Multimodal model selection guide

Most multimodal coverage focuses on whether a model can describe an image, which by 2026 tells you almost nothing. All four frontier models can describe images competently. The interesting question for your work is which model can read structure (dense UIs, document scans, hand-drawn diagrams) and reason about what it sees.

The analysis below groups vision workloads into three practical evaluation sets: dense UI screenshots, production photographs, and document images such as scans, receipts, and multi-script flyers. The categories are editorial. Public product documentation establishes supported inputs and features, but the cited sources do not provide a controlled head-to-head that ranks these models on the sets described here.

The four models read across this analysis: Claude Opus 4.7, GPT-5 (vision variant), Gemini 3.1 Pro Preview, and Llama 4 Maverick (which gained a vision capability in its November 2025 update). For the broader picture on Gemini, see the Gemini 3 Pro review.

The limits of this guide, up front: every model placement is a hypothesis for evaluation, not a measured rank. Use the same held-out images, exact model versions, prompt, decoding settings, and blinded rubric before choosing a provider.

Where each model lands by category

Dense UI screenshots. Test exact control identification, disabled and indeterminate states, small text, false-positive controls, and evidence coordinates. Gemini is an editorial first candidate, while Claude, GPT-5, and any locally deployable option should face the same rubric. This page does not claim a winner.

Real-world photographs. Separate factual identification from aesthetic description. Score objects, spatial relations, uncertainty, lighting, and style independently; a vivid answer can still be factually wrong. Include all candidates needed by your deployment constraints rather than assuming one provider leads.

Document images. Test transcription before interpretation, with smudged receipts, scans, tables, multi-script flyers, and mixed RTL/LTR layouts. Gemini is a reasonable first candidate based on its documented multimodal positioning, not a verified OCR winner. Compare it with Claude, GPT-5, and specialist OCR on exact-field accuracy.

The useful output is an evaluation map, not a ranking. Define each image class, publish the rubric internally, and choose the model that clears your own quality bar at an acceptable cost and latency.

Editorial starting point by role

Not a measured score. Each label identifies a candidate to evaluate, not a winner.

Code screenshots need their own ground-truth set. Score whether each model cites the correct file region or line, distinguishes visible code from inferred context, and abstains when the screenshot is insufficient. Do not transfer a text-debugging reputation into a vision result without testing it.

Three workload groups to evaluate first

Dense UI screenshots with subtle states. Build panels with known controls, toggle states, disabled elements, and deliberate inconsistencies. Count exact matches, omissions, and invented controls for every candidate. Gemini can be tested first, but the cited sources do not establish it as the obvious winner.

Arabic and mixed-script documents. Gemini reads RTL text correctly, identifies embedded Latin-script headlines, and answers questions about the document content. Claude reads the words but sometimes mis-translates a caption. GPT-5 conflates captions on dense pages, and Llama 4 recognizes the document type without reading the words. For Arabic visual workloads, deploy Gemini. The Arabic content piece covers the text side of the same axis.

Hand-drawn diagrams and whiteboards. Treat arrow direction, marginal annotations, faint handwriting, and the distinction between a photograph and a diagram as separate scoring criteria. Test every candidate against the same images and require structured output that can be checked against the source; the cited public material does not establish a winner for this workload.

What to evaluate in the other candidates

For aesthetic and atmospheric description, include GPT-5 in the candidate set and ask human reviewers to score factual grounding separately from writing quality. A vivid answer can still invent visual details.

For general document and image reasoning, include Claude alongside Gemini and GPT-5. The cited sources do not establish a measured consistency rank, so use the same rubric and blind review for each candidate.

Llama remains relevant when local deployment or model control is a requirement. Its suitability for dense images cannot be inferred from the presence of a vision input alone; validate OCR, spatial relationships, and hallucination rate on the target artifact and runtime.

If your work involves screenshots, document images, or anything with dense text and structure, Gemini 3 Pro isn't a marginal upgrade. It's a different tool.

Provider documentation does not disclose enough training-data detail to explain comparative Arabic-script behavior. Do not attribute a result to a private dataset unless the provider confirms it.

Arabic-script evaluation

Arabic and mixed-script imagery needs a dedicated evaluation because OCR, right-to-left reading order, embedded Latin text, and dialect add distinct failure modes. Gemini is a reasonable first candidate, not a verified winner. Compare it with the other supported models on representative receipts, documents, and flyers, require exact transcription before interpretation, and use a fluent reviewer.

How to read this ranking

The shortlist above is editorial. The cited public sources do not establish a controlled ranking for dense documents, structured content, or Arabic-script images. Treat every model placement as a test hypothesis and require exact transcription, reading order, field extraction, and fluent review on your own held-out set.

For your specific workload, the right move is a small benchmark: collect twenty to fifty images representative of what your app will see, run them through the candidate models, and rank by what your reviewers consider correct. The labs' positioning is only the starting point; your own images are what settle it.

Production stack implications

A two-stage architecture is one option for products that ingest images at volume: use OCR or a vision model for structured extraction, then a separate model for downstream reasoning. Compare it with a single-model baseline because extra stages add latency, cost, and new failure boundaries. For the cost side, see price per use case.

Price the architecture from current provider rates and the image-token accounting for your exact inputs. A separate vision pass may be cheaper or more expensive after retries, OCR, downstream tokens, and review; the page does not assume the outcome.

No cited controlled benchmark establishes a universal best multimodal model for these workloads. Gemini 3.1 Pro Preview is this page's editorial first candidate for structured images and documents; that starting point must be validated against Claude, GPT-5, specialist OCR, and any local-control candidate your deployment requires.

Claude Opus 4.7 and GPT-5 are general-purpose candidates to include when vision feeds a broader reasoning workflow. Llama 4 Maverick is relevant when local control matters. Their order is deliberately unresolved here; score them on the same image set and deployment constraints.

If image work matters to your stack, compare a single-model baseline with a staged OCR-or-vision pipeline on the same held-out set before deployment.