Grok Imagine API: Models, Pricing, and Image 2.0 Status

The Grok Imagine API is live and metered, and it currently exposes four models: two for images and two for video. What it does not expose, as of today, is Imagine Image 2.0. xAI announced that model on August 7, 2026 and shipped it to the consumer apps the same day, with the sentence developers care about buried in the post: "API access is coming soon." So if you came here to wire 2.0 into a pipeline this afternoon, the honest answer is that you cannot yet.

This is the developer view: which model ids exist right now, what each one actually costs including the input charges people miss, the rate limits and regions, and what changes when 2.0 lands. For the announcement itself and what Image 2.0 does, see the Imagine Image 2.0 launch post.

Is the Grok Imagine Image 2.0 API available yet?

No. Three independent checks say the same thing today:

• xAI's own announcement says so. The launch post ships Image 2.0 as Quality Mode on grok.com/imagine and the iOS and Android apps, and states that API access is coming soon.
• There is no model page for it. Every callable xAI model has a page under docs.x.ai/developers/models/. The URL for grok-imagine-image-2 returns a 404.
• It is absent from the pricing table. The xAI API pricing page lists grok-imagine-image, grok-imagine-image-quality, and grok-imagine-video, and no Image 2.0 row.

Full size

Worth naming the confusion this creates: xAI already uses the phrase "Quality Mode" for two different things. In the consumer app, Quality Mode now means Image 2.0. On the API, grok-imagine-image-quality is the older high-quality tier that shipped in April 2026. Same words, different model. If you benchmark the app against the API right now and the app looks noticeably better at typography and small text, that is not your prompt, that is a model you cannot call yet.

Practical planning advice: do not architect around a specific 2.0 model id before it exists. When xAI ships it, it will appear as a new entry in the models list with its own price, and anything already routing by model key picks it up as a one-line change. Hard-coded model strings scattered through a codebase are the thing that makes launch day expensive.

Grok Imagine API pricing, per model

xAI prices images per image and video per second, and bills inputs separately from outputs. That second part is the line item that surprises people running edit-heavy workloads, because an edit pays for both the image you send and the image you get back. Current rates from xAI's pricing page, as of August 7, 2026:

A few readings of that table that matter more than the numbers themselves.

The standard image model is absurdly cheap for exploration. Two cents per output image, and the API lets you "configure output count (up to 10 images per request)". Ten variations of one prompt costs twenty cents and one round trip. That is the single best argument for the API over the consumer app, where you get a subscription and an unpublished daily cap instead of a meter.

The quality tier is 2.5x the standard tier at 1K and 3.5x at 2K. That gap is large enough that routing every job to quality is a real budget decision, not a rounding error. Use the standard model for iteration and reserve the quality tier for the render you ship.

The two video models are not a simple old-and-new pair. grok-imagine-video-1.5 is the newer and more expensive one at $0.080 per second, and it takes text and image input. But only grok-imagine-video accepts video as an input modality, which is what reference-to-video and video extension need. Picking the higher version number is not automatically picking the more capable model for your job.

Grok Imagine API rate limits and regions

xAI publishes per-model rate limits on each model page, and they split cleanly by media type. Both image models are capped at 5 requests per second. Both video models allow 10 requests per second, which matters less than it sounds because video generation is asynchronous: you POST the job and poll for the result rather than holding a connection open.

Regions differ per model, and this is easy to miss. Per xAI's model pages, grok-imagine-image and grok-imagine-video are available in us-east-1, us-west-2, and us-saltlake-2, while grok-imagine-image-quality and grok-imagine-video-1.5 are listed only in us-east-1 and us-west-2. If you have a region pinned for latency or data-residency reasons, check it before you assume a model swap is a one-word change.

What the Grok Imagine API actually exposes

Four endpoints, all under https://api.x.ai/v1:

• POST /images/generations for text to image, with configurable output count, aspect ratio, resolution, and a URL or base64 response format.
• POST /images/edits for editing, which per xAI's docs supports "up to 3 source images in a single request for combining subjects, transferring styles, and composing scenes". Source images go in as a public URL or a base64 data URI.
• POST /videos/generations for video, with duration configurable up to 15 seconds. It returns a request id.
• GET /videos/{REQUEST_ID} to poll that job until it finishes.

On billing behavior, xAI is explicit that "image edits are billed for both the input image and the generated output image." An edit-heavy loop therefore costs roughly 10 percent more per call than the headline output price on the standard model, and about 20 percent more on the quality tier at 1K. Budget accordingly if your product is an editor rather than a generator.

Getting a Grok Imagine API key

Keys come from the xAI developer console, and the client story is deliberately boring: the API is compatible with the OpenAI SDK, so you point an existing OpenAI client at https://api.x.ai/v1 with your xAI key and most of your code stays put. xAI also ships its own xai_sdk Python package, and Grok models are resold through Azure AI Foundry, Oracle OCI Generative AI, and Google Cloud Vertex AI Model Garden if procurement prefers an existing vendor relationship.

The parts that are genuinely on you: key storage and rotation, retry and backoff against those per-second limits, polling loops for video, moving returned URLs into your own storage before they expire, and validating that what came back matches what you asked for. None of that is hard. All of it is code you write once per integration and then maintain forever.

Using Grok Imagine without wiring the API yourself

If the goal is a product feature rather than a direct provider relationship, there is a shorter path. Dream Pixel Forge runs Grok Imagine as one adapter behind a provider router, so generation is reachable from an agent or a script without an xAI key, a polling loop, or a storage story. grok-imagine-image is 2 credits per image with up to three input images for editing and seven aspect ratios, and grok-imagine-video is 24 credits for a 6-second 480p clip. When a call fails, the router falls back to another model instead of handing your agent an error to reason about.

Two surfaces, one account and one credit balance. The MCP endpoint at /api/mcp gives any MCP-capable client (Claude, Codex, or your own agent) tools like generate_image, edit_image, and generate_video directly. The dpf CLI is a zero-dependency npm package (npx dreampixel-cli login) that wraps the identical calls over HTTPS for anything that would rather shell out than speak MCP. The login flow mints the key on your machine and sends only its SHA-256 hash to the server, so a leaked server log yields a hash rather than a usable credential. The details are in the CLI guide and the agent API post.

To be precise about what runs today: the Grok model in production here is grok-imagine-image, not Image 2.0. Nobody has 2.0 on an API right now. It goes into the router when xAI ships the endpoint, and existing calls pick it up as a model key rather than a rewrite.

When the raw API is still the right call

Plenty of cases, and it is worth being straight about them. Go direct to xAI when you need the exact per-call cost on your own invoice, when you want the 10-images-per-request batching for a research or evaluation loop, when you need a specific model id and region pinned for compliance, or when generation is your core product and the routing logic is a thing you want to own. Two cents an image is hard to beat when you are prepared to build everything around it.

Go through a routed surface when generation is a feature inside a larger product, when an agent is the caller and you would rather it get a validated result than an image URL and a retry policy, or when you want to A/B Grok against Nano Banana and GPT Image without maintaining three integrations. Model comparisons are in the Grok Imagine prompt guide, which also covers the reference aspect-ratio trap that quietly distorts edit inputs, and the Grok image editing guide for the multi-image edit workflow specifically.

The short version

Four Grok Imagine models are callable today: grok-imagine-image at $0.02 per output image, grok-imagine-image-quality at $0.05 to $0.07, grok-imagine-video at $0.05 to $0.07 per second, and grok-imagine-video-1.5 at $0.080 per second. Images are limited to 5 requests per second, video to 10. Edits bill both directions. Imagine Image 2.0 is not among them, xAI says API access is coming soon, and there is no model page or price for it yet. Build against model keys rather than hard-coded strings and launch day is a config change.