GPT Image 2 on Kubeez: API, App, and MCP (Complete Guide)

GPT Image 2 is OpenAI’s next-generation class of image generation: strong prompt adherence, photoreal and illustrative styles, and support for text-to-image and image-to-image with up to 16 reference images. On Kubeez, you can run it the same way you use other flagship models: in the web app, over the REST API, or through the Kubeez Model Context Protocol (MCP) server—one credit balance, one account, and no separate vendor API keys for the creative stack.

#What “GPT Image 2” means on Kubeez

Kubeez wires provider-class image models behind a unified job system: you pick a model_id, send a prompt (and optional reference image URLs), and receive permanent HTTPS URLs on Kubeez’s CDN when the job finishes.

For GPT Image 2:

This is different from GPT Image 1.5 on Kubeez (gpt-1.5-image-medium / gpt-1.5-image-high), which exposes fixed aspect sets (e.g. 1:1, 2:3, 3:2) and a quality axis. If you need explicit 16:9/9:16 control, pick another family (e.g. Nano Banana 2, Imagen 4); if you want OpenAI-style reasoning in the image model and long prompts, GPT Image 2 is the right fit.

#Where you can run GPT Image 2

#1. Web app (Media / Images)

1. Open Media generation (or your locale equivalent).
2. Select GPT Image 2 in the model list.
3. Text only → text-to-image. Attach images (or paste from uploads) → the app switches to the edit path and passes references—same behavior the API encodes with source_media_urls.

The UI enforces the same limits and credits as the API; the difference is you get previews, history, and upload helpers without writing JSON.

#2. REST API (servers, scripts, CI)

Kubeez’s public REST base is https://api.kubeez.com/v1 (see interactive docs at the same host’s /docs and the site’s REST API overview).

Minimum flow for GPT Image 2:

1. Authenticate with X-API-Key: sk_live_… (or Authorization: Bearer—see your key type).
2. Discover models (recommended): GET /v1/models?model_type=image and confirm gpt-image-2 appears with scopes you need.
3. Start a job: POST /v1/generate/media with:
   • model: "gpt-image-2"
   • prompt: your text
   • generation_type: "text-to-image" or "image-to-image"
   • For edits: source_media_urls: array of https URLs the backend can fetch (upload first via POST /v1/upload/media if files are only on disk).
4. Poll GET /v1/generate/media/{id} until status is completed and outputs contain CDN URLs—not vendor temp links. (MCP: use get_generation_status instead of long-lived REST streams.)

Scopes: generate:media on the API key. For balance and job listing, the matching read scopes on your key (see your project’s API Keys and docs).

Why teams pick REST: long-lived keys in a secret manager, idempotent server workers, and tight integration with your own DAM or CMS.

#3. MCP (Cursor, ChatGPT-style hosts, agents)

Kubeez hosts an MCP server at:

https://mcp.kubeez.com/mcp

Typical tool sequence for GPT Image 2:

1. get_models with model_type: "image" — confirm gpt-image-2 and read usage_notes / cost_per_generation.
2. (Optional) get_balance before large batches.
3. generate_media with model: "gpt-image-2" and prompt (add source_media_urls for image-to-image).
4. Wait roughly estimated_time_seconds, then get_generation_status on the returned generation_id every ~5 seconds until completed or failed.

Scopes: the MCP key or OAuth connection must include generate:media (and read:generations to poll, read:balance if you call get_balance).

Local / private files: the MCP (and API) cannot read the user’s disk. For reference images, use get_upload_url → user uploads in browser → get_upload_session → pass URLs to source_media_urls—documented in the Kubeez MCP tool descriptions and MCP overview.

Why teams pick MCP: the assistant plans steps, rewrites prompts, and runs tools in a loop—ideal for “three variants, same brief” or “check balance, then generate.”

#Request shape: what to send (and what to skip)

• Model id — always use the string returned by get_models (gpt-image-2 on Kubeez).
• Do not rely on aspect_ratio for GPT Image 2: capabilities show an empty aspect_ratio_options list. Prefer describing composition (“wide banner feel”, “tall mobile poster”) in natural language, or use another model if you need strict 16:9/9:16 at the API.
• Image-to-image — set generation_type to image-to-image when you pass references, or let the platform infer from source_media_urls (REST/MCP may auto-adjust—still pass URLs in the shape your client expects).
• Long briefs — product legal copy, shot lists, and multi-paragraph art direction are allowed within the prompt_max_chars budget; trim only if you exceed the cap.

#Pricing and comparison (practical, not a price list)

• GPT Image 2 is billed as a flat cost per output image on Kubeez (see cost_per_generation in get_models / GET /v1/models). It is not duration-based.
• For rigid aspect ratios and marketing crops, you may want Nano Banana 2 or Seedream alongside GPT Image 2 in your toolkit; for open-ended OpenAI-style stills and heavy prompting, GPT Image 2 is the natural choice.

#Operations: polling, rate limits, and failures

• Polling: do not hammer status every second. Use estimated_time_seconds for a sensible first delay, then ~5s intervals until complete.
• Failures: on failed, read error_message on the generation record; credits are handled according to Kubeez’s standard refund rules for that job type (see account and support docs if you hit edge cases).
• CDN: completed jobs return Kubeez CDN URLs suitable for long-lived links in production (not time-limited vendor URLs).

#Security and content policy

• Use API keys and MCP connections the same way you would any production secret: rotate, scope minimally, and never commit keys to git.
• Generated content is still subject to platform and provider policies; for brand work, add human review for copyright, likeness, and campaign compliance.

#Links that stay current

• Available models (docs) — canonical list; IDs and availability can change.
• MCP overview and MCP settings — connect Cursor or PAT-based clients.
• API overview and API keys — REST for backends.

Bottom line: GPT Image 2 on Kubeez is a single model_id (gpt-image-2) across the app, REST API, and MCP. Call get_models first, run text-to-image or image-to-image with up to 16 references, poll until CDN-ready URLs appear, and keep one credit wallet for every integration path you use.