AI Image Generation Pricing: Compare Cost per Image

Official API prices checked September 10, 2026. All amounts are in USD.

A price of $30 per million tokens tells you very little about the cost of making a product photo. Neither does a “from $0.03” headline if the image you need is four times the advertised size.

To compare AI image generation pricing, start with a specific output: one square image, at a stated resolution and quality setting. Convert the provider’s billing unit into dollars per image, then add the input and workflow costs separately.

This guide compares OpenAI, Google, Black Forest Labs, Ideogram, BytePlus, Grok, Alibaba Cloud (Qwen), Meta, Microsoft, and Stability AI using their official pricing pages. It is a pricing comparison, not a claim that similarly priced models produce equivalent results.

The comparison baseline

Our main table uses 1024 × 1024 pixels, one output per request, and standard API processing. It excludes reference images, separately billed prompt tokens, generated text or thinking, streaming previews, search, retries, tax, and negotiated discounts.

That makes the shared unit image output cost, rather than a promise about the entire API invoice. For providers that charge a flat generation fee, the listed figure is that fee. For token billing, it is the image-output portion.

We keep the provider’s quality labels visible. “Medium” from one company is not a standardized quality score, and “low” should not quietly be compared with another model’s premium setting.

For clarity, this article uses:

• 1K square: 1024 × 1024 pixels.
• 2K square: 2048 × 2048 pixels.
• 4K square: 4096 × 4096 pixels.

A vendor’s “4K” landscape output may be a different size. It is not interchangeable with our 4K-square column.

AI image generation prices at 1K square

“Estimate” means the result displayed by the provider’s calculator. “Calculated” means published token counts multiplied by the published rate. Dollar totals for 1,000 outputs use the unrounded per-image amount shown here.

Sources: OpenAI’s image generation calculator, Black Forest Labs pricing, Google’s Gemini API pricing, and Ideogram’s API rate card.

Seedream Pro uses the lower output tier at 1024 × 1024. Seedream Lite does not support this 1K-square baseline; its $0.035 price appears in the 2K/4K table below. Sources: BytePlus price card, Seedream dimensions, Grok Imagine Image rates, and Grok Imagine Image 2.0 rates.

Qwen sources: Singapore / International rates and API size limits. Meta and Microsoft have separate pricing tables below because their verified quotes do not establish this same square-output baseline.

The OpenAI calculator currently groups Sunburst and Flare under “GPT Image 2.5.” These are its displayed estimates, not independently measured averages for each model. Actual usage remains the billing authority.

Google rounds some prices on its page: $0.0672 appears as approximately $0.067, for example. We retain the multiplication result so a 1,000-image budget does not accumulate rounding differences.

The cheapest row is a starting point for testing. It says nothing about how many attempts your brief will need.

What changes at 2K and 4K?

Here are comparable square-resolution scenarios. A dash means no supported, comparable figure is established in this table; it does not mean free generation.

Sources and calculations: OpenAI calculator and size limits, BFL’s first/additional-megapixel rates, and Google’s resolution-specific token counts.

Additional sources: BytePlus Pro rates, Lite rates, Seedream size constraints, and the official Grok base and 2.0 model cards. Seedream Pro supports 1K and 2K; Lite supports 2K through 4K. We map their fees to the explicit square dimensions above.

OpenAI’s documented maximum edge is 3840 pixels for GPT Image 2.5, so 4096 × 4096 is outside the documented limits. BFL caps FLUX.2 output at four of its billable megapixels, which includes 2048 × 2048 but not 4096 × 4096.

This is why “four times the pixels means four times the price” is not a safe shortcut. Each provider has its own billing curve.

Describe the video you want…

Pixo

Create AI videos with Pixo

Turn any idea into a publish-worthy video. One sentence is all it takes.

Start Creating

How to convert tokens into a per-image price

For a token-billed request, separate the bill into components:

Request cost =
  text input tokens × text input rate / 1,000,000
  + image input tokens × image input rate / 1,000,000
  + image output tokens × image output rate / 1,000,000
  + separately billed text, thinking, tools, or other charges

Use separate terms for cached and uncached input if caching applies. Do not charge the same input twice.

Worked example: GPT Image 2.5

The official calculator estimates 1,756 image-output tokens for a 1024 × 1024 image at high quality. With an image-output rate of $30 per million tokens:

1,756 × $30 / 1,000,000 = $0.05268

Now assume the image model receives exactly 200 billable text-input tokens, with no reference image or other charges. At $5 per million text-input tokens, the prompt adds $0.001. The resulting estimate becomes $0.05368 per request, or $53.68 for 1,000 such requests. The 200-token prompt is an illustrative assumption, not a measured token count for a particular sentence. Official token rates.

If a separate model plans or rewrites the prompt before invoking image generation, budget that model’s usage too.

Worked example: Nano Banana 2

Google lists 1,120 output tokens for a 1K image at $60 per million image-output tokens:

1,120 × $60 / 1,000,000 = $0.0672

Input, text, and thinking charges are additional. Google pricing.

The lesson is simple: the token rate alone cannot rank models by cost. You also need the number of tokens used to produce the image.

How to convert megapixels into a per-image price

BFL defines one billable megapixel as 1024 × 1024 pixels, and rounds the output up to a whole billing unit. Its first-megapixel price can differ from the rate for additional megapixels. BFL pricing rules.

For FLUX.2 [pro], a 2048 × 2048 output is four billing units:

Billable units = ceil((2048 × 2048) / (1024 × 1024)) = 4
Output cost = $0.03 + (4 − 1) × $0.015 = $0.075

Notice the two common mistakes: dividing by one million without checking the provider’s definition, and charging the first-megapixel rate for every unit. Both can distort a comparison.

Keep the original dimensions in your spreadsheet alongside the billing units. A label such as “2K” is useful for readers, but the pixel dimensions are what make the calculation reproducible.

Grok image pricing: specify the model and quality

Grok Imagine Image costs $0.02 per output at either 1K or 2K. Image inputs add $0.002 each. Official model card.

Grok Imagine Image 2.0 charges $0.04/$0.06 at 1K for low/medium, and $0.06/$0.08 at 2K. Each reference image adds $0.01. A 2K medium edit with one reference therefore costs $0.09, or $90 for 1,000 identical requests. Official 2.0 model card.

Keep the model identifier and quality setting in your budget: a generic “Grok image price” can refer to different products.

Seedream and Stability AI: translating other billing systems

Seedream 5.0 Pro: a resolution threshold

BytePlus lists $0.045 per output at or below 2.36MP and $0.09 above 2.36MP. That is $45 or $90 per 1,000 outputs. The first reference image is free; each additional input image costs $0.003. BytePlus’s official model price card.

This is a threshold, not a linear per-megapixel rate. A supported output above the threshold costs the higher amount; you do not multiply $0.045 by its megapixel count. The documented 1024 × 1024 output falls below the threshold, while 2048 × 2048 exceeds it. This gives $0.045 at 1K square and $0.09 at 2K square. Pro does not support 4096 × 4096. Official size constraints.

For example, three reference images add $0.006: the first is free and the next two are billed. That makes a lower-tier edit $0.051, or a higher-tier edit $0.096, before any other applicable charges.

BytePlus also lists Seedream 5.0 Lite at $0.035 per output, or $35 per 1,000. Its documented size range includes 2048 × 2048 and 4096 × 4096, but excludes 1024 × 1024. The resolution table therefore shows $0.035 at 2K and 4K, with no 1K quote. ModelArk pricing.

Stability AI: credits with a published dollar value

Stability AI defines one API credit as $0.01. Stable Image Core costs three credits and Stable Image Ultra costs eight, equivalent to $0.03 and $0.08 per generation, or $30 and $80 per 1,000. These are native-service fees, not verified fixed 1024 × 1024 quotes. Stability AI pricing.

Credits can be compared once their dollar value is known. Do not assume a credit on another platform has the same value.

Qwen image pricing: choose the deployment region

Alibaba Cloud lists Qwen-Image-3.0 at $0.03 per output for both 1K and 2K. Qwen-Image-3.0-Pro costs $0.04 at 1K and $0.075 at 2K. We use the Singapore endpoint, International deployment scope, in USD; other regions have different rates. Both models charge $0.003 per input image. Official regional pricing.

The API supports explicit 1024 × 1024 and 2048 × 2048 outputs within its documented pixel range. Neither model receives a 4K-square quote here. Official dimensions.

A 2K Pro edit with two reference images is $0.075 + 2 × $0.003 = $0.081, or $81 per 1,000 requests. Self-hosting another Qwen model or using a reseller is a separate cost scenario.

Meta and Microsoft: published prices with different comparison limits

These models belong in the comparison, but the sources below do not establish the same 1024 × 1024 output-cost baseline. Their quoted billing units are shown explicitly rather than ranked alongside the normalized 1K rows.

Meta Muse Image

Meta advertises Muse Image on Meta Model API at $0.01 per image, equivalent to $10 for 1,000 images. We have not established an exact supported-resolution matrix from the cited official announcement, so this is a published per-image quote, not a verified 1K, 2K, or 4K-square quote. Meta developer announcement listing.

Microsoft MAI-Image

Microsoft publishes the following token rates for MAI-Image-2.5 and Flash. These are USD per million tokens, not per-image prices. Official Microsoft pricing announcement.

For a request using T image-output tokens, the output charge is T × $47 / 1,000,000 for 2.5, or T × $19.50 / 1,000,000 for Flash. Add text and image inputs separately. We did not verify a fixed token count for a 1024 × 1024 output, so we do not substitute OpenAI or Gemini token counts to manufacture a per-image estimate.

Microsoft has also introduced MAI-Image-2.6 and 2.6-Flash in Foundry. The cited launch page does not establish numeric rates or a reproducible 1K-square cost; the 2.5 rates above must not be assigned to 2.6. Confirm the selected model, deployment, and current quote in Foundry. Official 2.6 announcement.

A usable image can cost more than a generated image

For a campaign, the number that matters is often cost per accepted asset:

Cost per accepted image = total generation and editing spend / accepted images

Consider a hypothetical project, not a model benchmark:

Workflow B has the higher generation fee but the lower cost per accepted result. Neither acceptance rate should be treated as typical of any named model.

Run a small comparison with your own briefs. Count rejected outputs, edits, and paid upscales. Check whether results preserve the product, subject, layout, and required text. Then compare the actual spend for a finished deliverable.

Our guide to AI image generators covers broader selection questions; the GPT Image 2.5 guide focuses on editing and visual control.

Build a budget you can actually use

Record the exact model, API provider, dimensions, quality setting, input-image count, and processing tier before comparing quotes. A reseller’s price is a price for that reseller’s service, even when the underlying model has the same name.

Keep standard and discounted processing in separate columns. A batch discount may be useful for overnight production, but it should not be presented as the regular price of an interactive workflow.

For subscription tools, divide the subscription spend by your actual accepted outputs over the same period. This gives an effective workflow cost, not a fixed API rate. Included generation quotas, unused capacity, and other bundled features make a subscription a different purchasing decision.

Several models covered in this guide are available on Pixo, including GPT Image 2.5, GPT Image 2, Nano Banana 2, Nano Banana Pro, and Grok Imagine Image. You can use them within your creative workflow. The official API prices above are for comparison; check the price displayed in Pixo for your selected model and settings.

Put the pricing comparison to work

• Compare tradeoffs in our GPT Image 2 vs Nano Banana 2 guide and GPT Image 2, Midjourney, and Imagen 4 comparison. Use this pricing guide for the September 10 API cost baseline; the older articles contain earlier estimates and different billing assumptions.
• Plan creative tests with the GPT Image 2 prompt guide and marketing application review.
• If texture artifacts are causing repeat attempts, see the GPT Image texture troubleshooting guide.

Frequently asked questions

What is the fairest unit for comparing image generation prices?

Use USD per image at the same dimensions, while keeping quality and processing tier explicit. Add separately billed inputs to estimate the request total. Use cost per accepted asset to compare complete workflows.

Are the prices in the main table my final invoice?

No. The table isolates image output or the provider’s flat generation fee. Token-billed inputs and any optional services may add to the total. Taxes, discounts, and retries are outside the baseline.

Can I assume 2K costs four times as much as 1K?

No. Four times as many pixels does not imply four times the charge. Billing may use token counts, a first-megapixel fee plus incremental units, or a resolution threshold.

Can a cheap model become expensive in practice?

Yes. A lower per-generation fee can be outweighed by additional attempts, edits, or upscaling. Track the cost of producing an asset you actually accept.

How should I compare an API with a subscription?

Keep their headline prices separate. For a shared operational measure, divide each workflow’s actual spend by its accepted outputs over the same period, and state which bundled services are included.

Methodology: This is a desk comparison of ten providers’ public API pricing, checked September 10, 2026. No paid generation benchmark was run. OpenAI estimates were read from its interactive calculator with explicit model, size, and quality settings; other conversions use published rates and documented units. Missing resolution mappings were not filled with assumed prices. Recheck the linked official rate cards before purchasing.