What the Claude API costs per YouTube video

In our runs, a full analysis of a 20-minute video cost $0.25 to $0.34 on Claude Opus 5.5. The same tokens at list prices come to $0.13 to $0.17 on Sonnet 5.5 and under 1 cent on Haiku 5.5. Length matters more than anything else except the model.

How far this goes. This is a small sample: 7 real runs on 5 videos. We ran Opus on three videos but have not yet run Sonnet or Haiku on them, so those numbers show price and say nothing about quality. Treat them as a guide.

Every model, per 20-minute video

Model20-minute videoBased on
Claude Opus 5.5$0.25 to $0.34measured
Claude Sonnet 5.5$0.13 to $0.17recalculated
Claude Haiku 5.5Under 1 centrecalculated
GPT-6.1 Sol$0.07 to $0.13measured
GPT-6.1 TerraAbout $0.05measured
GPT-6 LunaUnder 1 centmeasured
Grok 4.7About $0.24measured
Grok 4.3About $0.04measured
Gemini 4 ArgonAbout $0.22estimate
Gemini 3.8 FlashAbout $0.09estimate
Gemini 3.1 Flash-LiteAbout $0.02estimate

Measured means a real run on that model, scaled to 20 minutes. Recalculated applies that model's list price to the token counts of our Opus runs. Estimate applies that model's list price to our GPT-6.1 Sol token counts.

The runs behind it

Each run makes five calls per video: chapters, per-segment analysis, summary, facts and grouped insights. Costs are at list prices on 5 November 2026.

VideoLengthModelTimeInputCache writeCache readOutputCost
Marketing talk (English)28:35Claude Opus 5.571 s4,40026,63439,9519,975$0.358
Q&A (German)39:22Claude Opus 5.5not logged6,45852,25678,38418,258$0.668
Interview (German)53:47Claude Opus 5.5142 s7,40471,268106,90223,355$0.874
Kurzgesagt, immune system7:49GPT-6.1 Sol44 s5,40605,6944,049$0.052
Kurzgesagt, immune system7:49GPT-6.1 Terra29 s3,31107,5923,390$0.021
Kurzgesagt, immune system7:49GPT-6 Luna28 s3,23807,5924,907$0.0029
Linux Foundation keynote15:08GPT-6.1 Sol38 s5,712016,2564,143$0.054

Download the numbers as CSV, with the Sonnet and Haiku recalculations.

Why output is the expensive part

On Opus 5.5, Anthropic charges $4 per million input tokens, $5 to write the prompt cache, $0.20 to read it and $20 for output. Our 53-minute run used 23,355 output tokens, which cost $0.47 of its $0.87 total.

The transcript is sent with all five calls, but after the first it's read from the cache at a twentieth of the input price. Without caching, that same run would cost about $0.34 more.

One trap for very long videos: Haiku 5.5 costs five times as much once a prompt goes over 100,000 tokens. We haven't measured a video that long.

API or subscription?

These are separate. Claude Pro includes no API credits. Max includes $100 a month and Ultra $200, which at our rates would cover hundreds of videos a month. How credits and plans work.

What we noticed about cheap models

On one 7:49 video we compared GPT-6.1 Sol ($0.05), Terra ($0.02) and GPT-6 Luna ($0.003). Luna kept concrete details from the video and we saw no obvious quality gap. Sol added one hedge the video didn't contain. That's a single video, so don't treat it as a ranking.

All 11 models are available in Klaara, and it shows the estimated cost for the one you pick.

Klaara for Mac

Klaara is a Mac app that turns YouTube videos into summaries with chapters, using your own AI API key. It costs €19 once. 3 free summaries first.

Download for Mac

Related