Best LLMs for Vision and Document Understanding
Reading screenshots, scans, charts and photos alongside text.
Typical request this page prices
- Quality floor: Top 70% of priced models by Intelligence Index (≥ 9.2).
- Only models recorded as accepting image input.
- Costs use each provider’s current standard rates, with cached tokens at the cached-input rate where one is published. Speed is from Artificial Analysis where measured.
Lowest cost that meets the bar
Models that clear the quality floor, ordered by estimated cost for the request above.
| Model | Intelligence | Speed | Cheapest host | Input / output per 1M | Per 1,000 requests |
|---|---|---|---|---|---|
| Gemma 4 26B Google · 262K context | 16.7 | 101tok/s | OpenRouter 7 providers | $0.042 / $0.220 cached $0.021 | $0.236 |
| Qwen 3.5 9B Alibaba · 256K context | 13.3 | 99tok/s | Empirio Labs 6 providers | $0.090 / $0.130 cached $0.045 | $0.335 |
| Gemma 4 31B Google · 262K context | 14.7 | 47tok/s | Deep Infra 12 providers | $0.090 / $0.340 cached $0.050 | $0.440 |
| Gemini 2.5 Flash Lite Google · 1.0M context | 10.4 | - | Prime Intellect 3 providers | $0.100 / $0.400 | $0.500 |
| Gemini 3.1 Flash Lite Google · 1.0M context | 15.6 | - | Velokey 5 providers | $0.100 / $0.600 cached $0.010 | $0.600 |
| Qwen 3.5 35B Alibaba · 262K context | 19.3 | - | OpenRouter 6 providers | $0.080 / $0.750 cached $0.040 | $0.615 |
| Qwen 3 VL 30B Alibaba · 131K context | 9.5 | - | Deep Infra 4 providers | $0.150 / $0.600 | $0.750 |
| Llama 4 Maverick 17B Meta · 128K context | 10 | 50tok/s | OpenRouter 7 providers | $0.188 / $0.652 cached $0.050 | $0.889 |
Highest-scoring options
Ordered by Artificial Analysis Intelligence Index, with the cost of the same request.
| Model | Intelligence | Speed | Cheapest host | Input / output per 1M | Per 1,000 requests |
|---|---|---|---|---|---|
| Claude Fable 5 Anthropic · 1.0M context | 49.6 | - | Velokey 6 providers | $4.00 / $20.00 cached $0.400 | $22.00 |
| Qwen 3.8 Max Alibaba · 1.0M context | 45.4 | 36tok/s | Deep Infra 9 providers | $1.65 / $4.95 cached $0.206 | $7.43 |
| Kimi K3 Moonshot · 1.0M context | 43.6 | 40tok/s | OpenRouter 16 providers | $0.730 / $10.00 cached $0.550 | $7.19 |
| Claude Opus 4.8 Anthropic · 1.0M context | 41.8 | - | Velokey 6 providers | $2.00 / $10.00 cached $0.200 | $11.00 |
| Claude Opus 4.7 Anthropic · 1.0M context | 40.7 | - | Velokey 6 providers | $2.00 / $10.00 cached $0.200 | $11.00 |
| Gemini 3.7 Flash Google · 1.0M context | 39.6 | - | Deep Infra 5 providers | $0.750 / $3.75 | $4.13 |
| GPT-5.4 OpenAI · 128K context | 39 | - | Velokey 6 providers | $0.750 / $4.50 cached $0.075 | $4.50 |
| Gemini 3.6 Flash Google · 1.0M context | 34 | - | Velokey 5 providers | $0.600 / $3.00 cached $0.060 | $3.30 |
Compare every model and provider on the LLM API pricing table.
How to choose an LLM for vision & document understanding
Vision workloads send an image — a scanned page, a screenshot, a chart, a product photo — with a short instruction. Providers convert each image into input tokens, typically between several hundred and a few thousand depending on resolution, so the estimate treats an image plus prompt as a few thousand input tokens.
Only models recorded as accepting image input are listed. Some recent models support images but are not yet tagged that way in our catalog; they will appear once their modalities are recorded.
For text-heavy scans, a dedicated OCR step followed by a text-only model can be cheaper than sending images. For charts, diagrams and layouts where position carries meaning, a vision model avoids losing that structure.
Frequently Asked Questions
How many tokens does an image cost?
It depends on the provider and the image resolution. Most providers tile or downscale images and bill a fixed token count per tile. Downscaling images to the smallest size that keeps text legible is the most effective way to reduce cost.
Should I use OCR instead of a vision model?
For clean, text-only documents at high volume, OCR plus a text model is usually cheaper. For forms, tables, handwriting or charts, a vision model tends to extract structure more reliably.
Why might a model I know supports images be missing?
The list only includes models whose image support is recorded in our catalog. Missing modality data is treated as unknown rather than assumed.