AI Multimodal Model Comparison, which AI models support vision, audio & video

This is a lookup-and-filter tool, not a calculator. Each of the 14model rows stores a set of boolean flags, does the documented API accept text, image, audio, video or PDF input, and does it produce image or audio output, copied verbatim from that provider's own API guide and cited per row. There is no arithmetic; the “methodology” is careful data compilation plus transparent filter logic.

1. Modality flags.“Support” means the model's documented chat or generate API accepts or produces that modality, not a research demo and not a separate product. A provider's standalone text-to-speech endpoint, for instance, does not make its chat model “audio-out.” Unknown or unpublished support defaults to false, so a required-modality filter never over-promises.
2. Hard limits. Where a doc publishes a numeric cap, images per request, image size, audio minutes, video minutes, context tokens. It is copied exactly with its unit. Where the provider publishes no fixed cap, the value is recorded as unknown and rendered —, never a guessed number.
3. Filter logic. A model survives only if it supports every modality you tick (logical AND): visible = models.filter(m => required.every(k => m[k] === true)). Provider and name filters narrow the set further. Everything runs client-side and instantly, no network calls, no keys.
4. Cross-check. Every filter is evaluated by two independent implementations (a filter pass and a reduce pass) that must return the same model set, and all rows are validated to cite a source and carry no undefined limits. When both agree the card shows a “Cross-checked” badge, the same dual-method credibility idea used in our tax calculator.

The v1 list covers the major public-API models plus flagship open-weight models (Llama 4, Qwen2.5). Niche and long-tail models are deliberately omitted rather than implying exhaustive coverage. For token pricing, benchmark quality or free-tier quotas, see the related tools below, capability, cost and quality are three separate questions.

• OpenAI: Images & vision guide (GPT-4o, GPT-4.1, o-series, GPT-5)
• OpenAI: Realtime API (GPT-4o audio in/out)
• Anthropic: Claude vision & document (image / PDF input, no audio or video)
• Google: Gemini video understanding (native long-video input)
• Google: Gemini audio understanding
• Meta: Llama 4 model cards (vision variants)
• Alibaba: Qwen model cards (Qwen2.5-VL / Qwen2.5-Omni)

Every model row above links its exact provider doc, expand a row to open it. Modality flags and limits were last cross-checked against these sources on 2026-07-03, and the list is reviewed whenever a major model ships.