Image generation pricing per image across providers

When you’re evaluating image generation pricing per image providers, the difference between a $0.040 DALL·E 3 render and a $0.01 SD Turbo step is not just a line-item cost—it changes how you architect retries, caching, and user-facing latency. This post puts OpenAI, Stability AI, Google, and Midjourney side by side on the dimensions that matter to engineers shipping production multimodal features, with real API shapes and hard limits called out.

The Contenders

We compare four reachable paths for programmatic image generation:

OpenAI DALL·E 3 (official REST/SDK)
Stability AI SDXL 1.0 / SD Turbo (Stability REST API)
Google Imagen 2 (Vertex AI public preview/GA)
Midjourney v6 (unofficial API via Discord proxy or third-party gateway)

All numbers reflect public pricing as of early 2024. Image generation pricing per image providers shifts quarterly; verify against provider billing before scaling to millions of calls.

Capabilities

DALL·E 3 ships tight prompt adherence, native 1792×1024 panoramas, and a built-in moderation chain that rejects most unsafe content before generation. It also supports quality:"hd" for finer rendering at 2× cost.

SDXL 1.0 gives a 1024×1024 base with strong style flexibility, open weights, and a separate inpainting model (sd-xl-1.0-inpainting). SD Turbo trades coherence for speed: 1–4 step sampling, usable for storyboards but not hero shots.

Imagen 2 emphasizes photorealistic output, accurate lighting, and integrates with Google Cloud’s safety filters and Vertex Prompt Builder. It supports mask-based editing via imagegeneration@002.

Midjourney v6 leads subjective aesthetic quality and handles complex art-direction prompts, but offers no official API, weak text rendering control, and limited programmatic variate control beyond --seed and --chaos.

Price and Cost Model

The core of image generation pricing per image providers is the per-call fee. OpenAI charges $0.040 for a 1024×1024 DALL·E 3, $0.080 for 1792×1024 or 1024×1792. DALL·E 2 is still available at $0.020 for 1024².

Stability’s hosted SDXL 1.0 runs $0.040 per image; SD Turbo is $0.010 per step (typically 1–4 steps, so $0.01–$0.04). Self-hosting on A100s can drop marginal cost below $0.001 but adds ops overhead, queue management, and model download plumbing.

Google Imagen 2 on Vertex AI lists $0.04 per 1024×1024 image, with resolution tiers at $0.08 (1536²) and $0.16 (2048²). Midjourney’s subscription model obfuscates per-image cost: the $30/mo Standard plan yields ~15 fast hours, roughly $0.02–$0.06 per final grid depending on usage pattern and whether you burn fast or relaxed mode.

When surveying image generation pricing per image providers, note that only OpenAI and Google expose clean per-image metering; Stability uses credit steps; Midjourney is time-metered and throttled by concurrent jobs.

# OpenAI per-image cost is explicit in the response
from openai import OpenAI
client = OpenAI()
img = client.images.generate(
    model="dall-e-3",
    size="1024x1024",
    quality="standard",
    prompt="a rusty robot in a sunflower field"
)
# $0.040 billed exactly once, no token accounting

Latency and Throughput

DALL·E 3 averages 5–10s for a single 1024² image; concurrency is limited by tier (50 req/min on paid, 3 req/min on free). SDXL 1.0 on Stability’s GPU fleet returns in 2–5s; SD Turbo under 1s but may queue under burst. Imagen 2 sits at 3–6s with Vertex’s auto-scaling, and supports batched number_of_images up to 4 per call.

Midjourney queues behind Discord; a /imagine job takes 20–60s plus wait time under load, and concurrent jobs are capped by subscription tier (3 on Basic, 12 on Standard).

If you need >100 images/min sustained, only Stability’s API or self-hosted SDXL realistically hold up without enterprise contracts. Image generation pricing per image providers must be weighed against this throughput ceiling.

Ergonomics and API Design

OpenAI’s SDK is the cleanest:

from openai import OpenAI
client = OpenAI()
r = client.images.generate(
    model="dall-e-3",
    size="1792x1024",
    quality="hd",
    prompt="cinematic wide shot of a neon city"
)
print(r.data[0].url)

Stability requires JSON with engine ID and explicit parameters:

curl -X POST https://api.stability.ai/v1/generation/sd-xl-1.0/text-to-image \
  -H "Authorization: Bearer $STABILITY_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "text_prompts":[{"text":"cinematic wide shot of a neon city"}],
    "height":1024,
    "width":1024,
    "steps":30,
    "cfg_scale":7.5
  }'

Google uses Vertex AI client libraries with region pinning:

from google.cloud import aiplatform
aiplatform.init(project="my-proj", location="us-central1")
model = aiplatform.ImageGenerationModel.from_pretrained("imagegeneration@002")
images = model.generate_images(
    prompt="cinematic wide shot of a neon city",
    number_of_images=1,
    height=1024,
    width=1024
)

Midjourney has no REST; you script a Discord bot or pay a proxy. That breaks type safety, complicates retries, and forces you to parse message embeds for image URLs.

Ecosystem and Tooling

DALL·E 3 is native in LangChain, Vercel AI SDK, and most no-code builders. SDXL has ComfyUI, Automatic1111, and hundreds of LoRAs; you can swap hosted for local with one env var. Imagen 2 lives inside Google’s Vertex ecosystem—good if you already use BigQuery/GCS and need unified IAM. Midjourney output is siloed to Discord unless exported manually.

For routing logic, an OpenAI-compatible gateway (e.g., n4n.ai’s single endpoint covering 240+ models) can proxy DALL·E and SDXL behind one interface, but image-specific cache hints are provider-dependent and fallback must respect each provider’s content policy.

Limits and Quotas

OpenAI caps prompt length at 4000 chars and denies certain entities (public figures, explicit content). Stability enforces 64 max steps on Turbo and 150 on SDXL. Imagen 2 requires Google Cloud trust & safety review for production and blocks unnamed brands. Midjourney bans automated scraping of their Discord and revokes keys on abuse.

Head-to-Head Comparison

Provider	Capabilities	Price per 1024²	Latency	Ergonomics	Ecosystem	Hard Limits
OpenAI DALL·E 3	Best prompt adherence, HD option	$0.040 (HD $0.080)	5–10s	SDK, REST, typed	LangChain, Vercel	50 req/min, content filter
Stability SDXL 1.0	Open weights, inpaint, Turbo	$0.040 (Turbo $0.01–$0.04)	2–5s (Turbo <1s)	JSON/multipart	ComfyUI, A1111	64 steps max Turbo
Google Imagen 2	Photoreal, GCP integ	$0.040	3–6s	Vertex client	GCP native	Trust review required
Midjourney v6	Top aesthetics	~$0.02–$0.06 sub	20–60s+queue	Discord only	None programmatic	No official API

Which to Choose

High-fidelity marketing assets, low volume: DALL·E 3 or Midjourney. If you need API integration and audit logs, DALL·E 3 wins; if designers live in Discord and aesthetic lead matters more than latency, Midjourney’s look may justify the proxy hack.

High-volume, cost-sensitive generation: Stability SDXL Turbo. At $0.01 per step you can batch thousands of thumbnails hourly. Self-host if you already operate GPU fleets; otherwise the hosted API still undercuts DALL·E 3 by 4×.

Enterprise on Google Cloud: Imagen 2. You get unified IAM, logging, data residency, and a predictable $0.04 per image. The Vertex client is verbose but integrates with your existing GCP CI/CD.

Rapid prototyping with one codebase: Use an OpenAI-compatible gateway that fronts multiple backends, then switch models by config. This avoids rewriting the Discord glue when Midjourney changes its terms and lets you A/B image generation pricing per image providers without forked client code.

Image generation pricing per image providers is not a single ranking—it’s a routing decision. Match the cost model to your request pattern, match the latency to your user’s patience, and keep an escape hatch to a second provider when the first one rate-limits.