TL;DR

• Cheapest per-image for standard resolution: Gemini 2.5 Flash-Lite at roughly $0.0003/image, followed by Llama 4 Scout on Groq at $0.0005/image
• GPT-5 vision costs $0.00213/image at high detail (512-token baseline) - roughly 4x Gemini Flash-Lite at equivalent quality
• Claude Sonnet 4 vision pricing works out to $0.0048/image assuming 1,600 input tokens; Haiku 4 drops that to $0.0016/image
• For OCR-heavy pipelines (1M page scans/month), the spread runs from $300 at the cheap end to $4,800+ at the premium end - a 16x range
• "Vision" does not always mean good OCR: Llama 4 and several open-weight options handle natural images well but degrade on handwritten text and dense tables

The Bottom Line

Vision API pricing is the least-understood cost center in multimodal pipelines. Everyone models text token costs. Almost nobody models image costs correctly - until they get their first bill.

The reason is that vision pricing is harder to normalize than text. Providers charge in fundamentally different units: OpenAI charges by "detail level" (low vs. high), which maps to a fixed tile-based token count. Anthropic and Mistral charge by input tokens, and an image at 1,600 tokens costs the same rate as 1,600 text tokens. Google charges per image directly, with size-tier multipliers. AWS Nova charges per 1,000 input tokens with images billed at a token-equivalent rate. Once you normalize everything to cost per image at a defined resolution, the actual pricing spread is dramatic.

The other thing most comparisons miss: the cheapest vision model is often not cheap enough if it requires downstream post-processing to fix recognition errors. An OCR pipeline that's 95% accurate versus 99% accurate can mean 5x the human review time on a high-volume document batch. I've flagged where vision quality diverges meaningfully from the pricing tier.

For benchmark data on actual vision accuracy, see the multimodal LLM leaderboard. For guidance on which tool fits which use case, see best multimodal AI tools 2026.

How Image Costs Are Calculated

Before the tables: the math matters more for vision than any other API category.

OpenAI tile model: GPT-5 and GPT-4.1 split images into 512x512 pixel tiles. Low detail mode uses a fixed 85 tokens regardless of image size. High detail mode uses 85 tokens for a base image tile plus 170 tokens per 512x512 tile covering the image area. A 1024x1024 image at high detail = 85 + (4 tiles × 170) = 765 tokens. At GPT-5's $1.25/MTok input rate, that's $0.00096/image. A 2048x2048 image at high detail = 85 + (16 tiles × 170) = 2,805 tokens = $0.0035/image.

Anthropic token model: Images are billed at the same rate as text tokens. Image size is the key variable: a 200x200 image costs roughly 65 tokens; a 1,568x1,568 image (the highest-resolution Anthropic recommends for non-upscaled images) costs roughly 2,400 tokens. A 1024x1024 document page runs approximately 1,600 tokens as a practical baseline.

Google per-image model: Gemini charges a fixed rate per image with a surcharge above 384x384 pixels. The fixed rate structure makes per-image cost modeling much simpler - you do not need to tile-count.

What "low detail" actually means: Low detail / low resolution modes halve cost but also halve (or worse) the practical accuracy on detail-sensitive tasks. Good for: "does this photo show a dog or a cat." Bad for: "extract the line items from this invoice."

Ranked Pricing Table

Sorted by estimated cost per standard 1024x1024 image. Prices verified April 19, 2026. Where providers use token-based billing, I've normalized using a 1,600-token baseline for a 1024x1024 image.

All per-image costs are estimates based on a 1,600-token equivalent for a 1024x1024 input image, or the vendor's published tile count for that image size. Actual costs vary with image complexity, resolution, and compression.

Cost at Scale

This is where vision APIs diverge sharply from their headline prices. A pipeline that processes scanned documents - invoices, contracts, medical records, forms - at high volume should run these numbers before picking a provider.

1,000 images

100,000 images

1,000,000 images

The Gemini Flash-Lite to Claude Opus 4 gap at 1M images is $300 vs. $8,000. That 26x spread matters enormously for document-heavy production workloads. The right question is not "which is cheapest" but "which is cheapest at the accuracy level my pipeline actually needs."

Per-Provider Breakdown

OpenAI - GPT-5 and GPT-4.1 Vision

Pricing: GPT-5 charges $1.25/MTok input. GPT-4.1 charges $2.00/MTok input. GPT-5 Nano charges $0.05/MTok. GPT-4.1 mini charges $0.40/MTok. Image tokens are billed at the same per-token rate as text.

Low detail mode: fixed 85 tokens regardless of image size. At GPT-5 rates, that's $0.000106/image - very cheap, but low detail means the model receives a compressed 512x512 version of the image.

High detail mode: 85 base tokens plus 170 tokens per 512x512 tile. A 1024x1024 image = 765 tokens ($0.00096 at GPT-5 rates). A 2048x2048 image = 2,805 tokens ($0.0035 at GPT-5 rates). Document pages with fine print typically benefit from high detail.

GPT-5 auto mode selects low or high detail based on the prompt context. For cost-predictable batch pipelines, I recommend specifying the detail level explicitly.

What you get: GPT-5 vision is among the best for complex visual reasoning - reading charts, interpreting technical diagrams, answering questions about ambiguous images. GPT-4.1 is marginally behind on complex tasks but meaningfully cheaper for document-structured tasks like invoice extraction. GPT-4.1 mini is the value play for high-volume pipelines where the task is well-defined.

Best fit: Mixed visual workloads where you need one model to handle both natural image understanding and document OCR. Batch API gives 50% off all image token costs.

Gotchas: Tile-based pricing creates nonlinear cost scaling. A 4096x4096 image at high detail generates 10,625 tokens ($0.013 at GPT-5 rates). Always downsample before sending unless your task genuinely needs maximum resolution. The 85-token minimum for low detail mode is fixed - you cannot go cheaper by sending smaller images in that mode.

Source: OpenAI API Pricing

Anthropic - Claude Opus 4, Sonnet 4, Haiku 4

Pricing: Images are billed at the same rate as input text tokens. Opus 4: $5.00/MTok. Sonnet 4: $3.00/MTok. Haiku 4: $1.00/MTok. A 1024x1024 image runs approximately 1,600 tokens.

At those rates: Haiku 4 is $0.0016/image. Sonnet 4 is $0.0048/image. Opus 4 is $0.0080/image.

There is no equivalent to OpenAI's low detail mode - you send the image, you pay for the token equivalent of its size. Smaller images cost less proportionally: a 200x200 thumbnail costs roughly 65 tokens (Haiku: $0.000065; Sonnet: $0.000195).

What you get: Claude's vision capability is genuinely strong on document understanding - multi-page contracts, financial tables, forms with mixed text and checkboxes. Sonnet 4 is the most cost-practical tier for document pipelines. Opus 4 is justified when the task is complex interpretation (legal document analysis, medical imaging descriptions) rather than bulk OCR.

Batch API provides 50% off. At 1M images on Sonnet 4 with batch: $2,400 instead of $4,800. That's a legitimate consideration for large-scale document processing where latency tolerance is high.

Best fit: Document understanding pipelines where accuracy on structured data extraction matters more than cost. Haiku 4 is the right starting point for experimenting; Sonnet 4 for production; Opus 4 for highest-stakes interpretation tasks.

Gotchas: No low detail mode means no easy cost shortcut. Image size is the only lever. If your pipeline processes images that vary widely in resolution, implement server-side resizing before the API call. Sending a 3000x3000 raw scan of an A4 page burns 4,000+ tokens when a 1,568-pixel downsampled version gives equivalent OCR results at half the cost.

Source: Anthropic Claude Models

Google - Gemini 2.5 Pro, Flash, Flash-Lite

Pricing: Gemini charges per image with a size-based tier. Below 384x384 pixels: fraction of a cent. Above 384x384: the standard image rate applies. Gemini 2.5 Flash: $0.0009/image above size threshold. Gemini 2.5 Flash-Lite: $0.0003/image. Gemini 2.5 Pro: $0.0020/image. These are direct image charges, not derived from token counts, which makes cost modeling much simpler.

Video frames billed per frame at the same image rate. Audio has separate token pricing. Mixed multimodal prompts (image + text) combine the image charge with per-token text costs.

What you get: Gemini's multimodal architecture was designed with vision as a first-class input, not an add-on. Flash-Lite is the most price-competitive managed option in this entire comparison for bulk image work. Flash is the sensible production choice when you need better accuracy than Flash-Lite without paying Pro rates. Gemini 2.5 Pro's 2M context window is relevant for pipelines that need to send many images in a single context window - long PDF analysis, multi-page document understanding.

Best fit: High-volume image classification, product catalog tagging, content moderation, or any workload where you process images at scale and need predictable per-image costs without the tile-math complexity of OpenAI.

Gotchas: Flash-Lite's accuracy on handwritten text and low-contrast scans is noticeably weaker than Flash or Pro. Run a validation pass on your actual document corpus before committing. The generous free tier (Gemini API, not Vertex AI) is limited by rate caps: 10-15 requests per minute on the free tier. Production workloads need the paid API. Above 128K tokens per request, Gemini 2.5 Pro doubles its input price - relevant if you're batching many images in one call.

Source: Google Gemini API Pricing

Mistral - Pixtral Large and Pixtral 12B

Pricing: Pixtral Large: $1.00/MTok input, $3.00/MTok output. Pixtral 12B: $0.15/MTok input, $0.15/MTok output. Image token equivalents follow the same per-MTok rate as text. A 1024x1024 image at ~1,600 tokens: Pixtral Large costs $0.0016/image; Pixtral 12B costs $0.00024/image.

What you get: Pixtral is Mistral's vision-native architecture - not a retooled text model with a vision adapter bolted on. Pixtral 12B is particularly interesting for its price-to-capability ratio on document and chart understanding. Pixtral Large has 128K context and handles high-resolution documents well.

For OCR-specific workloads, Pixtral Large outperforms models of similar price on dense tabular data, multi-column layouts, and mixed-language documents (Mistral's European language coverage is notably broad).

Best fit: Document processing pipelines that need strong multilingual OCR at moderate cost. Pixtral 12B is worth testing as a Haiku 4 alternative - at roughly one-seventh the per-image cost, even with lower accuracy, it can be cost-positive for bulk classification tasks.

Gotchas: Pixtral 12B is not available on all regions of the Mistral API. Self-hosted deployment via Ollama or vLLM is an option for teams that want to avoid per-token costs entirely. Pixtral Large is available via mistral.ai/api but enterprise volume pricing requires contact.

Source: Mistral AI Pricing

Meta Llama 4 Vision - Scout and Maverick

Pricing: Llama 4 Scout on Groq: $0.11/MTok input. Llama 4 Scout on Together AI: $0.18/MTok. Llama 4 Maverick on Together AI: $0.27/MTok. Image tokens billed at the same input rate as text. At 1,600 tokens per image: Scout on Groq = $0.000176/image; Scout on Together = $0.000288/image; Maverick on Together = $0.000432/image.

Fireworks AI also hosts Llama 4 Scout at similar rates. Pricing across hosts varies by roughly 30-50% for the same model.

What you get: Llama 4's native multimodal architecture handles natural images and visual Q&A well. Scout at 17B active parameters is fast and cheap. Maverick's 128K context window matters for multi-image tasks. Open-weight status means you can also self-host these on your own hardware with zero per-token cost once compute is sunk.

Best fit: High-volume natural image understanding - content moderation, product classification, scene description. Less suitable as the primary OCR layer for scanned business documents.

Gotchas: "Vision" on Llama 4 means strong performance on clear photographs and web-scale image understanding. It means below-average performance on handwritten text, blurry scans, and densely formatted tables compared to models specifically fine-tuned for document understanding. If your pipeline handles both use cases, test separately before assuming Llama 4 covers everything. Groq's rate limits on Scout are lower than its text model limits due to compute intensity.

Sources: Groq Pricing, Together AI Pricing, Fireworks AI Pricing

xAI - Grok 4 with Vision

Pricing: Grok 4 standard: $3.00/MTok input, $15.00/MTok output. Vision inputs are billed at the same token rate as text. At 1,600 image tokens, Grok 4 costs $0.0048/image - identical on a per-token basis to Claude Sonnet 4.

What you get: Grok 4 gained native vision capability with its April 2026 release. Visual understanding quality is competitive with other frontier models on natural image tasks, but independent benchmarks on document OCR and structured data extraction are limited - xAI's primary showcases have been analysis of scientific figures and social media image understanding.

Best fit: Teams already running Grok 4 for text tasks who want a single API for multimodal pipelines rather than routing vision requests to a separate provider. The 256K context window handles moderately large multi-image prompts.

Gotchas: Grok 4's vision capability is newer than GPT-5, Claude Sonnet 4, or Gemini 2.5 Pro, and third-party benchmarks are still accumulating. At the same per-token cost as Sonnet 4, the choice between them should be driven by task evaluation on your specific document types, not headline pricing. xAI's batch API for vision is not yet publicly available.

Source: xAI Developer Models

Amazon Nova Pro Multimodal

Pricing: Amazon Nova Pro on Bedrock: $0.80/MTok input, $3.20/MTok output. Images billed as token equivalents at the input rate. A 1024x1024 image at approximately 2,250 tokens (AWS uses a different image-to-token conversion than Anthropic or OpenAI): $0.0018/image.

Nova Lite ($0.06/MTok input) processes images at an even lower rate, though with reduced accuracy. Nova Pro sits between Claude Haiku 4 and Sonnet 4 in the cost/capability ladder.

What you get: Nova Pro is the enterprise-grade multimodal option for teams building on AWS infrastructure. Native Bedrock integration means existing IAM, VPC, CloudTrail, and compliance configurations apply without additional setup. Nova Pro handles documents, images, and video frames. AWS's image-to-token ratio is slightly less favorable than competitors' at the same resolution, but the per-token rate compensates.

Best fit: Enterprise pipelines already on AWS that need a fully managed vision API with compliance guarantees. The Bedrock infrastructure handles at-rest encryption, audit logging, and data residency requirements out of the box.

Gotchas: AWS's per-image token count is higher than OpenAI or Anthropic for the same image size - factor this into cross-provider cost comparisons. Nova Lite is significantly cheaper but should be validated on your document types before relying on it for accuracy-critical tasks. The Bedrock console and AWS pricing calculator are the authoritative sources; third-party aggregators sometimes lag pricing updates.

Source: Amazon Nova on Bedrock - AWS

Open-Source Options via OpenRouter - InternVL and MiniCPM-V

Pricing: InternVL-2.5 on OpenRouter: varies by host, approximately $0.06-$0.10/MTok input. MiniCPM-V 2.6 on OpenRouter: approximately $0.05-$0.08/MTok depending on provider. At 1,600 image tokens: InternVL ~$0.00014/image; MiniCPM-V ~$0.00010/image.

What you get: InternVL-2.5 and MiniCPM-V are the leading open-weight vision models outside the Llama 4 family. Both are available self-hosted (transformers library, vLLM, Ollama) or via inference providers through OpenRouter. InternVL-2.5 78B scores within striking distance of GPT-4.1 on document OCR benchmarks - it was trained heavily on Chinese and English document datasets, which makes it unusually strong for mixed-language business documents.

MiniCPM-V is smaller (8B) and faster, trading accuracy for throughput. At $0.00010/image via OpenRouter, it's the cheapest option in this comparison. For high-volume image classification tasks where accuracy requirements are moderate, MiniCPM-V competes with Flash-Lite at a lower price.

Best fit: Teams comfortable with open-source tooling who want maximum control over the inference environment. Self-hosted deployment eliminates per-image costs after the hardware sunk cost. InternVL-2.5 in particular is worth evaluating for document-heavy workloads - the accuracy-to-cost ratio beats several proprietary options at this price tier.

Gotchas: OpenRouter pricing varies by the underlying provider routing the request - you don't always know which inference cluster you're hitting. Latency and throughput guarantees are weaker than managed APIs. Self-hosted InternVL-2.5 78B requires significant GPU memory (40+ GB VRAM for inference without quantization). MiniCPM-V is runnable on 16GB VRAM in 4-bit quantized form.

Source: OpenRouter Models

Qwen3-VL via Alibaba Cloud / Together AI

Pricing: Qwen3-VL 72B on Together AI: approximately $0.60/MTok. Via Alibaba Cloud's DashScope API (native provider): lower rates available but requires a DashScope account. At 1,600 image tokens: approximately $0.00096/image on Together.

What you get: Qwen3-VL is Alibaba's flagship vision-language model. It performs strongly on multilingual document tasks, particularly Chinese-language documents, mathematical notation in images, and diagram understanding. The 72B variant is competitive with Pixtral Large on document benchmarks.

Best fit: Pipelines handling Asian-language documents, technical diagrams, or mixed math/text images where specialized training on those data types matters. DashScope provides the lowest pricing for teams that can set up the account.

Gotchas: Together AI hosting is the most accessible western deployment path but pricing fluctuates. DashScope requires regional account setup that complicates compliance for non-Asian deployments. Vision quality on Western-script handwritten text and poor-quality scans is below Pixtral Large at similar price points.

Sources: Together AI Pricing

DeepSeek-VL3

Pricing: DeepSeek's vision model pricing via api.deepseek.com is approximately $0.27/MTok input for their multimodal endpoint. At 1,600 image tokens, that's $0.00043/image - competitive with the mid-tier options. API availability has been intermittent and DeepSeek has not formally published a dedicated pricing page for VL3 as of April 2026.

What you get: DeepSeek-VL3 is a solid multimodal model for general visual understanding. It handles document images and natural photos adequately. Performance is consistent with similarly priced models (Pixtral 12B, Llama 4 Scout) across standard benchmarks.

Best fit: Teams already using DeepSeek's text API who want to add vision capability without adding a second provider relationship. The unified billing and token pricing simplify accounting.

Gotchas: API availability is the main caveat. DeepSeek's servers have experienced capacity constraints during high-demand periods, and formal SLA commitments for vision endpoints are not published. For production pipelines requiring uptime guarantees, treat DeepSeek-VL3 as a supplementary option rather than a primary path until documentation improves.

Source: DeepSeek API Docs

Hidden Costs

Resolution Normalization

The single most effective cost optimization for vision APIs is pre-processing image resolution server-side before the API call. Most document scanning workflows produce 300 DPI images at A4 size - that's 2,480x3,508 pixels. Sent directly to GPT-5 at high detail: 5,015 tokens per image ($0.0063 at GPT-5 rates). Downsampled to 1024x1024 first: 765 tokens ($0.00096). Same OCR result. 6.5x cost reduction.

For Anthropic models, resize to 1,568 pixels on the longest edge before sending. That's the recommended maximum before quality benefits of higher resolution diminish. Going larger costs more without accuracy gain.

Batch API Discounts on Vision

OpenAI, Anthropic, and Google all apply their 50% batch discount to vision requests. At 1M images per month, that difference is substantial: Sonnet 4 at $4,800 high-detail becomes $2,400 with the Batch API, at the cost of 24-hour latency. Any document processing pipeline that can tolerate asynchronous processing should use the batch endpoint.

OCR Pipelines vs. Vision APIs

For pure text extraction from standardized documents - especially scanned forms, receipts, invoices - a dedicated OCR tool may beat vision APIs on both cost and accuracy. Google Document AI, AWS Textract, and Azure Document Intelligence are purpose-built for structured document extraction and often cheaper per page than sending the same document to a general vision model.

Use a vision LLM when the task requires understanding beyond text extraction: "summarize this contract clause," "identify anomalies in this financial table," "describe what's unusual about this medical scan." For pure "pull all the text off this page" tasks, evaluate Document AI pricing separately.

Context Window Management for Multi-Image Prompts

Sending 10 images in a single API call versus 10 separate calls affects cost differently per provider. Anthropic and OpenAI bill the same total tokens either way. Google's Gemini charges per image regardless of how many appear in one call. The optimization opportunity is in the output tokens: a single call asking "find all invoices in these 10 images" generates one output response versus 10. Output tokens are typically 3-5x more expensive than input tokens - combining image analysis into batched prompts reduces total output cost meaningfully.

Which Model for Which Vision Workload

FAQ

How is image pricing calculated for vision APIs?

It depends on the provider. OpenAI uses a tile model: images are split into 512x512 tiles and each tile costs 170 tokens at high detail, plus an 85-token base fee. Anthropic converts images to an equivalent token count based on image dimensions and bills at the same rate as text tokens. Google Gemini charges a fixed per-image fee with size tiers above 384x384 pixels. Amazon Nova charges at a token-equivalent input rate.

What is the cheapest vision API for bulk image processing?

Gemini 2.5 Flash-Lite at approximately $0.0003/image is the cheapest managed option. For open-source models hosted via OpenRouter, MiniCPM-V runs around $0.00010/image. If you can self-host, zero marginal cost is achievable with InternVL-2.5 or Llama 4 Scout on your own GPU infrastructure.

Does low detail mode actually save money?

Yes, significantly. On OpenAI models, low detail reduces any image to 85 tokens regardless of size. For a 2048x2048 image, that's a 33x cost reduction versus high detail. The trade-off is that the model receives a 512x512 downsampled version. Low detail works well for broad classification tasks. It fails on anything requiring text legibility - don't use it for OCR.

Should I use a vision LLM or a dedicated OCR tool for document processing?

For pure text extraction from standardized forms, dedicated OCR tools (Google Document AI, AWS Textract, Azure Document Intelligence) are typically cheaper and more accurate than general vision LLMs. Vision LLMs justify their cost when the task requires reasoning about document content - understanding context, summarizing clauses, flagging anomalies - rather than just transcribing characters.

How much does Claude cost per image?

Anthropic does not charge a separate per-image fee. Images are billed at the same per-token rate as text. A 1024x1024 image is approximately 1,600 tokens. At Claude Haiku 4 rates ($1.00/MTok), that is $0.0016/image. At Sonnet 4 rates ($3.00/MTok), it is $0.0048/image. At Opus 4 rates ($5.00/MTok), it is $0.0080/image. Resize images before sending to reduce token count.

What vision models support batch API pricing?

OpenAI's Batch API applies to all GPT-5 and GPT-4.1 vision requests at 50% off. Anthropic's Batch API covers all Claude models including vision inputs at 50% off. Google Gemini's async batch endpoint also provides discounts on image processing. xAI's batch API for vision is not yet available as of April 2026.

Sources:

• OpenAI API Pricing
• Anthropic Claude Models
• Google Gemini API Pricing
• Mistral AI Technology
• Groq Pricing
• Together AI Pricing
• Fireworks AI Pricing
• OpenRouter Models
• xAI Developer Models
• Amazon Nova - AWS
• AWS Bedrock Pricing
• DeepSeek API Pricing
• Cohere Pricing

Also see: LLM API Pricing Comparison and Image Generation API Pricing.