AI Image Generation Cost Calculator: Per-Image Pricing Across Every Major Model (2026)

Subscription pricing (Midjourney, Adobe Firefly, Krea, Leonardo, consumer Ideogram) charges a flat monthly fee for a fixed allotment of generations or credits. Midjourney Basic at $10/month bundles roughly 200 fast generations; Standard at $30/month bundles ~1,500 fast plus unlimited 'relaxed' generations.

API pricing (OpenAI Images, Replicate, Fal, Stability, Vertex AI) charges per image, with rates varying by model and resolution. No monthly minimum, no maximum — usage scales linearly with cost.

The break-even point: subscriptions win when you consistently generate enough images to use the full allotment. Midjourney Standard at $30/month for 1,500 fast generations works out to $0.02/image — cheaper than every per-image API at comparable quality. If you generate 100 images/month, the same $30 buys nothing on a subscription you only used 7% of, while $4 of OpenAI gpt-image-1 mini calls gets the same job done.

Rule of thumb: under 200 images/month, APIs win on raw cost. Over 1,000 images/month, subscriptions win for creative workflows; APIs still win for product integrations where end-user count varies.

Reference workload: 100 images, 1024×1024 standard quality, single project.

Midjourney Basic ($10/mo): the cheapest subscription tier covers 100 generations easily, so the effective cost is $10/100 = $0.10/image. If you cancel after one month, total is $10.

OpenAI gpt-image-1: 100 × $0.040 = $4.00.

Flux Pro 1.1 (Replicate): 100 × $0.040 = $4.00.

Flux Schnell (Together): 100 × $0.003 = $0.30.

Stable Diffusion XL (Together): 100 × $0.005 ≈ $0.50.

OpenAI gpt-image-1-mini: 100 × $0.015 = $1.50.

For 100 images, APIs are clearly cheaper than the cheapest subscription. The quality decision matters more than the price decision here — $0.30 to $4 are all small absolute numbers. Pick by aesthetic fit and prompt-adherence requirement: Midjourney for stylized art, Flux Pro for photorealism, Ideogram if text in the image matters, gpt-image-1 for tightly-controlled product output.

Reference workload: 1,000 images per month, standard quality, mixed prompts.

Midjourney Standard ($30/mo, ~1,500 fast included): effective cost $30/1,000 = $0.030/image. Plus relaxed mode is unlimited at slower speed — adds extra capacity for free.

Midjourney Pro ($60/mo, ~3,000 fast): effective cost $60/1,000 = $0.060/image. Worth the upgrade only if you need stealth mode or hit the 1,500 limit.

OpenAI gpt-image-1: 1,000 × $0.040 = $40/month.

Flux Pro 1.1: 1,000 × $0.040 = $40/month.

Flux Schnell: 1,000 × $0.003 = $3/month.

OpenAI gpt-image-1-mini: 1,000 × $0.015 = $15/month.

At 1,000 images/month Midjourney Standard wins on creative quality per dollar ($30/mo all-in). Flux Schnell wins on raw cost ($3/mo). Most teams using image generation for marketing content land on Midjourney Standard; product features integrating image gen for end users land on Flux or OpenAI APIs. For prompt-quality help, our Midjourney prompt builder and Stable Diffusion prompts tighten generation prompts so fewer regenerations are needed.

Reference workload: 10,000 images per month for a product feature — say, AI-generated avatars or product mockups in a consumer app.

Midjourney is no longer the right fit — even the Pro plan's 3,000 fast generations falls short, and most subscription terms restrict commercial API use without enterprise agreements.

API pricing: OpenAI gpt-image-1 = 10,000 × $0.040 = $400/month. gpt-image-1-mini = $150/month. Flux Pro 1.1 = $400/month. Flux Schnell = $30/month. Stable Diffusion XL via Together = $30-80/month.

At this scale, the open-weights stack (Flux Schnell, SDXL, sometimes self-hosted) is materially cheaper and usually wins on cost. Trade-offs: lower aesthetic quality on subjective evals than Midjourney or gpt-image-1, less prompt-adherence than gpt-image-1, less photorealism than Flux Pro.

Hybrid pattern that works well: use Flux Schnell or SDXL for first-pass generation, gate the output through a quality filter (a smaller LLM scores the result), and regenerate failures on Flux Pro. The blended cost lands around $0.005-$0.015 per delivered high-quality image — $50-150/month at 10k volume.

Aesthetic / artistic quality: Midjourney v7 still leads on subjective evals for stylized illustration, painterly looks, and unconventional aesthetics. Aided by community-trained styles and rich parameter system.

Photorealism: Flux Pro 1.1 Ultra leads as of June 2026 — published comparisons against gpt-image-1 and Imagen 4 Ultra typically favor Flux on skin, hair, hands, and incidental detail. Imagen 4 Ultra is a close second and stronger on landscapes.

Prompt adherence (does the image match what you asked for): OpenAI gpt-image-1 and Imagen 4 Ultra lead. Both correctly handle multi-subject prompts ('a red ball next to a blue cube on a yellow table') more reliably than Midjourney or Flux.

Text in image (signs, labels, posters): Ideogram 3.0 leads by a wide margin, followed by gpt-image-1. Midjourney still struggles with anything beyond short readable words; Flux is improving but inconsistent.

Cost per high-quality image at scale: Flux Schnell and Stable Diffusion XL via Together or Fal lead — under $0.005/image — with quality acceptable for many production use cases.

Commercial safety: Adobe Firefly is the only widely-deployed model trained exclusively on licensed/owned content, making it the default for advertising agencies and regulated industries. The other models all have some training-data ambiguity.

Vector / brand style: Recraft v3 is the only production-grade option for vector output and tight brand-style control via custom style packs.

Higher resolution costs more, often non-linearly. OpenAI gpt-image-1 at standard 1024×1024 is $0.040; at HD 1024×1024 it doubles to $0.080; at 1792×1024 (widescreen HD) it jumps to $0.120 — a 3x premium for ~1.75x the pixels.

Flux Pro 1.1 at 1024×1024 is $0.040; the Ultra variant at higher resolution is $0.055 — a smaller premium because the Ultra variant uses a different model rather than just more pixels.

Midjourney charges 'fast hours' for higher resolutions and upscales, which deducts from your monthly allotment. A 2x upscale typically costs 0.5-1 minute of fast time, equivalent to 1-2 standard images.

Practical advice: generate at 1024×1024 for ideation, upscale only the winners. Most teams over-resolve by default and burn 2-3x the budget for marginal visible improvement on screens. For print or hero-image use, plan one upscale pass on the chosen winner.

Image generation pricing only tells half the story for content teams in 2026, because most production workflows now mix stills with short video clips. Video models price on a fundamentally different unit — per second of output rather than per image — and the per-second number lands 50-200x higher than the per-frame cost of an equivalent image model. That gap matters once you start building social ads, product explainers, or storyboard previews, because the budget math no longer carries over from image work.

OpenAI Sora 2 sits at the premium end of the API market. Standard 1080p output is priced around $0.30 per second of finished video; the 720p tier drops to roughly $0.10 per second. A 10-second 1080p clip runs $3.00, a 30-second clip runs $9.00, and a one-minute clip lands at $18.00. The 720p tier turns the same one-minute clip into $6.00 — a meaningful saving when the clip is destined for vertical social where the viewer will never see the difference between 720p and 1080p on a phone screen.

Google Veo 3 prices by clip rather than by raw second. A standard-quality 8-second clip costs roughly $0.50, and the high-quality tier (sharper detail, better motion coherence, more accurate physics) is around $1.20 for the same 8-second window. That works out to $0.06 per second standard and $0.15 per second high quality — meaningfully cheaper than Sora 2 at the standard tier, with the trade-off that Veo 3 caps individual generations at 8 seconds and you have to stitch clips for longer outputs. A 30-second ad assembled from four Veo 3 high-quality clips runs about $4.80, versus $9.00 on Sora 2 at 1080p.

Runway Gen-4 sells two ways. The subscription tiers are $15/month (Standard, ~625 credits), $35/month (Pro, ~2,250 credits), and $95/month (Unlimited, generation cap plus Explore mode for unmetered runs). Credits convert roughly 1:1 with seconds of generated video at base quality, so the $35 Pro tier effectively delivers about 37 minutes of video monthly — equivalent to $0.016 per second if you actually consume the full allotment. The Runway API is priced separately at roughly $0.05 per second, putting a 30-second clip at $1.50. The subscription wins by a wide margin for an in-house creative team; the API wins if you are embedding generation into a product where end-user demand is unpredictable.

Pika 2.0 is the price leader on the API side at roughly $0.04 per second, putting a 30-second clip at $1.20 and a 10-second teaser at $0.40. The trade-off versus Sora 2 or Veo 3 is shorter individual clip lengths (usually 5-10 seconds before quality degrades) and weaker physics on complex motion. Kling 2.0 splits the difference at around $0.08 per second — a 30-second piece costs $2.40 — with stronger motion coherence and better handling of human subjects than Pika, but less polish than Sora or Veo on cinematic shots.

Two pricing dynamics carry across every model. First, resolution scales cost roughly 3-4x: 1080p typically costs 3-4 times the per-second rate of 720p across Sora, Runway, and Pika. Veo 3 hides this in its standard-versus-high-quality tier split, but the multiplier is comparable. Duration scales linearly within a single generation, but most models charge a small re-initialization overhead per clip, so generating one 16-second clip is marginally cheaper than two 8-second clips when the model supports the longer single take. Second, the cost-per-second on the cheapest video tier (Pika 2.0 at $0.04) is still about 13x the cost of a single Flux Schnell image ($0.003) — and a single second of video is 24-30 rendered frames. The per-frame video cost is actually competitive; the human-visible second is what makes the bill jump.

Worked example for a 30-second product ad at 1080p: Sora 2 direct = $9.00 per take, and most teams run 3-5 takes for a usable result, so the budget lands at $27-45 for one finished ad. Veo 3 stitched from four 8-second high-quality clips = $4.80 per take, $14-24 for a usable result. Runway Gen-4 Pro subscription at $35/month covers 37 minutes of generated material, so a single 30-second ad amortizes to under $0.50 if you are producing more than one ad in the month. Pika 2.0 across six 5-second clips = $1.20 per take, $4-6 for a usable result, with the caveat that quality will not match Sora 2 or Veo 3 on cinematic compositions. Most startup commercial-content workflows in 2026 follow a consistent three-stage pattern: generate base footage cheaply (Pika or Kling), pay for an upscale and motion-smoothing pass (Topaz Video AI at $0.03 per second, or Runway's Upscale endpoint at $0.04 per second), then manually select the keepers. A blended budget of $0.10-0.15 per delivered second of finished 1080p video is achievable, which puts a 30-second ad at $3-5 finished rather than the $27-45 of running everything on Sora 2 directly.

Commercial licensing is the part that catches teams off guard. OpenAI Sora 2 grants commercial rights to API customers but prohibits generating likenesses of real people without explicit consent, and outputs must include the embedded C2PA provenance signal. Google Veo 3 includes a SynthID watermark on every generation, and the commercial license through Vertex AI is straightforward for enterprise customers but more limited on the consumer Gemini tier. Runway Gen-4 grants full commercial rights on the Pro and Unlimited subscription tiers; the Standard tier restricts commercial use to a single account holder. Pika 2.0 grants commercial rights on its paid tiers but explicitly disallows training derivative models on outputs. Kling 2.0, operated by Kuaishou, grants commercial rights but has had documented enforcement inconsistency outside China — most US-based ad agencies treat Kling output as draft-only and re-shoot the keeper with Sora, Veo, or Runway for the deliverable. Always re-read the active terms before a campaign launches, because every one of these vendors has changed terms at least once in the last 12 months.

Tighten prompts. Most teams generate 4-8 candidates per intended deliverable, then pick the best. Tightening prompt specificity (subject, composition, style, lighting, lens) cuts the regen rate by 30-50% in our experience. See the Midjourney prompt builder and DALL-E prompt creator for structured prompt templates that reliably hit the bullseye on the first try.

Match resolution to delivery surface. Social media and thumbnails do not need HD or 2x upscales; default to standard 1024×1024 and only upscale the hero shots.

Cascade quality tiers. For product features serving 10k+ images/month, run first-pass on Flux Schnell or SDXL, gate output through an automated quality check, regenerate failures on Flux Pro or gpt-image-1. Blended cost is usually 60-80% lower than running everything on the premium tier.

Reuse generations. Image generation often produces near-duplicates of prior generations. Caching at the prompt-hash level catches 5-15% of redundant generations in most workloads.

Right-size the subscription. If you used 250 of your 1,500 monthly Midjourney generations last month, downshift to the $10 tier. If you blew past the cap, upshift. Most teams either over-pay or run out and patch with one-off API calls.