves partly like a chat model and partly like an image API.

This guide is about what that pairing gives you in practice, how to call it, and how to stop the bill being a surprise.

What the listing is, and what it is not

Start with the parts. GPT Image 2 is OpenAI's image generation and editing model, released on 21 April 2026 with the API id gpt-image-2 and the snapshot gpt-image-2-2026-04-21. On OpenAI's own API it is served on the image generation and image edit endpoints, plus Batch, and supports inpainting.

GPT-5.4 is the general flagship released in March 2026. OpenAI's API already lets GPT-5.4 call image generation as a tool inside the Responses API, so the combination is not new as a capability. What the OpenRouter listing adds is a single id, openai/gpt-5.4-image-2, that takes text, images and files as input and can return text and images as output.

Two catalogue details are worth noting. The context window is listed at 272,000 tokens, smaller than the 1.05M window documented for GPT-5.4 on its own, and the listing does not advertise tool calling. If your workflow depends on function calls in the same conversation as image generation, test that before you build on it.

What GPT Image 2 brings to the pairing

The image half is the reason to use this listing at all. Rather than rely on launch claims about it, test the four things that decide whether an image model is usable in production:

• Instruction following under constraints. Write prompts with several hard requirements, such as an exact object count, a fixed layout and a named colour, and score how many of them hold in each output.
• Text inside the image. Labels, posters, diagrams and interface mock-ups are where this matters. Include every script you actually ship, non-Latin ones too, and check spelling at full size.
• Resolution and aspect ratio. Record the pixel dimensions you actually get back for each aspect_ratio and quality setting instead of assuming a maximum.
• Consistency across a set. OpenRouter's image endpoint accepts an n of 1 to 10, so you can ask for several images in one request and check whether style and characters hold across them.

Public preference leaderboards are a reasonable first filter, but rankings move quickly and reflect strangers' prompts rather than yours; Arena Elo explained covers what those ratings do and do not measure.

Two ways to call it

OpenRouter exposes the model through two routes, and choosing the right one saves a lot of awkward parsing.

Chat Completions, at /api/v1/chat/completions, is the conversational route. It accepts the usual options, including reasoning_effort, response_format and streaming. Use it when the model should decide what to draw from a longer conversation, or when you want text and images interleaved: a short brief, then the image, then alt text.

The images endpoint, at /api/v1/images, is the direct route. It takes a prompt plus options such as aspect_ratio, quality, background, n, output_compression, and up to 16 input_references for image-to-image work. It returns base64 image data. Use it for production pipelines that generate assets from a template, where you already know exactly what you want.

Two documented constraints save debugging time: providers that only produce one image at a time reject n greater than 1, and a transparent background needs PNG or WebP output. The model is also not available through OpenRouter's EU in-region routing, which matters if you have data residency requirements.

Budgeting when one model writes and draws

OpenRouter lists the model at $8 per million input tokens and $15 per million output tokens at the time of writing, with cached input at $2. Those token rates do not tell you what a single image costs, because image output is priced on its own terms and thinking adds reasoning tokens on top.

The reliable approach is to read cost from the response, not to estimate it. The images endpoint returns a usage.cost field with each result. Log it with the prompt, the size, the quality setting and n, and after a few hundred generations you will have a real cost table for your own prompts. OpenRouter also documents that failed generations are not billed, so a retry after an upstream failure does not double-charge you.

The habits that keep cost down are predictable: generate at lower quality while iterating and only re-render the final at full quality, ask for one image until the prompt is stable, and reserve thinking for prompts with strict layout or text requirements.

Where it still goes wrong

OpenAI's own image generation guide notes that its GPT Image models can still struggle with precise text placement and clarity. Beyond that, test for the weak spots common to current image models: artefacts in close-cropped photorealistic faces, inexact logo geometry, long blocks of text that degrade, and style that holds within a batch but drifts between separate sessions.

Each has a workaround. Composite real logos in afterwards rather than asking the model to reproduce them. Keep long copy out of the image and set it in HTML or a design tool. Generate a set in one request when you need visual consistency, and keep reference images to feed back in when you need to match it later.

When to pick it over the alternatives

• Pick GPT-5.4 Image 2 when one call should both reason about content and produce the visual, such as turning a spec into a labelled diagram with an explanation.
• Pick plain GPT-5.4 when the work is text or code and an image is rare; you keep the larger window and the full tool set.
• Pick another image model when you need cheaper bulk generation or a different look. The Nano Banana Pro guide covers Google's equivalent, and the trade-offs are worth comparing on your own prompts. OpenAI's image guide now also documents newer GPT Image 2.5 models, worth checking for a direct image pipeline.

Whichever you choose, judge it on twenty prompts from your real backlog, scored blind, with cost logged per image. A leaderboard tells you what strangers preferred; your own set tells you what you will ship.