Skip to content

model-profiles.md

Latest commit

 

History

History
84 lines (71 loc) · 3.07 KB

model-profiles.md

File metadata and controls

84 lines (71 loc) · 3.07 KB
Raw

Model Profiles

Strengths, weaknesses, and routing heuristics for each model in the orchestration pool.

GPT-5.4 (OpenAI) — Default Analyst

  • Alias: gpt5
  • Model ID: gpt-5.4
  • Cost tier: $$$
  • Strengths:
    • Deep structured analysis with chain-of-thought reasoning
    • Finding issues, scoring, categorizing problems
    • Comprehensive critiques with specific line-level references
  • Weaknesses:
    • Can over-compress when rewriting (tends to cut data tables)
    • May smooth over nuance in creative tasks
  • Default role: Primary analyst
  • Route here for: Analysis, structured critique, scoring, identifying problems

Claude Opus 4.6 (Anthropic) — Default Counter-Analyst + Implementer

  • Alias: opus
  • Model ID: claude-opus-4-6
  • Cost tier: $$$
  • Strengths:
    • Deep analytical reasoning and counter-analysis
    • Challenging assumptions — finds what other models miss
    • Precise multi-constraint instruction following
    • Complete, lossless rewrites preserving all data
    • Long-context comprehension
  • Weaknesses:
    • Higher cost per token
    • Can be overly thorough when brevity is needed
  • Default role: Counter-analyst and implementer
  • Route here for: Counter-analysis, implementation, tasks requiring precision

Claude Sonnet 4 (Anthropic) — Lighter Claude

  • Alias: sonnet
  • Model ID: claude-sonnet-4-20250514
  • Cost tier: $$
  • Strengths:
    • Same instruction-following quality as Opus, faster and cheaper
    • Good for medium-complexity analysis and review
  • Weaknesses:
    • Less depth on deeply analytical tasks compared to Opus
  • Default role: Available on request
  • Route here for: Claude-quality analysis when Opus cost isn't warranted

GPT-4o (OpenAI) — Legacy

  • Alias: gpt4o
  • Model ID: gpt-4o
  • Cost tier: $$
  • Strengths:
    • Fast response times
    • Fluent prose and creative rewriting
  • Weaknesses:
    • Aggressively over-compresses when rewriting
    • Less precise on multi-constraint tasks than GPT-5.4
  • Default role: Available on request (superseded by GPT-5.4)
  • Route here for: Quick creative tasks, or when GPT-5.4 is unavailable

Gemini 2.0 Pro (Google) — Validator

  • Alias: gemini
  • Model ID: gemini-2.0-pro
  • Cost tier: $
  • Strengths:
    • Massive context window (1M tokens) — compare large documents side by side
    • Good at cross-referencing and spotting inconsistencies
    • Cost-effective for validation and comparison
  • Weaknesses:
    • Less nuanced on creative writing
    • Shorter max output (8K) limits complex generation
  • Default role: Available for validation and second opinions
  • Route here for: Comparing versions, regression detection, third perspective

Adding a New Model

  1. Add to DEFAULT_MODELS in scripts/orchestrate.py.
  2. Optionally add to ~/.claude/llm-config.json (see references/config-setup.md).
  3. Add a profile section here following the format above (Alias, Model ID, Cost tier, Strengths, Weaknesses, Default role, Route here for).
  4. Run python scripts/orchestrate.py check to verify.