obsidian-penny/docs/api-costs.md

API Costs

PENNY can use multiple LLM providers. Anthropic Claude charges per token, while Ollama runs locally at no API cost. This guide explains how costs work and how to manage them.

How PENNY Uses the API

PENNY makes one API call per annotation, not per chapter. If a chapter has 5 annotations, that is 5 API calls. Each call includes:

1. System prompt -- PENNY's instructions, your voice rules, and the processing template.
2. Context -- Your style guide, voice tests, relevant character sheets, plot outline, and other project files loaded by priority.
3. Chapter text -- The full text of the current chapter.
4. Annotation -- The specific passage and your instruction.

The API response contains only the revised passage.

NOTE and RESEARCH annotations are never sent to the API. They are free.

Token Counts for Different Operations

Token usage depends on how much context PENNY loads. Here are typical ranges:

Input tokens (what you send)

Component	Typical Size
System prompt + template	500-1,000 tokens
Style guide	500-2,000 tokens
Voice tests (matched section)	300-1,500 tokens
Character sheets (1-3 characters)	500-3,000 tokens
Plot outline	500-2,000 tokens
Full chapter text	2,000-10,000 tokens
Wiki/lore entries	0-2,000 tokens
Series bible + themes	0-3,000 tokens
Total per annotation	4,000-20,000 tokens

A typical annotation with moderate context uses 8,000-12,000 input tokens.

Output tokens (what comes back)

Operation	Typical Output
REWRITE (paragraph)	100-500 tokens
EXPAND (add a scene)	300-1,500 tokens
CUT (condense)	50-200 tokens
TONE (adjust passage)	100-500 tokens
DIALOG (rework exchange)	200-800 tokens
PACING (restructure)	200-1,000 tokens
CHARACTER (fix voice)	100-500 tokens
Typical per annotation	100-800 tokens

Thinking tokens

PENNY uses adaptive thinking, which means Claude may use additional tokens for internal reasoning before producing the output. Thinking tokens are billed at the input token rate. Typical thinking usage is 500-3,000 tokens per annotation.

Context Assembly and Token Usage

The biggest variable in cost is context size. PENNY loads context by priority:

Priority	Source	Impact on Cost
1	Full chapter	Always loaded. Longer chapters cost more.
2	Voice tests	Matched section only. Usually small.
3	Style guide	Always loaded. Keep it concise.
4	Plot outline	Always loaded. Bigger outlines cost more.
5	Character sheets	Loaded per character. More characters = more tokens.
6	Wiki entries	Loaded by term match. Varies widely.
7	Series bible	Only if budget permits.
8	Themes	Only if budget permits.

If you want to reduce costs, the most effective lever is the context budget setting. Lowering it from 800,000 to 100,000 tokens means PENNY loads fewer low-priority sources (wiki, series bible, themes) and each call costs less.

Cost Estimates by Model

Current Anthropic pricing (as of early 2026):

Model	Input Price	Output Price	Thinking Price
Claude Opus 4 (claude-opus-4-6)	$5 / 1M tokens	$25 / 1M tokens	$5 / 1M tokens
Claude Sonnet 4 (claude-sonnet-4-6)	$3 / 1M tokens	$15 / 1M tokens	$3 / 1M tokens
Claude Haiku 3.5 (claude-haiku-3-5)	$1 / 1M tokens	$5 / 1M tokens	$1 / 1M tokens

Cost per annotation (estimated)

Using typical token counts of 10,000 input + 2,000 thinking + 400 output:

Model	Input Cost	Thinking Cost	Output Cost	Total per Annotation
Opus	$0.05	$0.01	$0.01	~$0.07
Sonnet	$0.03	$0.006	$0.006	~$0.04
Haiku	$0.01	$0.002	$0.002	~$0.014

Cost per chapter (estimated, 5 annotations)

Model	Cost per Chapter
Opus	$0.20-$0.60
Sonnet	$0.10-$0.30
Haiku	$0.04-$0.10

Cost per full novel revision (estimated, 25 chapters, 5 annotations each = 125 calls)

Model	Cost per Novel Pass
Opus	$5-$15
Sonnet	$3-$8
Haiku	$1-$3

These are rough estimates. Actual costs depend on your chapter lengths, context size, and how much output each annotation generates.

Note: Anthropic pricing changes over time. Check anthropic.com/pricing for current rates.

Ollama (Local Models) -- Free

Ollama runs models locally on your machine. There is no per-token cost. The only costs are:

• Electricity and hardware -- running large models requires a capable GPU. Smaller models (7B parameters) run well on most modern machines.
• Quality trade-off -- local models are generally less capable than Claude Opus for nuanced prose work, but excellent for simple edits.

When to use Ollama

• CUT annotations -- condensing is mechanical. A local model handles it well.
• PACING annotations -- restructuring paragraphs requires less voice nuance.
• Bulk first passes -- process all annotations with a fast local model, then re-annotate the ones that need more finesse and process those with Claude.
• Privacy-sensitive projects -- nothing leaves your machine.

Hybrid routing for cost reduction

With model routing, you can send different annotation types to different providers in the same chapter pass:

Tier	Provider	Cost
Light (CUT, PACING)	Ollama/llama3.2	Free
Standard (TONE, EXPAND, PLOT)	Anthropic/Sonnet	~$0.04/annotation
Heavy (REWRITE, DIALOG, CHARACTER)	Anthropic/Opus	~$0.07/annotation

For a chapter with 2 CUT, 2 EXPAND, and 1 REWRITE annotation, the hybrid cost is roughly $0.15 vs. $0.35 for all-Opus. Over a full novel revision pass, this adds up.

Tips for Reducing Costs

Use model routing

PENNY's model routing is the most effective cost lever. Light tasks (CUT, PACING) rarely need a top-tier model. Set them to Haiku or a local Ollama model and reserve Opus for REWRITE, DIALOG, and CHARACTER annotations where voice precision matters.

Use the right model for the job

• Opus for voice-critical scenes, complex emotional beats, and chapters where precision matters most.
• Sonnet for routine revisions, dialogue fixes, and general prose tightening. Good balance of quality and cost.
• Haiku for bulk first-pass processing, simple cuts, and when you plan to do a follow-up pass with a better model.
• Ollama for free local processing of simple edits and privacy-sensitive work.

Model routing handles this automatically once configured -- you do not need to manually switch models.

Write fewer, better annotations

Five well-written annotations cost less and produce better results than fifteen vague ones. Each annotation is one API call.

Instead of:

%% REWRITE: fix this %%
%% TONE: wrong %%
%% DIALOG: fix the dialogue %%

Write one:

%% REWRITE: the tone is off and the dialogue doesn't match her voice. Rewrite to be casual, blunt, and technically specific. Check the Tim voice test. %%

Reduce context budget

Lower the context budget in settings. A budget of 100,000 tokens still loads the chapter, style guide, voice tests, and character sheets -- the most important sources. It just skips the wiki, series bible, and themes.

Keep your style guide concise

A 500-word style guide loaded into 125 API calls adds up. A 5,000-word style guide costs 10x more in input tokens across the same run. Say what matters. Cut what does not.

Process in batches

Instead of annotating every chapter and running Process all chapters, process 3-5 chapters at a time. Review the results, adjust your approach if needed, then continue. This avoids paying for a full novel pass when your voice rules need tuning.

Use dry runs

PENNY: Dry run shows what would change without making API calls. Use it to verify your annotations are well-formed and targeting the right passages before spending tokens.

Monitoring in the Activity Log

PENNY writes a JSONL log file to your activity log folder (default: .penny-log/activity.jsonl). Each line records:

{
  "timestamp": "2026-04-16T14:30:00Z",
  "book": "book-1",
  "chapter": "ch-05",
  "versionFrom": 3,
  "versionTo": 4,
  "annotationsProcessed": 4,
  "annotationsSkipped": 2,
  "wordCountBefore": 3200,
  "wordCountAfter": 3450,
  "tags": ["REWRITE", "EXPAND"],
  "durationMs": 45000,
  "provider": "anthropic",
  "model": "claude-opus-4-6"
}

Use this to track:

• How many annotations you process per session.
• Which models you use most.
• How long processing takes.
• Word count trends across revisions.

For precise token and cost tracking, check your usage dashboard at console.anthropic.com.