AI image generator pricing: plans, APIs and real costs

The number that matters for a project is cost per accepted image: total relevant spend divided by the images you can actually use. A cheap generation that needs several replacements or extensive repair may cost more to finish.

AI image subscriptions: monthly bills versus annual commitments

Prices are in USD and were checked on September 10, 2026. Taxes, region, promotions and checkout terms can affect the final bill. Annual equivalents describe the cost spread over a year; they are not one-month purchase prices. Sources: Magic Hour, Midjourney, ChatGPT Plus, Ideogram and Firefly.

Magic Hour publishes this guide and is one of the products compared. These prices establish costs and allowances, not a ranking of image quality.

Which kind of billing fits your work?

A subscription for interactive creation

A subscription can make sense when you generate and edit images regularly in an app. It pays for the workflow as well as generation: interfaces, saved work, editing tools and sometimes other media features.

Magic Hour combines image generation, image editing and video tools. Check the selected model’s credit quote; do not treat the plan page’s illustrative image total as the output count for every model and resolution. Paid plans permit commercial use. The free tools have their own allowances.

Midjourney meters Fast use in GPU time. Basic’s 200 minutes are not a guaranteed 200 finished images. Standard adds a slower image-generation queue, while Pro and Mega add Stealth mode. Its web interface means Discord is an option rather than a requirement.

ChatGPT Plus provides image creation alongside other ChatGPT features. Its usage guidance says limits can vary with system conditions, so a fixed claim such as “50 image prompts every three hours” is not a reliable allowance. OpenAI introduced ChatGPT Images 2.5 on September 8, 2026; older articles naming Image 1.5 as the current ChatGPT image model are out of date.

Ideogram separates priority credits from slow-queue access. Its live pricing page lists Plus, Pro and Team purchase options, with credit consumption depending on model and rendering settings. Read that page before relying on older Basic-plan prices or a fixed daily free quota. Its plan documentation also directs purchase decisions to the live pricing flow.

Adobe Firefly combines its own models with selected partner models. Firefly’s plans distinguish standard and premium generation access. Buying Firefly access does not automatically mean every Creative Cloud application or partner-model generation is included without limits.

If you already subscribe to Canva or a Google AI plan, check the image tools and remaining allowance included in that account before adding another subscription. Their Canva AI allowances and Google AI plan benefits vary by plan and feature. The app subscription is not a universal per-image API price.

An API for programmatic generation

APIs are useful when generation needs to run inside a product, catalog pipeline or repeatable workflow. Compare the exact provider, model, input type, output dimensions and quality setting. An API service with the same model name can still have different rates and billing rules from another host.

Primary rate references: Stability API pricing, Gemini API pricing and OpenAI API pricing. Web subscriptions and API billing can be separate. OpenAI explicitly bills API usage independently of ChatGPT Plus.

For OpenAI’s token-based model, a hypothetical request billed for 400 text-input tokens and 1,000 image-output tokens, with no image input, costs $0.002 + $0.03 = $0.032 at those rates. This illustrates the calculation; it is not a promise that a particular resolution uses 1,000 output tokens. Use the image cost calculator for the selected model and settings.

Worked budget: 100 finished images

Suppose you need 100 images and budget for two billable generations per accepted result. That gives 200 billable generations. Two attempts is a planning assumption, not an observed success rate.

This is a billing example, not a claim that these models produce equivalent images or achieve the same acceptance rate. Include any additional editing, upscaling, stock assets and manual work in the project budget. Minimum credit purchases can also make checkout spend higher than the usage charge for a small job.

For subscriptions, start with the full amount you actually pay. If a $19 monthly plan is used only for one project and produces 100 accepted images within its allowance, the subscription allocation is $0.19 per accepted image. If it produces only 20, that allocation is $0.95. If you also use the plan for videos or other projects, allocate its cost consistently instead of charging the whole subscription to each job.

A simple cost-per-accepted-image formula

Cost per accepted image = (subscription allocation + API or credit charges + editing and other project costs) ÷ accepted images.

Do not add both the price of a credit pack and the same credits’ dollar usage a second time. Track cash paid and credits consumed separately when you need both a cash budget and a project-cost estimate.

To compare two tools, give them the same brief, required dimensions and acceptance criteria. Keep the image only if its subject, text, important details and export meet the actual requirement. Record the attempts and correction time. That produces a useful comparison without pretending that every generated file is finished work.

Is running an image model locally free?

A downloadable model can remove a hosted per-request charge, but running it still uses hardware, electricity and your time. Cloud GPU rental adds an hourly cost; setup, storage and idle time can matter as much as generation speed.

For electricity alone, use: average measured power in kilowatts × runtime in hours × your electricity price per kWh. For example, an assumed 0.4 kW system running for 10 seconds at $0.20/kWh uses about $0.00022 of electricity. That is an illustrative calculation, not a benchmark or an all-in image cost. It excludes hardware, idle time and labor.

Licensing is another separate check. The Stability Community License allows covered Core Models to be used under its conditions, with commercial eligibility tied to an annual-revenue threshold. Other models can have different licenses. “Downloadable,” “free to run” and “unrestricted commercial use” are not interchangeable descriptions.

What can increase the real cost?

Reference images and revisions. Editing an existing image can add input charges or consume a different number of credits from text-to-image generation. A requested revision may be another paid generation even when most of the image stays the same.

Resolution and quality. A low-cost preview setting may not match the final export requirement. For Google’s Nano Banana 2 Standard API output, the listed approximate image-output equivalent rises from $0.067 at 1K to $0.151 at 4K, before other applicable charges. Generate drafts at the needed working size, then budget for final production settings.

Private generation and team requirements. Privacy can be a paid feature. Midjourney’s Stealth mode requires Pro or Mega; its commercial terms also require those plans for companies above $1 million in annual gross revenue. Check the plan that actually fits your organization rather than assuming the cheapest plan applies.

Unused allowances and billing cadence. A large credit allocation only lowers your effective cost if you use it. Check renewal dates, expiration, top-up terms and the full annual commitment. A discounted yearly price can be a poor fit for a one-week project.

Commercial permission and source rights. Magic Hour’s paid plans permit commercial use; free outputs are for personal use. Adobe’s Firefly indemnification is conditional on qualifying agreements, features, surfaces and export events. It is not automatic protection for every free account or every partner model. You still need to assess the rights to reference photos, recognizable people, logos and other inputs.

Which option should you start with?

For occasional personal images, try a free tool with an export you can inspect. For frequent interactive work, choose the workflow you can finish in, then buy the smallest plan that covers your demonstrated usage. For a product integration, compare API costs with the exact model and settings you intend to ship.

If you need both images and video, start with Magic Hour’s image generator, check a representative result, and use the image editor for a focused correction. A finished image can then feed image-to-video. Budget for each stage rather than assuming a photo credit also pays for animation.

For broader selection criteria, see our AI image-generator guide and image and video API comparison.