What does a script actually cost on the Claude API?
Most quoted figures for AI script writing are a single number with no pipeline behind it. This works the cost the way it actually accrues: every generation stage costs input tokens and output tokens, retries multiply both, and the model you pick changes the token count as well as the rate. Every rate below is editable, because the published price is a starting point and yours may differ.
Introductory pricing, published as running through 31 August 2026. From 1 September 2026 the published rate is $3 / $15 per million tokens, which is roughly a 50% increase on the figures used here.
Result
- API cost per script
- $0.112
- $0.022 input, $0.090 output.
- Cost per 1,000 finished words
- $0.075
- Estimated output tokens per script
- 8,970 tokens
- Across 3 stages, including retries.
- Estimated input tokens per script
- 11,213 tokens
- Where the cost goes
- 80% output
- Output dominates, so shorter scripts or fewer stages move the bill more than trimming your prompt.
- Monthly cost at 20 scripts
- $2.24
What this assumes
Each of these is a choice that moves the answer. They are listed so you can disagree with a specific one rather than with the total.
- Tokens are estimated from word count at roughly 0.75 words per token, the ratio Anthropic's own pricing documentation gives for English. Your real ratio moves with formatting, proper nouns and non-English text.
- Each generation stage is modelled as producing output roughly the length of the finished script. A pure outlining stage produces less than that and a polish pass produces about the same, so a three-stage pipeline is costed as three script-lengths of output. If your stages are lighter than that, this overestimates.
- Models from Claude 4.7 onward use a newer tokenizer that produces approximately 30% more tokens for the same text. Selecting one of those models scales the token estimate accordingly, so the comparison between models is like-for-like on the text you produce rather than on raw token counts.
- The retry allowance is applied to input and output alike, on the basis that a retry re-sends the prompt as well as regenerating the answer.
- Cached input is priced at one tenth of the base input rate, per the published prompt-caching multipliers. Cache writes are not modelled, so a pipeline that writes its cache more often than it reads it will cost more than this shows.
- The batch option applies the published 50% discount to input and output. It is only valid if your pipeline can tolerate asynchronous processing, which a same-day publishing workflow usually cannot.
What the figure covers
Counted
- Input tokens for every generation stage, including the prompt and any context you pass
- Output tokens for every generation stage
- Retries, at the rate you set
- The prompt-caching discount on whatever share of input you mark as reused
- The batch processing discount, where you select it
Not counted
- Voiceover, images, music, stock footage and editing: this is the script stage only
- Research, titles, thumbnails and description writing, unless you count them as generation stages above
- Any subscription you already pay for, including a Claude consumer plan
- Your own time reviewing and rewriting the output, which is usually the largest real cost
- Tax, and any regional or data-residency pricing multiplier your account is subject to
- Cache write charges, and any server-side tool use such as web search
On comparing this to a subscription
There is no built-in comparison against a monthly chat subscription here, because the two are not measuring the same thing: a consumer plan is rate-limited rather than token-metered, cannot be driven by a script, and its limits are not published as a token figure that could be divided into a per-script cost. Any calculator that prints such a comparison has picked numbers to make it work. Instead, enter whatever subscription you are weighing this against and the calculator tells you how many scripts a month it would take for the API to be the cheaper option.
Where the rates come from
Rates last checked against the source , and re-checked monthly. Published prices change without notice, so check the source below before relying on a figure from this page.
- Anthropic: Claude model and feature pricing · accessed
Every rate on this page is editable in the form, because the published price is a starting point and your account may not be on it. How figures are sourced, how often they are re-checked and what is explicitly not tested is set out on the methodology page.
Read next
- What "own your stack" actually means
Owning your content stack means holding the API keys, the prompts, and the logic yourself, so no vendor can price you out or shut you down.
Terms used here:TokenInference costContext windowBYOK
Ultra-Long-Form YouTube Script Engine
Generate structured YouTube scripts of 15,000 words or more using your own Claude API key. The engine writes the script section by section, carries context forward between chapters, and helps reduce the repetition and narrative drift that often appear in very long AI-generated scripts.
Get the build guides as they go out.
One email when a new guide ships. The full method, not a teaser. Unsubscribe whenever.