model-audit

Audit every AI model call in the codebase. Output a routing table with real cost math, not just model names.

Phase 1 — Find model configuration

Search for where models are defined:

grep -r "gpt-4\|claude\|gemini\|anthropic\|openai\|mistral\|groq\|llama" \
  --include="*.py" --include="*.ts" --include="*.js" --include="*.env*" \
  -l 2>/dev/null | head -30
grep -rn "MODEL\|model_config\|AI_MODEL\|LLM_MODEL\|OPENAI_MODEL\|ANTHROPIC_MODEL" \
  --include="*.py" --include="*.ts" --include="*.env*" -l 2>/dev/null

Read every file returned. Map alias → actual model string. Example: AI_MODEL="claude-sonnet-4-6" or const FAST_MODEL = "gpt-4o-mini".

Phase 2 — Find every AI call

# Anthropic
grep -rn "messages\.create\|anthropic\.messages\|client\.messages" \
  --include="*.py" --include="*.ts" --include="*.js"

# OpenAI
grep -rn "chat\.completions\.create\|openai\.chat\|completion\.create" \
  --include="*.py" --include="*.ts" --include="*.js"

# Google / Gemini
grep -rn "generate_content\|GenerativeModel\|genai\." \
  --include="*.py" --include="*.ts" --include="*.js"

# Embeddings
grep -rn "embeddings\.create\|embed_content\|embedding" \
  --include="*.py" --include="*.ts" --include="*.js"

For each call, record:

• Feature / file / function triggering it
• Model alias or literal model string
• Task type: generation, classification, summarization, embedding, tool use, reasoning
• Rough prompt size: small (<1K tokens), medium (1–10K), large (>10K)
• Frequency: per-request, per-user-action, background job, one-time

Phase 3 — Pricing reference (2026)

Anthropic Claude

Model	Input $/1M	Output $/1M	Cache write $/1M	Cache read $/1M	Context
claude-fable-5	$10.00	$50.00	$12.50	$1.00	1M
claude-opus-4-8	$5.00	$25.00	$6.25	$0.50	1M
claude-sonnet-4-6	$3.00	$15.00	$3.75	$0.30	1M
claude-haiku-4-5	$1.00	$5.00	$1.25	$0.10	200K

Prompt caching saves ~90% on cache read vs uncached input. Requires 1024+ tokens in the cacheable prefix.

OpenAI

Model	Input $/1M	Output $/1M	Cached input $/1M	Notes
gpt-4o	$2.50	$10.00	$1.25	Vision capable
gpt-4o-mini	$0.15	$0.60	$0.075	High-volume workhorse
gpt-4.1	$2.00	$8.00	$0.50	Longer context, strong coding
gpt-4.1-mini	$0.40	$1.60	$0.10	Good at instruction following
gpt-4.1-nano	$0.10	$0.40	$0.025	Cheapest, short tasks only
o4-mini	$1.10	$4.40	$0.55	Reasoning, budget version
o3	$10.00	$40.00	$2.50	Heavy reasoning tasks

Batch API: 50% discount on all models. Min latency ~24h. Use for offline/async workloads.

Google Gemini

Model	Input $/1M (≤200K)	Input $/1M (>200K)	Output $/1M	Notes
gemini-2.5-pro	$1.25	$2.50	$10.00	Best for reasoning
gemini-2.5-flash	$0.15	$0.15	$0.60	Fast, multimodal
gemini-2.0-flash	$0.10	$0.10	$0.40	Cheapest Gemini

Context caching: free to cache, billed per-token-per-hour (~$1/1M tokens/hour for 2.5 Pro).

Phase 4 — Build the routing table

Output this table:

AI MODEL ROUTING — [product name]
════════════════════════════════════════════════════════════════════════════
FEATURE                  TASK TYPE        MODEL                   $/1K CALLS
────────────────────────────────────────────────────────────────────────────
[feature]                [task]           [actual model string]   $[estimate]
[feature]                [task]           [actual model string]   $[estimate]
════════════════════════════════════════════════════════════════════════════

COST MATH (per call):
  Input tokens × (input rate / 1M) + Output tokens × (output rate / 1M)
  Example: 2K in, 500 out on claude-sonnet-4-6 = (2000 × $3/1M) + (500 × $15/1M)
                                                = $0.006 + $0.0075 = $0.0135/call

UNUSED CONFIG (defined, never called):
  [alias] → [model string]

DEPRECATED / RENAMED MODELS:
  ⚠️  [any model string that no longer exists or has been superseded]

Phase 5 — Check for caching opportunities

For each call with a large, stable prefix (system prompt, docs, schema, examples):

• Is the prefix ≥1024 tokens? → Prompt caching candidate
• Does the prefix change per-user or per-request? → Not cacheable
• Same system prompt on every call? → Cache it; saves ~90% on those tokens

Flag calls that have a large system prompt but aren't using cache_control.

Anthropic pattern:

{"role": "system", "content": [{"type": "text", "text": "...", "cache_control": {"type": "ephemeral"}}]}

OpenAI (automatic): input caching is automatic on gpt-4o / gpt-4.1 for repeated prefixes.

Phase 6 — Check for batch API candidates

Any call that is:

• Not user-facing (background job, scheduled task, dataset processing)
• Can tolerate hours of latency
• Volume > 100 calls/day

→ Flag as batch candidate. Batch = 50% cheaper (OpenAI) or up to 50% (Anthropic).

Phase 7 — Right-sizing flags

For each call, check if the model is over- or under-powered for the task:

Downgrade candidates (expensive model on a simple task):

• Classification or extraction → gpt-4.1-nano, haiku, or gemini-2.0-flash
• Short Q&A with a known answer domain → gpt-4o-mini or haiku
• Embedding generation → text-embedding-3-small (OpenAI) at $0.02/1M
• Summarization of short text → haiku, gpt-4.1-mini

Keep the powerful model when:

• Multi-step reasoning or tool use chains
• User-facing creative or long-form generation
• Code generation where correctness matters
• Tasks with high retry cost if wrong

Upgrade candidates (cheap model failing silently):

• Tool use with complex schemas → needs sonnet/gpt-4o minimum
• Long-context reasoning (>50K tokens) → needs a model with proven long-context recall
• Structured extraction from noisy input → cheap models hallucinate field names

Phase 8 — Final flags

OPTIMIZATION FLAGS:
  CACHE   [feature] — large static system prompt, not cached. Estimated savings: $X/day
  BATCH   [feature] — background job, 50% savings available
  OVERFIT [feature] — using [expensive model] for [simple task], [cheap model] would work
  SCATTER [feature] — hardcoded model string outside central config, won't update on model changes
  DEPR    [model string] — deprecated, may stop working; migrate to [replacement]
  INCON   [task type] — same task using different models in different files: [file A] uses X, [file B] uses Y