Claude model migration — retired IDs, new versions, and behaviour deltas

# Find every hardcoded Claude model ID in the repo
rg -n '"claude-[0-9a-z-]+"' \
   --type py --type ts --type js --type go --type rb --type yaml --type json

# Find bare model references (aliases without a version)
rg -n 'model["\s]*[:=]\s*["\']?claude' \
   --type py --type ts --type js

# Common patterns in configs
rg -n 'CLAUDE_MODEL\|ANTHROPIC_MODEL\|OPUS_MODEL\|SONNET_MODEL' \
   --type py --type ts --type js --type yaml --type env

# Once you have the list, consolidate to a single constants file
# constants/models.py:
#   OPUS   = "claude-opus-4-7"          # or a pinned @date if you need stability
#   SONNET = "claude-sonnet-5"
#   HAIKU  = "claude-haiku-4-5-20251001"
#   FABLE  = "claude-fable-5"

# Replace every hardcoded ID with the constant. Now migration is a one-file change.

"""Run a prompt suite against old + new model and diff results."""
import json
import time
import anthropic
from pathlib import Path

client = anthropic.Anthropic()

OLD_MODEL = "claude-3-5-sonnet-20241022"   # currently in production
NEW_MODEL = "claude-sonnet-5"              # target

def run(model: str, prompt: dict):
    t0 = time.time()
    r = client.messages.create(
        model=model,
        max_tokens=prompt.get("max_tokens", 1000),
        system=prompt.get("system"),
        messages=prompt["messages"],
    )
    elapsed = time.time() - t0
    return {
        "model": model,
        "text": r.content[0].text if r.content else "",
        "output_tokens": r.usage.output_tokens,
        "input_tokens": r.usage.input_tokens,
        "elapsed_s": round(elapsed, 2),
        "stop_reason": r.stop_reason,
    }

# Load your prompt suite
prompts = json.loads(Path("prompt_suite.json").read_text())

results = []
for p in prompts:
    old = run(OLD_MODEL, p)
    new = run(NEW_MODEL, p)
    results.append({
        "name": p["name"],
        "old": old,
        "new": new,
        "diff": {
            "output_tokens_delta": new["output_tokens"] - old["output_tokens"],
            "elapsed_delta_s":    round(new["elapsed_s"] - old["elapsed_s"], 2),
            "same_stop_reason":   old["stop_reason"] == new["stop_reason"],
        },
    })

# Summarise
total_old = sum(r["old"]["output_tokens"] for r in results)
total_new = sum(r["new"]["output_tokens"] for r in results)
print(f"Output tokens: old={total_old}, new={total_new}, delta={total_new-total_old:+d}")
print(f"Avg latency: old={sum(r['old']['elapsed_s'] for r in results)/len(results):.2f}s, "
      f"new={sum(r['new']['elapsed_s'] for r in results)/len(results):.2f}s")

# Manually review the diff on outputs where old vs new differ substantially

"""Strategy: pin versions in prod, use aliases in dev/staging."""
import os

# Environment-aware model selection
ENVIRONMENT = os.getenv("APP_ENV", "development")

MODEL_CONFIG = {
    "production": {
        # Pinned — behaviour stable, migration is a deliberate PR
        "opus":   "claude-opus-4-7-20250514",
        "sonnet": "claude-sonnet-5-20251015",
        "haiku":  "claude-haiku-4-5-20251001",
    },
    "staging": {
        # Alias — validates that a future rollout of the new version works
        "opus":   "claude-opus-4-7",
        "sonnet": "claude-sonnet-5",
        "haiku":  "claude-haiku-4-5",
    },
    "development": {
        # Alias — fastest to try new features
        "opus":   "claude-opus-4-7",
        "sonnet": "claude-sonnet-5",
        "haiku":  "claude-haiku-4-5",
    },
}

def model(tier: str) -> str:
    return MODEL_CONFIG[ENVIRONMENT][tier]

# Usage
response = client.messages.create(
    model=model("sonnet"),
    max_tokens=500,
    messages=[...],
)

# Migration path:
# 1) Run test suite against latest alias in staging weekly
# 2) When alias behaviour changes, decide: adopt or hold
# 3) On adopt, bump the pinned version in production and deploy