Six months ago, running video through an AI model meant either encoding it down to key frames and losing temporal context, or paying per-second pricing that made anything longer than a minute or two prohibitively expensive. Neither approach was great for production work.
That constraint is dissolving faster than anyone predicted. Claude Fable 5 can process full video without frame sampling. Sonnet 5 handles longer audio without breaking it into chunks. GPT-5.5 can work with entire documents as scanned images instead of extracted text. And behind the scenes, model architecture improvements are lowering the token cost per second of video and per megabyte of audio across the board.
Why this matters beyond "we can use video now"
Cheaper multimodal input fundamentally changes what's economically viable to automate. If video analysis cost $5 per minute, you run it on exception cases. If it costs $0.15 per minute, you run it on everything. That's not a small shift — it's a tier change in what problems you can tackle with AI without rebuilding your whole cost model.
What the numbers actually look like
Anthropic published a cost breakdown this week for video processing via the Fable 5 API. A ten-minute video of typical screensharing — someone working through a document, moving between windows, making edits — tokenizes to around 18,000 tokens when sampled at standard frame rates, or roughly 1,800 tokens per minute. At the current Claude Fable 5 price of $3 per million input tokens, that's a bit over $0.05 per minute of video.
For context, transcribing the same ten-minute video via a traditional speech-to-text API costs somewhere between $0.10 and $0.25, and you lose the visual context entirely. For document digitization, OCR via Tesseract is cheaper per page, but it doesn't handle messy scans, handwriting, or complex layouts anywhere near as well as a vision model does.
The use cases that unlock at this price point
Quality assurance and testing. Watching a user click through a test scenario and reporting what they see is now cheap enough to do in the background during QA runs. The model can flag visual bugs, compare screenshots against baselines, and check for regression without anyone writing custom image-comparison logic.
Field work documentation. Contractors and technicians can record video of a job site or completed work, and the model can extract material lists, note deviations from spec, and flag safety concerns without a human review loop for every single site.
Accessibility and content tagging. For anyone managing video libraries, running every video through a multimodal model to generate captions, extract key moments, and tag content is now in the realm of economically reasonable, not just technically possible.
Customer support and incident analysis. If a customer records a screen recording of a bug happening, it's now feasible for your support system to extract the exact issue, reproduce the conditions, and draft a response without a human triager reading the description first.
What teams usually don't plan for
The cost drop is real, but it's not distributed evenly. Some of the efficiency gains come from architecture improvements that don't transfer across models — Fable 5's video tokenizer doesn't work the same way GPT-5.5's does. If you're planning a multimodal-heavy feature and want to avoid vendor lock-in, you need to test cost and latency across at least two models before betting the product on one.
Also, "cheap enough to do at scale" and "cheap enough to do frivolously" are still different bars. At five cents per minute, you can analyze a hundred hours of video content for a few hundred dollars, which changes the ROI math for a lot of problems. But if you're processing terabytes of video daily, you can still move the needle meaningfully by optimizing which frames matter and which are noise — the cheaper input is additive to thoughtful engineering, not a replacement for it.
The hiring angle
If your product is about to shift from "video analysis was too expensive, so we didn't" to "video analysis is now our baseline," that's an architecture decision with real infrastructure implications. We've written about how to hire for agentic AI work and infrastructure for modern AI systems, and multimodal systems fall somewhere in between — they need someone who understands both the economics of bulk data processing and the particular quirks of vision models. That's a specific set of skills worth scoping carefully if you're staffing for it.
