Solutions · Vision
Best LLMs for vision & multimodal
Charts, PDFs, screenshots, and OCR-grade document parsing.
Vision routes score MMMU, DocVQA, ChartQA, and internal screenshot-to-JSON benchmarks. Use auto:vision when you pass image_url or PDF parts — the router picks the strongest visual reasoner within your latency budget.
| # | Model | Score | MMMU | Why |
|---|---|---|---|---|
| 1 | gpt-5.5 | 96 | 82.4% | Best overall visual reasoner; strong on charts and diagrams. |
| 2 | gemini-3.1-pro | 95 | 81.7% | Best on video + long PDF understanding. |
| 3 | claude-opus-4.5 | 94 | 80.9% | Highest accuracy on complex document layouts. |
| 4 | claude-sonnet-4.5 | 91 | 78.2% | Default vision workhorse; best $/image at production scale. |
| 5 | gemini-3.6-flash | 87 | 74.5% | Cheap, fast vision — ideal for high-volume OCR. |
| 6 | qwen3-vl-235b | 85 | 72.8% | Open-weights vision leader; excellent Chinese OCR. |
| 7 | gpt-4o | 83 | 71.4% | Mature multimodal with audio in/out. |
| 8 | pixtral-large | 80 | 68.6% | Mistral vision; strong EU-hosted document pipeline. |
| 9 | llama-3.2-90b-vision | 76 | 65.1% | Open weights, self-hostable for regulated data. |
| 10 | qwen3-vl-72b | 74 | 63.7% | Cheaper Qwen vision for background enrichment. |
Why these models
- Composite of MMMU, DocVQA, ChartQA, and RealWorldQA.
- PDF and multi-image support required; single-image-only models excluded.
- OCR accuracy scored separately for document-heavy pipelines.
Drop-in example
const res = await client.chat.completions.create({
model: "auto:vision",
messages: [{
role: "user",
content: [
{ type: "text", text: "Extract the totals table as JSON." },
{ type: "image_url", image_url: { url: invoicePngUrl } },
],
}],
});