Stable Diffusion 3 API guide to SD3.5 migration and requests

Stable Diffusion 3 API requests using sd3-large, sd3-large-turbo, or sd3-medium are rerouted to their SD3.5 equivalents under Stability AI's April 17, 2025 migration. For a new integration, select a current model such as sd3.5-large and send an authenticated multipart request to /v2beta/stable-image/generate/sd3. Stability AI launched the original SD3 and SD3 Turbo API on April 17, 2024. Migration notice, API reference, launch

What are the key facts about the Stable Diffusion 3 API?

What does the current SD3.5 API support?

The launch emphasized text understanding, typography, and complex prompt following through the Multimodal Diffusion Transformer architecture.

These describe the original SD3 release; they should not be treated as measured results for a current application. Launch explanation

The current service supports text-to-image and image-to-image requests. It exposes choices such as model, output format, aspect ratio, seed, and negative prompt in its documented request schema. Current API

A hosted request is useful when you want to evaluate image generation without first selecting and operating local inference hardware.

Keep your initial application small: submit a prompt, handle the response, and display a reviewable result.

Use a real creative brief for that first integration. For example, ask for a ceramic desk lamp against a pale wall with room for a heading. Review whether the result would actually fit your page layout.

Record the request fields with each image. This creates an understandable experiment log and makes it easier to explain which model and settings were used for an accepted design.

What changes when SD3 API calls migrate to SD3.5?

Stability's migration notice maps sd3-large to sd3.5-large, sd3-large-turbo to sd3.5-large-turbo, and sd3-medium to sd3.5-medium. An old model string therefore does not preserve original SD3 behavior. Deprecation notice

If you are comparing against saved SD3 outputs, label those results with their actual generation date and access route. Reusing the old request body is not evidence that you have repeated the original experiment.

API authentication, request validation, and moderation are part of the service.

The reference documents unsuccessful responses as well as successful image output; handle those paths before connecting generation to an automatic publishing step. Response reference

The launch announcement is not a source for present-day latency or prices. Consult current service documentation and measure observed request time for your workload before setting a user-facing expectation.

Stability publishes current charges in credits per successful generation. Do not substitute a price from another provider or a model-weight license into the API cost estimate. Service charges

How do you call the SD3.5 API and estimate its cost?

1. Sign in to the Stability AI Developer Platform, obtain an API key, and review the current generation reference. Store the key in an environment variable for your client. Authentication reference

2. Choose an explicit current model value, such as sd3.5-large. Keep the model name in your configuration and experiment log so the endpoint path does not become its only identifier. Model schema

3. Submit a multipart request to the documented endpoint. Request image bytes if you want to write a file directly, or JSON if your application expects an encoded response. Request format
   

curl --fail-with-body \
  https://api.stability.ai/v2beta/stable-image/generate/sd3 \
  -H "Authorization: Bearer $STABILITY_API_KEY" \
  -H "Accept: image/*" \
  -F 'model=sd3.5-large' \
  -F 'prompt=A ceramic desk lamp against a pale wall, soft window light' \
  -F 'output_format=png' \
  -o lamp.png

This is a current SD3.5 request through the documented /sd3 path. The client sets the multipart boundary; the API reference cautions against manually supplying an incomplete multipart content-type header. Endpoint instructions

Treat a nonzero curl exit as a failed generation and inspect the error response. Do not assume that a file named lamp.png contains an image when the request failed.

For a production client, check the HTTP status, content type, and finish-reason response header. The schema includes CONTENT_FILTERED for a blurred output even in a successful HTTP response; JSON responses carry this value in finish_reason. Keep that result separate from an accepted asset. Response schema

For image-to-image, the reference requires the prompt, input image, strength, and mode=image-to-image. Test that route separately with an image you are authorized to use before exposing it in a general upload flow. Image-to-image schema

As of September 6, 2026, the reference lists SD3.5 Large at 6.5 credits per successful generation. Review the current model-specific charge before estimating a batch; the endpoint serves more than one model option. Published charges

Build the estimate around attempts, not only final selections. Record generated candidates, accepted images, and any additional editing operations in separate columns so the budget reflects your actual review process.

To assess a migration, rerun a fixed set of representative briefs and inspect the resulting images. Keep the prior assets and notes intact; decide whether the new outputs still satisfy the product's requirements.

Use the AI model timeline for release context, and retain the dates from Stability's notice in technical documentation for a migrated integration.

How does the Stability API compare with self-hosted weights?

The SD3 Medium licensing sibling explains weight access. The ComfyUI pillar covers organizing a local workflow.

For an AWS deployment route, see the SD3.5 Bedrock sibling. Treat each provider's model identifier, authentication, request schema, and billing as its own integration contract.

What else should you know about SD3 API migration?

Do old SD3 API model names still select the original models?

Stability's SD3 API migration reroutes sd3-large, sd3-large-turbo, and sd3-medium to their matching SD3.5 model names. The transition took effect on April 17, 2025. Release notes

Which endpoint do I use for SD3.5 image generation?

Stability documents /v2beta/stable-image/generate/sd3 for SD3.5 generation. Send an authenticated multipart POST request and set the model field explicitly, such as sd3.5-large. API reference

What does SD3.5 Large cost through Stability's API?

Stability's published API rate for SD3.5 Large is 6.5 credits per successful generation as of September 6, 2026. The reference states that failed generations are not charged. Published charges

Does an SD3 API key permit weight redistribution?

An SD3 API key authenticates requests to Stability's hosted service. Downloading or redistributing SD3 Medium weights follows the separate repository conditions and Community License. API, weights

Sources

• Stable Diffusion 3 API launch announcement
• Stability AI developer release notes
• Stability AI current API reference
• Stability AI live OpenAPI schema
• Stable Diffusion 3 Medium model card

Related guides on PromptZone

• AI Model Releases Timeline