How much does it cost to make a video with Claude?

The numbers

We built 72 long films and 90 Shorts with Opus 5.5 in Claude Code and summed the tokens of every session. One subagent built each documentary film, so each transcript is exactly one video. Prices are Anthropic's API list rates for Opus 5.5: $4 per million input tokens, $20 per million output tokens, $5 per million for 5 minute cache writes and $0.20 per million for cache reads.

Our sessions actually ran on a Claude subscription, so what we paid was lower than these API figures. On a subscription the cost shows up as usage limits rather than a bill, but the token counts are the same, and they tell you how much work a video really is.

Where the money goes

The surprise is that writing the video is a small part of the bill. An average documentary film session used about 150,000 output tokens, which includes Claude's thinking, but re-read 33.6 million tokens of cached context and wrote 1.66 million tokens into the cache. At list prices that is about $3.00 of output (17 percent), $6.72 of cache reads (37 percent) and $8.28 of cache writes (46 percent).

In plain terms, each turn of an agent loop re-reads the whole conversation so far: the system prompt, your instructions, every file Claude opened and every tool result. By the hundredth turn that is about 260,000 tokens per turn. The cache makes re-reading cheap, but not free, and a cache entry only lives five minutes.

Why renders make it expensive

A final render of a 2.5 minute film takes several minutes; a 3D film took 634 seconds on our Mac for 147 seconds of video. If Claude waits for that render inside its turn, the five minute cache expires, and the next turn writes the entire context back into the cache at the write price. That is roughly a dollar each time at 200,000 tokens of context. Our single most expensive film, at $85.81, was this pattern repeated across 272 turns.

Failure adds a quarter on top. Across the programme, one film in fifty was abandoned and another blew up, 14 of 50 needed a paid narration fix after a fact check, and every one of the 90 Shorts was redone once after a layout rejection. Budget for it.

The other costs: voice, music, sound, render

Everything except Claude is small, at November 2026 list prices.

The last row is why code-drawn video is cheap. A video model bills per generated second; a page drawn by code bills nothing per frame. Mixing in a few seconds of generated footage for a hero shot is affordable; generating every shot is not.

How to cut the cost of an agentic video

• Run renders outside the Claude turn, in a separate terminal or a background job, so the cache never sits idle for five minutes. This attacks the 46 percent that is cache writes.
• When a human approval sits between turns, use the 1 hour cache instead of the 5 minute one. It costs more to write but survives the wait.
• Keep a shared kit of helpers and a fixed contract between videos. Our second run of documentary films used a shared kit and cost 26 percent less on average, even with more turns.
• Start each video in a fresh session with a short brief, rather than continuing a long conversation. A forked session that inherited a huge parent context cost us $7.45 for a 25 second promo.
• Approve a storyboard and stills before the full render. Most expensive failures are visible in a still.
• Do not expect much from switching the agent to a cheaper model. Opus 5.5 and Sonnet 5 have the same cache read price, so we estimate the saving in an agent loop at about 10 percent, not half.

Why a template engine costs dozens of times less

The agent loop is expensive because Claude does everything, including waiting and checking. A template engine splits the work: Claude writes the script and a storyboard as structured data once, which is a few thousand tokens, and deterministic code does the rest. The styles are already written and tested, so there is no film code to debug. Rendering, polling, loudness, sync and frame checks are code, not tokens. The context tax disappears.

That is how Video Studio works, and why a 30 second video costs us about $0.10 to $0.15 to make instead of several dollars. The trade-off is flexibility: an agent can invent any look you describe, while a template engine renders the styles it has. For a channel that posts a consistent style every day, that trade usually favours templates. For a one-off launch film, an agent session can be worth its $10.