Skip to main content

ADR-003: Anthropic Claude as Primary LLM

Status: Accepted Date: February 2026

Context​

ReGenesis's AI engine (Sasha) requires a large language model for session analysis, insight generation, evidence pack creation, and 24/7 companion interactions. The choice of LLM provider has significant implications for data privacy, compliance, cost, and output quality.

Decision​

Use Anthropic Claude as the primary LLM, with architecture designed to support fallback to OpenAI GPT-4 or other providers.

Specifically:

  • Primary: Anthropic Claude (latest available model)
  • Fallback: OpenAI GPT-4 (for provider outage resilience)
  • Architecture: LLM-agnostic adapter pattern so any provider can be swapped

Alternatives Considered​

Alternative 1: OpenAI GPT-4 Only​

  • Pro: Widest adoption, most tool integrations, well-documented API
  • Con: OpenAI's data practices have been scrutinized; default 30-day data retention on API; less emphasis on safety/alignment in public positioning
  • Rejected as primary because: Anthropic's Constitutional AI approach and stronger data privacy stance better align with enterprise coaching data sensitivity

Alternative 2: Self-Hosted Open Source (Llama, Mistral)​

  • Pro: Full data control, no third-party data exposure, no per-token cost
  • Con: Massive infrastructure cost, engineering overhead, lower quality for nuanced coaching analysis, harder to maintain and update
  • Rejected because: Startup stage — can't afford to manage ML infrastructure. Quality gap for coaching-specific tasks is significant.

Alternative 3: Multiple Providers by Task​

  • Pro: Best model for each task (e.g., Claude for analysis, GPT-4 for code, Gemini for search)
  • Con: Complexity, more DPAs to manage, more attack surface, harder to audit
  • Rejected because: Simplicity preferred at startup stage; re-evaluate at GA

Consequences​

Positive​

  • Anthropic by default does not train on API data (unless opted in)
  • Constitutional AI approach aligns with coaching safety requirements
  • Strong enterprise privacy positioning
  • Claude's long context window is ideal for processing full session transcripts
  • Fallback to OpenAI provides resilience

Negative​

  • Vendor dependency on a single provider (mitigated by adapter pattern)
  • Anthropic is younger/smaller company — risk of instability (mitigated by OpenAI fallback)
  • Per-token costs for intensive analysis (full transcript processing) can add up
  • Need DPA with Anthropic covering data handling commitments

Cost Considerations​

  • Anthropic Claude API pricing (as of early 2026): approximately $3-15 per 1M tokens depending on model
  • Typical session transcript: 10,000-30,000 tokens
  • With evidence pack generation (multiple passes): 50,000-100,000 tokens per session
  • Estimated cost per coaching session: $0.50-$3.00 in LLM tokens
  • At scale (1,000 sessions/month): $500-$3,000/month in LLM costs

Technical Requirements​

  • Implement LLM adapter interface: generateInsights(transcript, config) → InsightResult
  • Claude implementation as primary
  • OpenAI implementation as fallback
  • Health check: if Claude API returns 5xx for 3+ consecutive calls, auto-switch to fallback
  • Set no-training flags on all API calls
  • Log model version, token usage, and latency for every call
  • DPA with Anthropic securing data handling commitments

References​