If you are generating images at scale, the model choice is mostly a cost decision. Quality differences across the four primary image models are real but small at typical use cases — the spread in price is 10x or more.
Below is the actual cost per image for GPT Image 2, Nano Banana 2, Seedream 5.0 Lite, and Grok Imagine, plus the math for the cost cliffs that change the answer at 100, 1,000, 10,000, and 100,000 images per month. Everything is sourced from official docs and aggregator pricing as of June 2026.
The goal of this post is not "which is cheapest." It is "which is cheapest for your scale and your quality floor." Those are two different questions.
The pricing per image, June 2026
All prices in USD. Per-image cost is calculated from the public API price for a standard 1024×1024 or equivalent output. Quality-parameter costs are noted where they differ.
| Model | Lowest quality | Standard quality | High quality | Notes |
|---|---|---|---|---|
| GPT Image 2 (OpenAI) | ~$0.006 (low) | ~$0.020 (medium) | ~$0.060 (high) | Token-based; 1024×1024 at low ≈ 1,300 tokens output |
Nano Banana 2 (gemini-3.1-flash-image, Google) | $0.020 (1K) | $0.045 (2K) | $0.090 (4K) | Free tier available with rate limits; per-image price is for paid API |
| Seedream 5.0 Lite (ByteDance) | ~$0.030 (2K) | ~$0.030 (2K) | ~$0.045 (3K) | Pricing is per-request, not token-based; fixed per size tier |
| Grok Imagine (xAI) | $0.020 (standard) | $0.050 (quality, 1K) | $0.070 (quality, 2K) | grok-imagine-image and grok-imagine-image-quality are separate models |
Sources (in priority order — official first):
- OpenAI pricing page (openai.com/api/pricing): GPT Image 2 token-based pricing ($8/1M input tokens, $2/1M output tokens for the image model variant)
- Google Gemini API pricing (ai.google.dev/gemini-api/docs/pricing): Nano Banana 2 image pricing per 1M tokens, with 1K/2K/4K output tiers
- xAI Grok docs (docs.x.ai/developers/models/grok-imagine-image): Grok Imagine per-image pricing
- Replicate and fal.ai for Seedream 5.0 Lite per-image pricing (Replicate charges ~$0.03 per 2K image)
- Aggregator pricing (OpenRouter, mem0, laozhang) used to cross-check and find the typical production rates
The hidden costs nobody quotes
The per-image price is not the cost. There are at least three other costs that change the real number:
1. Re-rolls. Most workflows are not "1 prompt → 1 perfect image." They are "1 prompt → 8 candidates → 1 chosen." If you are generating 8 candidates per accepted image, your real cost is 8x the per-image price. Some models are better at first-pass quality than others, so the re-roll multiplier is model-specific.
2. Failed outputs. Some models return a charged image that does not meet your bar (wrong aspect, broken text, identity drift). For text-heavy images, GPT Image 2 has a high success rate and a low re-roll multiplier. For identity-preserving edits, Nano Banana 2 tends to need more iterations.
3. Edit passes. Image-to-image or inpaint workflows charge per pass. A 3-pass edit workflow is 3x the per-image cost. Models with strong single-pass editing (GPT Image 2) have a real cost advantage here.
The real cost per delivered image is:
real_cost = per_image_price × candidate_count × edit_pass_count
If a model has a higher per-image price but a 2x lower candidate count, it can be cheaper in practice.
The cost cliffs at scale
Below is the math for what you actually pay at common production scales. Assumes standard quality, 2 candidates per accepted image, no edit passes.
At 100 images/month:
| Model | Per-image | 2 candidates | Monthly cost |
|---|---|---|---|
| GPT Image 2 (medium) | $0.020 | 200 calls | $4.00 |
| Nano Banana 2 (2K) | $0.045 | 200 calls | $9.00 |
| Seedream 5.0 Lite (2K) | $0.030 | 200 calls | $6.00 |
| Grok Imagine (standard) | $0.020 | 200 calls | $4.00 |
At small scale, the per-image price dominates. GPT Image 2 and Grok Imagine tie for cheapest.
At 1,000 images/month:
| Model | Per-image | 2 candidates | Monthly cost |
|---|---|---|---|
| GPT Image 2 (medium) | $0.020 | 2,000 calls | $40 |
| Nano Banana 2 (2K) | $0.045 | 2,000 calls | $90 |
| Seedream 5.0 Lite (2K) | $0.030 | 2,000 calls | $60 |
| Grok Imagine (standard) | $0.020 | 2,000 calls | $40 |
Same order. GPT Image 2 and Grok Imagine lead. Nano Banana 2 is 2.25x more expensive at this scale.
At 10,000 images/month:
| Model | Per-image | 2 candidates | Monthly cost |
|---|---|---|---|
| GPT Image 2 (medium) | $0.020 | 20,000 calls | $400 |
| Nano Banana 2 (2K) | $0.045 | 20,000 calls | $900 |
| Seedream 5.0 Lite (2K) | $0.030 | 20,000 calls | $600 |
| Grok Imagine (standard) | $0.020 | 20,000 calls | $400 |
Same order, $500 spread between cheapest and most expensive. This is where volume discounts from the providers start to matter.
At 100,000 images/month:
| Model | Per-image | 2 candidates | Monthly cost |
|---|---|---|---|
| GPT Image 2 (medium, volume) | $0.014 (est.) | 200,000 calls | ~$2,800 |
| Nano Banana 2 (2K, volume) | $0.035 (est.) | 200,000 calls | ~$7,000 |
| Seedream 5.0 Lite (2K) | $0.025 (est.) | 200,000 calls | ~$5,000 |
| Grok Imagine (volume) | $0.018 (est.) | 200,000 calls | ~$3,600 |
At 100K/month, GPT Image 2 holds the cost lead, but the spread is now $4,200/month. This is the scale where model choice stops being a per-image decision and starts being a contract decision — committed-use discounts, reserved capacity, and enterprise pricing can move any of these by 30-50%.
The quality-cost trade-off matrix
The price table above assumes "standard quality." Most real workflows mix qualities: low for iteration, high for finals. The optimal mix is:
Iteration volume (low quality, lots of candidates):
- GPT Image 2 low: $0.006 per image
- Nano Banana 2 1K: $0.020 per image
- Seedream 5.0 Lite 2K: $0.030 per image
- Grok Imagine standard: $0.020 per image
GPT Image 2 at low quality is 3-5x cheaper than the alternatives for iteration. If you do a lot of prompt experimentation, this is the workflow that wins on cost.
Final assets (high quality, fewer candidates):
- GPT Image 2 high: $0.060 per image
- Nano Banana 2 4K: $0.090 per image
- Seedream 5.0 Lite 3K: $0.045 per image
- Grok Imagine quality 2K: $0.070 per image
At the high-quality end, Seedream 5.0 Lite is actually the cheapest for final 3K output. If you generate at 2K for iteration and 3K for final, Seedream 5.0 Lite wins on the final pass. The model split is now: GPT Image 2 for iteration, Seedream 5.0 Lite for finals.
Hero assets (4K, premium):
- GPT Image 2 high: $0.060 (max 2K reliable)
- Nano Banana 2 4K: $0.090
- Seedream 5.0 Lite 3K: $0.045 (no 4K)
- Grok Imagine quality 2K: $0.070
For 4K output, Nano Banana 2 is the only primary model that does 4K natively. It is 50% more expensive than 2K output, but the only path to 4K. If you need 4K, the cost decision is made for you.
The decision framework: pick by use case
Stop asking "which is cheapest overall." Ask "which is cheapest for this use case."
You generate 1,000+ social images per month, 80% throwaway, 20% finals:
- Use GPT Image 2 at low quality for iteration, medium for finals.
- The low-quality iteration cost is 3-5x cheaper than competitors.
- Monthly cost: ~$30-50 for 1,000 delivered images.
You generate 100 high-quality marketing assets per month, mix of 1K/2K/4K:
- Use Nano Banana 2 for any 4K needs, Seedream 5.0 Lite for 2K/3K finals, GPT Image 2 for text-heavy images.
- Three models, picked by use case. Monthly cost: ~$15-30.
You generate 10,000+ product photos per month for ecommerce:
- Use Seedream 5.0 Lite for the bulk of 2K product shots, Grok Imagine for variant generation.
- The reasoning: Seedream 5.0 Lite has predictable per-request pricing and is optimized for product photography.
- Monthly cost: ~$300-500.
You generate 100,000+ images per month, automated pipeline:
- Negotiate committed-use pricing with OpenAI for GPT Image 2.
- The per-image price is 30-50% lower at volume, and the low-quality iteration tier is the cheapest in the field.
- Monthly cost: ~$1,500-3,000 at volume rates.
You generate under 100 images per month, ad hoc:
- Use Grok Imagine standard at $0.020/image. Cheapest at small scale, no minimum commit.
- Monthly cost: <$5.
What the price table does not tell you
Three things to verify before locking in a model on cost alone:
1. Free tier availability. Nano Banana 2 has a free tier via Google AI Studio. Grok Imagine has free credits. GPT Image 2 does not have a meaningful free tier. For prototyping, the free tiers change the math entirely.
2. Regional pricing differences. Some providers vary pricing by region. The xAI Grok docs explicitly call out "regional based pricing." Check the pricing for your deployment region before assuming the price in the table.
3. Token vs per-image pricing. OpenAI's GPT Image 2 charges by tokens; the others charge by image or by resolution tier. At very large sizes, GPT Image 2 token pricing can be cheaper than expected (a 1024×1024 image is ~1,300 output tokens ≈ $0.0026, but the API charges a per-image minimum that brings it to $0.006). At very small sizes, the per-image pricing of Seedream/Grok may be cheaper.
The summary
For most production workloads, the cost order from cheapest to most expensive is:
- GPT Image 2 (low and medium quality)
- Grok Imagine (standard and quality)
- Seedream 5.0 Lite (2K and 3K)
- Nano Banana 2 (1K, 2K, 4K — 4K is the unique capability)
But the right model depends on:
- Volume. Below 1K/month, per-image price dominates. Above 100K/month, contract pricing dominates.
- Quality mix. 80% iteration / 20% finals favors GPT Image 2. 100% finals at 2K/3K favors Seedream 5.0 Lite. Any 4K needs forces Nano Banana 2.
- Re-roll behavior. Models with stronger first-pass quality effectively have a lower per-delivered-image cost even at higher per-image prices.
- Free tier eligibility. For prototyping, free tiers on Nano Banana 2 and Grok Imagine change the math.
Pick by workflow, not by sticker price. The cheapest model for the wrong workflow is more expensive than the most expensive model for the right one.
