How many tokens — and how many dollars — does it take to analyze a video with Claude, GPT-5.6, or Gemini? Enter your video below. Formulas and prices are pulled from official docs, verified 2026-08-16.
Estimated, not a quote · prices & formulas last verified 2026-08-16 · GPT-5.6 rows are patch counts (OpenAI hasn't published the patch→billed-token multiplier for 5.6) · Gemini token rates are Google's published approximations — for billing-exact numbers use each provider's count-tokens API.
Claude reads images in 28×28-pixel patches: ceil(w/28) × ceil(h/28) tokens. A 720p frame = 46×26 = 1,196 tokens; 1080p = 2,691 tokens (official worked example). Images beyond 2,576px long edge or 4,784 tokens get downscaled first. Haiku 4.5 uses the older tier (1,568px / 1,568-token cap). Anthropic vision docs
ceil(w/32) × ceil(h/32) patches with no downscaling at default detail. A 720p frame = 40×23 = 920 patches. OpenAI hasn't published the patch→token multiplier for the 5.6 series, so treat results as estimates; the older 512px-tile math applies to GPT-4o/4.1-era models only. OpenAI vision docs
Gemini is the only one of the three that eats video files directly: ≈300 tokens per second at default resolution (258/frame at 1fps + 32/sec audio), ≈100 tokens/sec at low resolution. Gemini video docs
Claude and GPT-5.6 don't accept video — you feed them frames plus a transcript. That sounds like a downside, but it's also the control knob: which frames you keep decides both cost and quality. Scene-cut extraction covers a fast-edited clip with ~10x fewer frames than blind 1fps sampling. Gemini native mode is simpler to call but samples at a fixed 1fps — you pay for every second whether anything happens or not.
Claude 5: 600 images per request (100 on 200k-context Haiku 4.5), 1M-token context. GPT-5.6: pages list 1,500 images / 512MB per request, 1.05M context. Anything longer than ~10 minutes at 1fps needs chunking: per-chunk summaries, then an overall pass. The calculator flags this per model.
A 3-minute 720p reel, scene-cut sampled (≈24 frames), with transcript, on Claude Sonnet 5: 24×1,196 ≈ 28.7k image tokens + ~0.6k transcript ≈ 29.3k input ≈ $0.06, plus a 1k-token answer ≈ $0.01. The same reel at blind 1fps (180 frames) would run ~$0.43 input — 7x more for the same understanding.
No — it's a sourced estimate. Providers' count-tokens endpoints and your actual usage receipts are the ground truth. We label every number that relies on an unpublished multiplier.
Patch math: 69×39 vs 46×26 patches. If your content is legible at 720p, downscale before sending.
Not directly — you transcribe first (Whisper is cheap/free locally) and pay for the transcript text tokens.
Output usually costs 5x input per token but is small in volume; the toggle above adds a ~1,000-token answer so totals aren't systematically understated.
claude-real-video is a free, open-source command that does the whole pipeline — scene-cut frame extraction, timestamped transcript, timeline fusion — and hands your LLM something it can actually cite by the second.
claude-real-video on GitHubSources: Anthropic pricing · Anthropic vision · OpenAI pricing · OpenAI vision · Gemini pricing · Gemini video · model data in models.json · changelog: 2026-08-16 first release. Gemini 3.7 Flash promo pricing runs through 2026-12-31 ($1.50/$7.50 after).
