Worked examples, in AI in Markets
Financial Document Token Estimator: Examples
The tokenization difference between providers matters at document scale. Anthropic counts roughly 3.5 characters per token; OpenAI and Google count about 4, so the same 10-K body of 72,000 characters is 20,571 tokens for Claude and 18,000 for GPT and Gemini. These scenarios fix a 10-K body and 2,000-token output, then show how that gap affects per-filing cost across providers. One-pass cost reads the filing once; synthesis re-ingests it with peer filings for a comparative analysis. List rates, no caching.
4 examples. Published May 26, 2026. By AI Fin Hub Research.
Worked examples
See the inputs and outcome together
Each scenario keeps the starting point, the outcome, and the actual lesson in one place so the page reads like a decision notebook, not a data dump.
- 1
One-pass 10-K on a cheap model
Reading a single 10-K body once on Claude Haiku 4.5 to extract a summary. The standard first-pass extraction job.
20,571 input tokens, one-pass cost $0.0306, fits in context.
Document
10-K body (~72K chars)
Output tokens
2,000
Peers
1
Model
Claude Haiku 4.5
A full 10-K reads for three cents on Haiku. The input dominates at $0.0206 because the filing is long and the output is short. For high-volume extraction this is the right tier; the filing comfortably fits the context window.
- 2
Same filing, flagship model
The identical one-pass read on Claude Fable 5.1, the choice when extraction accuracy on dense financials matters.
20,571 input tokens, one-pass cost $0.3057 on Claude Fable 5.1, Anthropic's flagship, fits in context.
Document
10-K body (~72K chars)
Output tokens
2,000
Peers
1
Model
Claude Fable 5.1
Fable 5.1 is ten times Haiku at $0.306 per filing, the ratio of their input and output prices since the token count is identical. The decision is quality against roughly 28 cents per filing, not a context or capability limit.
- 3
Tokenizer difference on the same document
Reading the same 10-K on GPT-5.5, which tokenizes at about 4 characters per token instead of Claude's 3.5.
18,000 input tokens, one-pass cost $0.15, fits in context.
Document
10-K body (~72K chars)
Output tokens
2,000
Peers
1
Model
GPT-5.5
GPT counts 18,000 tokens for the same filing that costs Claude 20,571, because its tokenizer packs more characters per token. The tokenizer alone gives a 12 percent token discount before any price difference; comparing models on price per filing, not price per token, is the only fair comparison.
- 4
Five-peer synthesis
A comparative analysis that ingests the target filing alongside five peer 10-Ks on Claude Haiku 4.5, then writes a synthesis. The realistic competitive-analysis workload.
Synthesis input 123,426 tokens, synthesis cost $0.1334 (versus $0.0306 one-pass).
Document
10-K body (~72K chars)
Output tokens
2,000
Peers
5
Model
Claude Haiku 4.5
Adding five peers multiplies the input six-fold (20.6K to 123K tokens) and roughly quadruples the cost to $0.133 — still cheap but no longer trivial. Multi-filing synthesis is where input volume, and therefore context window, starts to matter for model choice.
Patterns
Run the numbers next
- Generators SEC Filing Chunk Optimizer
Pick a filing archetype, tune chunk size and overlap, and see chunk count, embedding cost, and structural-boundary warnings across three chunking strategies.
- Calculators Earnings-Call Summarization Cost Calculator
LLM cost per stock per quarter to summarize earnings transcripts across Sonnet, Opus, GPT-5.5, Gemini 2.5 Pro/Flash. Cache-hit-rate aware. Snapshot pricing.
Sources
- Anthropic API Pricing , Anthropic (2026)
- EDGAR Full-Text Search and Filing Access , U.S. Securities and Exchange Commission
Keep the topic connected
- Learn Agent-Cost Envelope
The agent-cost envelope: the loop of (calls × tokens × retries × model_price) that determines the dollar cost of an LLM-driven trading agent per decision.
- Learn MCP (Model Context Protocol)
Model Context Protocol: Anthropic's open standard for letting LLMs discover and call tools — the interface, why it matters, and finance MCP server checks.
- Learn LLM Hallucination Detection in Finance
How to detect LLM hallucinations in financial outputs: citation grounding, verifiable-claim checks, and cross-model agreement that flag fabricated data.