What it costs
Nothing goes to me. This is MIT-licensed and it always will be — there's no paid tier, no upsell, and no feature behind a paywall. The Buy Me a Coffee link exists if you're feeling generous, and that's the extent of it.
The only money involved is whatever your chosen AI provider charges for the tokens. You pay only on the weeks you actually play, and there are two ways to get it to zero: run a local model, or draw on a Claude Code or ChatGPT subscription you already pay for.
What a session costs
| Routing | Per session | What it is |
|---|---|---|
| Subscription everywhere | $0.00 | Every phase on a Claude Code or Codex plan you already pay for. |
| OpenRouter prose + subscription grounding | $0.42 | A Claude Code plan on the mechanical phases, OpenRouter on the prose. |
| OpenRouter everywhere | $0.44 | Every phase on OpenRouter, chosen for cost against measured quality. |
| Gemini measured hybrid | $0.78 | Derived from a controlled A/B on a real session rather than from guesswork. |
| Gemini Flash everywhere | $0.98 | The cheapest all-Gemini route. Prose is noticeably less polished than Pro. |
| Gemini Smart Budget | $2.77 | Pro for the chronicle, Flash for the mechanical phases, Flash-Lite to condense. |
| Gemini Pro prose + subscription grounding | $3.60 | A Claude Code plan carries the mechanical phases; Gemini writes the prose. |
| Gemini Pro everywhere | $4.85 | The reference number. Every phase on the strongest Gemini model. |
| OpenRouter, strongest models | $10.39 | The same catalogue, routed for quality rather than for price. |
Figures are USD for a 3-hour session — about 190k characters of transcript with a furnished campaign’s lore in context.
Priced on 2026-08-24 against the live OpenRouter catalogue (411 models), and regenerated every time the site is published.
These numbers are generated, not typed. Every time the site is published, the build prices each routing against the live OpenRouter catalogue using the same estimator the app shows you before a run — so a model getting cheaper shows up here without anyone editing this page.
An estimate is an estimate. It assumes a three-hour session and a furnished campaign's lore in context; a shorter session with a thin glossary costs less, a marathon with a large lore folder costs more. The Run button shows the estimate for your transcript and your routing, which is the number to trust before committing to a long run.
Where the money actually goes
Thinking tokens, mostly. They bill at the output rate, and on the chronicle phase they run at close to the model's ceiling on every call — on a reconciliation against two real bills they accounted for the large majority of the spend.
Turning thinking off is not the saving it looks like. With it disabled, a fast-tier model misattributed dialogue to the wrong character. A cheaper chronicle that puts words in the wrong character's mouth is not cheaper, so thinking stays on in every shipped routing.
The levers that do work, and are already applied:
- Prompt caching. System instructions, lore, and DM answers sit in a cacheable prefix, so the lore is not re-billed on every chunk. This was the single largest cost driver before it landed.
- Chunk sizes keyed to the model tier. A fast-tier model runs chunks about half the flagship size: cheaper per call, and it degrades less on long input, so quality holds at the lower token count.
- Routing the mechanical phases somewhere cheap. Grounding, audit and extras do not need a flagship. Putting them on a cheap model, a subscription CLI, or a local runner is where most of the difference between the rows above comes from.
Getting it to zero
| Route | Cost | Trade-off |
|---|---|---|
| A subscription you already pay for — Claude Code or Codex | No API spend | Metered in rolling windows rather than per token. One session can use a large share of a five-hour window, so budget one or two per window. Exhausting it pauses the run rather than failing it. |
| A local model — Ollama, LM Studio or Unsloth | Electricity only | Prose quality below roughly 15B parameters is noticeably weaker. Mixing — local for the mechanical phases, cloud for the prose — is usually better than going fully local on modest hardware. |
What is never paywalled
Every feature, forever. There is no pro tier, no seat count, no export limit, and no account to create. If a future version adds something, it will be in the same MIT-licensed repository as everything else.
Next
- Choosing a provider — which connection to set up
- Per-phase routing — how to build a cheaper recipe by hand
- Recommended settings — sensible defaults to start from