RXed AI News

What it is — in one coffee-break

Multimodal means the model handles images, audio or video directly, rather than having someone convert them to text first. You upload a photo of a supplier's delivery note and it reads the note. You drop in a 45-minute recorded call and it answers a question about minute 31. You type a description and it renders video frames.

The word doing the work is natively. Plenty of products claim multimodal because they run Tesseract on your PDF and feed the resulting text to a language model. That pipeline loses layout, table structure and handwriting. A native model looks at the pixels, and on messy scans the gap is measurable: Parsli's 2026 benchmark puts Gemini 2.5 Pro at 94% on scanned invoices against 82% for AWS Textract on line-item extraction and 80–85% for Tesseract 5 on degraded inputs.

When you actually need it (and when you don't)

You need it when the source material is not text and never will be. Photos from the shop floor, the warehouse or a site visit. Scanned paperwork from suppliers who still fax. Recorded sales and support calls. Product photography and short-form video you would otherwise pay an agency for. In each case the alternative is a person retyping, and that person costs €25–40 an hour.

You do not need it when your data is already clean digital text — invoices arriving as structured XML, CRM exports, spreadsheet rows. Paying image-token rates to read a PDF that already has a text layer is money set on fire; extract the text layer and use a cheap text model. You also do not need it for one-off jobs. Ten photos a month is a job for a person, not a pipeline. And you do not need generative video at all unless video is already a channel you sell through.

How to recognize good vs bad implementations

Good implementations accept your file, in your format, at your quality, and tell you when they are unsure. Bad ones work beautifully on the vendor's curated sample and fall apart on a 200-dpi scan with a coffee ring on it.

Three questions expose a weak implementation inside five minutes of any demo. "Can I upload my own file right now — not the one in your deck?" A vendor who needs to "prepare" your file is showing you a demo, not a product. "Is this native, or are you running OCR or transcription first? Show me the API call." If there is a separate transcription step in the trace, you are paying multimodal prices for a single-modality product. "What confidence score comes back with each extracted field, and what happens when it is low?" A system with no confidence output has no safe place to put a human, which means every output needs checking, which erases the saving.

What this costs

Reading is cheap. Per Google's Gemini API pricing, Gemini 2.5 Flash charges $0.30 per million tokens for text, image and video input and $1.00 per million for audio — audio is consistently the expensive modality across the table. Document extraction lands near $0.17 per 1,000 pages on Gemini Flash; transcription through OpenAI runs $0.006 per minute, or $0.36 an audio-hour.

Generating is where budgets die. Runway charges 12 credits per second of Gen-4.5 video: the $12/month Standard plan's 625 credits buy about 52 seconds, and the $76/month Max plan's 9,500 credits buy roughly 13 minutes. Image generation on Gemini runs $0.045–$0.151 per image.

Self-hosted, the reading side is essentially free. Whisper's code and weights are MIT-licensed; a roughly $600 Mac mini pays for itself after about 1,670–2,500 audio-hours versus the API. Qwen3-VL 7B reads documents in around 5 GB of VRAM quantized, which an RTX 3060 handles. Renting instead, an RTX 4090 goes for $0.20–$0.48 an hour. Self-hosted video generation at Runway's quality is not realistically available — that one you rent.

Where to see it scored

Sixty-two tools were audited and 61 carry a score on this element. Four sit at the top with 9.5: Runway (9.5) and Krea (9.5) for generation, Gemini / Google (9.5) for ingesting every modality through one API, and Freepik / Magnific (9.5) for image work.

The bottom of the list is more useful. Zapier AI (4) scores low not because it is a bad product but because it is a text-and-JSON automation layer that hands files to other services — if your workflow starts with a photo, Zapier is the plumbing, not the eyes. DeepSeek (2.5) is the clearest case: strong, cheap text reasoning, effectively no native handling of anything else. Neither score is a verdict on the tool. It is a verdict on this one column, which is exactly how the table is meant to be read.