These are real production figures from the last ten days of episodes, taken on 31 October 2026. Model prices change every few weeks, so read them as a snapshot of that date, not a fixed price list.
Where the money goes
Every episode starts with a prompt and ends as about 25 minutes of finished audio. In between, a pipeline researches the topic, plans it, has an AI model write the whole conversation, voices every line on rented GPUs, makes cover art, and publishes it. One choice moves the bill more than anything else: which model writes the script. So here are two versions of the same episode, one with a premium writer and one with a cheaper one.
- Script writer (premium or cheaper model)
- Other AI steps (research, planning, tags, metadata)
- GPU voices and orchestration (Modal)
- Web search (Exa)
- Cover art
Only the script writer changes between the two. The GPU, search and cover-art costs are the same either way; they take a bigger share of the cheaper episode because its total is smaller. Each pie is drawn to scale with its episode's cost.
| Part | Claude Sonnet 5.5 | DeepSeek V4.1 Flash |
|---|---|---|
| Script writer (premium or cheaper model) | $0.12 | $0.004 |
| Other AI steps (research, planning, tags, metadata) | $0.025 with either writer | |
| GPU voices and orchestration (Modal) | $0.099 with either writer | |
| Web search (Exa) | $0.019 with either writer | |
| Cover art | $0.011 with either writer | |
| Total | $0.27 | $0.16 |
What each part is
- The script writer. One long call: the model reads a prompt of about 26,000 tokens (the research, the plan, the hosts' character notes, the show's rules) and writes the whole two-host conversation in one go. With Claude Sonnet 5.5, the show's writer as of today, that's about 12 cents. DeepSeek V4.1 Flash does the same job for well under a cent. The gap is the whole difference between the two charts.
- Every other AI step runs on cheap, fast models: cleaning up the prompt, research, planning, tagging, the episode description, the memory the hosts carry between episodes. Together, about 2.5 cents.
- The voices. Every line is rendered by an open-source text-to-speech model on rented L4 GPUs on Modal, two at a time. That's about 8 cents of GPU time, plus about 2 cents for the machine that runs the pipeline end to end. A typical episode takes about 13 minutes from prompt to published (half take between 10 and 25 minutes).
- Web search through Exa, so the hosts work from current sources rather than memory: about 2 cents.
- Cover art, one generated image: about 1 cent.
Not counted: website hosting, audio storage and bandwidth, the database, and the time Daniel spends sending prompts and listening back. At today's pace of about 14 episodes a day, that's roughly $115 a month with the premium writer, or $67 with the cheaper one.
How much one provider could take off
The costs bunch up in two places, so a single sponsor can make a large dent:
- AI model credits (from a model lab, or a router such as OpenRouter) would cover the writer and every other AI step: about 52% of a premium-written episode.
- GPU credits would cover the voices and the pipeline: about 37% of a premium-written episode, and about 63% of a cheaper one.
Together, those two would cover close to 90% of what it costs to make an episode. Credits are also what would let us keep a premium writer on every episode, which is where the conversation gets its life.
How these were measured. AI costs come from what our model router billed for each episode, and GPU costs from the pipeline's own per-episode records. Both are medians over the 139 episodes made between 21 and 30 October 2026. The non-writer AI steps (about $0.025) are what an episode costs in total when its writer costs well under a cent. The Sonnet writer figure is its list price, because it became the default writer today. The DeepSeek figure is measured on the episodes it wrote. The search figure is the mean of the last 17 runs.