Per-Image Pricing: Examples & Companies

What is it

Per-image pricing is a billing unit where each AI-generated image is metered, common in image generation APIs and multimodal AI platforms.

The headline “one image, one charge” is only a starting rate. What a vendor actually bills depends on which model generated the image (a cheap diffusion draft versus a flagship multimodal render) and, where exposed, the resolution or quality tier — dimensions that stack, so a tier or model swap silently resets the cost.

The corpus splits per-image pricing across two buyer types. Developer-facing APIs like Reka AI, Hyperbolic, and xAI price images as a modality alongside tokens — a separate line item per image submitted to or returned by a multimodal model, which lets a buyer forecast cost-per-output directly. Application platforms like Stability AI price images indirectly through credit pools, so the end user sees credits consumed per generation and has to back-calculate the underlying rate.

How it works

Per-image pricing works by charging a fee each time an image is generated and returned. The fee is set at the API endpoint (or model) level, so the caller controls cost by choosing the model and quality settings before the request is sent. Most platforms expose the rate in dollars per image; some expose it in credits per generation, and OctoAI historically layered an additional GPU-compute-second component that scaled with inference effort.

The dimensions that drive rate variation across the corpus:

Unit math: Total image spend = (images generated) × (rate per image for that model/quality tier). For diffusion models with a variable compute meter (OctoAI historically): Total = (images) × (base per-image fee) + (total compute seconds) × (per-second rate).

A worked example using FLORA’s Pro plan: a 4-seat Pro team costs $200/month in seats (4 × $50/seat) and includes a $200/month pooled usage budget (4 × $50/seat included). At FLORA’s published Nano Banana 2 rate of $0.151/image, that pool funds roughly 1,325 images per month before any pay-as-you-go overage. If the team switches to a more expensive model from FLORA’s 60+ integrated options (such as Veo 3.1 for video), the same $200 pool covers far fewer generations. Because FLORA bills at published dollar rates rather than credits, the per-image rate — not a hidden credit conversion — is the variable to manage.

Companies using this

Seven companies in the corpus bill on a per-image unit, spanning developer inference APIs, multimodal platform APIs, and creative application subscriptions. Stability AI is the category-defining company — the originator of Stable Diffusion, which established per-image metering as the standard for AI image generation.

Patterns observed

Per-image is a line item bolted onto a token meter, never standalone. Every API-facing company here bills images alongside per-token text on the same card. Reka Core’s $2/M input tokens and its per-image rate ride one multimodal API; xAI’s grok-imagine sits beside grok-4.3’s $1.25/M input rate; MiniMax draws image-01 and its MiniMax-M2/M3 $0.30/M token rates from one prepaid USD balance. Per-image pricing is an extra meter layered on existing token infrastructure, not its own billing system.

Transparency at the API layer, opacity at the application layer. The API vendors — Reka AI, xAI, MiniMax, and the discontinued OctoAI — all publish explicit dollar rates. The application and infrastructure vendors don’t: Stability AI’s Brand Studio hides the rate behind credits, and Hyperbolic serves SDXL and FLUX but lists only its token and GPU-hour rates, never the per-generation image price. FLORA breaks the layer’s convention by billing at published dollar rates, selling that transparency to agencies that pass costs through to clients.

The rate spread runs past 40x. MiniMax’s $0.0035/image floor and FLORA’s $0.151 Nano Banana 2 ceiling bracket the corpus at roughly 43× apart — a developer embedding images in a pipeline versus a creative team paying for flagship quality. The gap tracks genuine model-capability and inference-cost differences, not pricing philosophy.

Counterexamples & variants

OctoAI’s dual meter shows the limit of a flat per-image rate. Before its October 2024 sunset (NVIDIA acquired the team), OctoAI priced SDXL, SD 1.5, and Stable Video Diffusion on a per-image fee plus a per-GPU-compute-second component — so a high-fidelity, high-step render cost more than a fast draft at the same resolution. It was the most honest model for diffusion, where two images at one resolution can differ sharply by inference depth, but commercially awkward: buyers budgeting per image couldn’t forecast the compute-second line, and after the acquisition the rate card vanished.

Credit abstraction breaks like-for-like comparison. Stability AI’s Brand Studio sells $50/month for 5,000 credits (free 1,000-credit Trial), but the images-per-credit ratio depends on model and quality settings governed by a separate, unpublished Platform API. A developer can’t compare providers on a level field when one quotes dollars-per-image and the other quotes credits-per-month with no public conversion — opacity that softens price competition for the vendor and burdens the buyer, and exactly the pattern FLORA’s dollar-rate billing rejects. xAI is a different variant: grok-imagine is a feature extension on a platform whose real meters are token rates and per-1k-call agentic tools like X Search ($5/1k), so per-image is a convenience add-on, not a core value metric.

What this means for buyers vs vendors

For buyers

Price the model tier you will actually ship at, not the headline rate — the same platform can be 4× apart across its own model family. Confirm the published rate covers your resolution and quality, and flag anything you can’t forecast from the public page: a compute-second component (OctoAI’s old model) or an unpublished image rate (Hyperbolic). Weighing a credit subscription against pay-as-you-go API billing, model the break-even: for spiky or exploratory workloads, transparent per-image billing usually beats prepaying a credit bundle and forfeiting the unused balance. See the usage-based pricing fundamentals guide for the break-even math and choosing the right usage metric for when a per-output unit fits better than per-token or subscription. For xAI, the xAI pricing calculator models token-plus-media cost.

For vendors

Tier your rate card by model and quality — one flat rate either underprices flagship renders or overprices drafts and pushes buyers away. A per-model ladder (Reka AI’s) lets buyers self-select their tier while you hold margin at the top. Decide deliberately between raw dollars and credits: credits blunt direct comparison but breed the opacity sophisticated buyers, especially agencies, distrust — FLORA shows transparency can be the sales argument. For variable-depth diffusion workloads, OctoAI’s per-image-plus-compute hybrid was more accurate but harder to sell; a fixed draft / standard / high-fidelity ladder is the legible alternative.

Explore this theme in the knowledge graph →