Claude: Anthropic Constitutional AI

Claude is Anthropic family of AI assistants, built using Constitutional AI training methods.

Claude 1 (March 2023)

  • Context window: 9,000 tokens
  • Training method: Constitutional AI CAI

Claude 2 (July 2023)

  • Context window: 100,000 tokens 10x Claude 1
  • Bar exam: 76.0 percent 70th percentile
  • Code generation: 71.2 percent on HumanEval

Claude 3 Family (March 2024)

Claude 3 Haiku

  • Speed: Fastest Claude model
  • Cost: Most affordable 0.25 dollars per million tokens
  • Context: 200K tokens

Claude 3 Sonnet

  • Balance: Best cost-performance ratio
  • Cost: 3 dollars per million tokens

Claude 3 Opus

  • Capability: Most intelligent Claude model
  • Cost: 15 dollars per million tokens
  • HumanEval: 93.7 percent

Claude 3.5 Sonnet (June 2024)

  • MMLU: 88.7 percent exceeds GPT-4 Turbo
  • HumanEval: 92.0 percent excellent coding
  • Speed: 2x faster than Claude 3 Opus
  • Context: 200K tokens

Which Claude Model for What?

  • Real-time chat: Claude 3.5 Haiku fastest cheapest
  • Document analysis: Claude 3.5 Sonnet best value
  • Complex research: Claude 3 Opus most capable
  • Code generation: Claude 3.5 Sonnet 92 percent HumanEval

SEBI Disclaimer

This article is for educational purposes only. Trading involves substantial risk.

Reading Financial PDFs Without the Headache

Claude models, especially 3.5 Sonnet, became the default tool for Indian retail analysts because of their document handling. A quarterly results PDF with a 40-page balance sheet, a chairman's letter and six annexures is exactly the workload the model is strongest at, with two practical conditions:

  • Upload the table-heavy pages as images or native PDF pages rather than text-dump, because the vision path handles Indian numbering (lakhs, crores) more reliably.
  • Ask for a reconciliation table: figures as printed, and the computed ratios with formulas, so you can spot the model's unit errors.

Where Claude 2 struggled with multi-page tables bleeding into each other, 3.5 Sonnet holds the coherence window across long annexures, which is the single biggest reproducibility win for report reading.

Claude in Agentic Trading Loops

Claude models are well suited to the browser- and tool-use pattern of an LLM agent: reading a dashboard, comparing two chains, drafting an execution checklist. The 2024-era training is particularly strong at:

  • Structured JSON output with low schema-violation rates.
  • Multi-step booking sequences when prompted with explicit state.
  • Honest refusal on missing data, which beats a confident guess in an automated loop.

The practical caution is token cost: long PDF ingestion in an agent loop burns context quickly, so slice documents to the specific factor you need before letting the agent hold them in memory.

API Pricing and Latency From India

Hosted Claude endpoints are reachable from Indian data centres, but two costs shape the budget for a trading-side workflow:

  1. Input and output token pricing differ sharply; a daily batch job that reads 200 pages costs nearly all of its bill in input tokens.
  2. Latency on long outputs climbs to tens of seconds during peak hours; schedule research batches off-market hours, and keep only short verification calls inside the trading session.

For a typical retail quant running one daily report, the Claude API bill stays below the monthly cost of a mid-tier market data subscription, making it cheap relative to the analyst hours it replaces.

An Earnings-Call Prompt Pattern

Earnings calls are the highest-value unstructured input in Indian equities. A copy-paste pattern worth reusing:

  1. Extract the transcript text, then ask Claude to list every number with its context: revenue, margins, debt, guidance.
  2. Ask it to separate management rhetoric from verifiable statements.
  3. Generate three scenario summaries: beats guidance, meets guidance, misses guidance, each with the hedge implied by Q&A.
  4. Score the call's tone on a 1-5 scale and compare the tone score against the subsequent options market move to calibrate the model for future calls.

Constitutional Guardrails and Advice Refusal

Claude's constitutional training produces a predictable boundary: it will analyse a position but hesitate to issue unhedged "buy now" directives. Treat this as a feature:

  • Design prompts that request the analysis and the risk framing, not the verdict.
  • When the model refuses, rephrase into a factual query it will answer, then draw your own conclusion.

The refusal behaviour effectively forces a human checkpoint, which is the right design for a market where the model will never carry the liability of the trade.

Building a Tiny Claude Bench on NSE Data

Before giving Claude real-money work, spend a weekend building a ten-question bench from data you already possess. Take three older Nifty option chains, three earnings-call transcripts and four SEBI or company filings, and ask the model exactly what it will face live: extract the put-call ratio, compute a spread's maximum loss, separate management rhetoric from verifiable numbers, and pull the rupee-denominated figures out of a messy annexure.

Score on exactness, not fluency, and run each prompt twice. A model that answers beautifully but flips its answer between runs is worse than a slower, consistent one, because an automated loop cannot tell the difference. Re-run the bench after every Anthropic release or API version change, since two 'Claude 3.5' aliases on the same day can behave differently.