Bedrock Pricing

Bedrock Model Pricing

Critical Rules

• ALWAYS ask for region first. No default region — never assume one.
• ALWAYS ask for model family. If user hasn't specified (e.g., "Sonnet", "Opus", "Haiku"), ask.
• ALWAYS use list_models() before get_model_prices(). Present versions to user, let them pick. Never guess "latest."
• NEVER use training data for prices. All prices must come from the local pricing cache files at runtime via query_model_pricing().
• NEVER implement cost formulas manually. Always use estimate_cost() or calculate_agent_session_compounded_cost().
• NEVER implement token formulas manually. Always use calculate_compounded_tokens_for_agent(), calculate_rag_subagent_tokens(), or calculate_research_subagent_tokens() for token calculations — if the user is only asking about token usage, not pricing.
• NEVER generate your own calculations. Present function output verbatim — it handles all token math internally.
• If user asks for detailed explanation, read the report file at result["file_path"]. Present the information as-is, then explain as needed. Do NOT recompute or manually derive calculations.
• ALWAYS run capacity fit check after calculating pricing using get_tier_limits_for_model() and check_capacity_fit(). This is not optional — every pricing result must include a capacity verdict for each model in the workload.
• All values in code examples are illustrative only. Always use user-specified values when provided. Prices must always come from the pricing cache, never from examples in this document.
• If user describes an agentic workload ("agent", "multi-agent", "sub-agent"), load agentcore-pricing and present combined costs.

Quick Reference

# §Load Script
import sys, os
sys.argv = ['bedrock_pricing.py']
script = ("tco_bva_capacity_skills/skills/bedrock-pricing/scripts/bedrock_pricing.py"
          if os.environ.get("USE_IN_KIRO") or os.environ.get("USE_IN_CLAUDE_CODE")
          else os.path.expanduser("~/.quickwork/skills/bedrock-pricing/scripts/bedrock_pricing.py"))
exec(open(script).read())
home = os.path.expanduser("~/bedrock_cache")

# 1. Check freshness
cache_status = check_pricing_data_status()  # handle "missing" → tell user to --refresh

# 2. Resolve model
models = list_models(home, "us-west-2", "Sonnet")  # → present to user, let them pick

# 3. Create session directory (all reports go here)
session_dir = create_report_session(model_name="Claude Sonnet 4.6", volume=10000)

# 4a. FAST PATH (single agent ONLY — raises error if subagents passed):
result = estimate_cost(home, "us-west-2", "Claude Sonnet 4.6", 10000, output_dir=session_dir)

# 4b. FULL PATH (multi-agent or custom params — REQUIRED for sub-agents):
prices = get_model_prices(home, "us-west-2", "Claude Sonnet 4.6")
result = calculate_agent_session_compounded_cost(
    main_agent_config={**prices, "agent_sessions_per_month": 10000, ...},
    subagents=[...],
    output_dir=session_dir,
)

# 5. Capacity (ALWAYS — not optional)
tier_limits = get_tier_limits_for_model(home, "Claude Sonnet 4.6", "us-west-2")
cap = check_capacity_fit(
    capacity_profile=result["capacity_profile"]["main_agent"],
    questions_per_month=50000, tier_limits=tier_limits, output_dir=session_dir,
)

Workflow

1. Check Pricing Data Freshness (once per session)

cache_status = check_pricing_data_status()

• "ok" — proceed.
• "stale" — warn user, suggest refresh, proceed with available data.
• "missing" — stop. Tell user to run the refresh command from cache_status["refresh_command"].

2. Resolve Model

models = list_models(home, "us-west-2", "Sonnet")
# → ["Claude 3 Sonnet", "Claude 3.5 Sonnet", ..., "Claude Sonnet 4.6"]

• Present the list to user, ask which version.
• Empty list → no models for that family/region.
• User provides exact model name upfront → skip to Step 3.

3. Present Assumptions

Before calculating, confirm parameters with the user:

• Sessions per month
• Questions per session (default: 5)
• Tools passed / invoked per question
• System prompt tokens
• Input/output tokens per question
• For multi-agent: which sub-agents, their models, invocation pattern

Only proceed after user confirms or adjusts values.

4. Calculate Cost

Always create a session directory first — all reports (pricing, capacity, agentcore, BVA) go here:

session_dir = create_report_session(model_name="Claude Sonnet 4.6", volume=10000)

Fast path — single agent ONLY. For multi-agent, you MUST use the full path below:

result = estimate_cost(home, "us-west-2", "Claude Sonnet 4.6", 10000, output_dir=session_dir)
# Pass any override as keyword: questions_per_agent_session=3, system_prompt_tokens=4000

Full path — multi-agent or custom sub-agents:

main_prices = get_model_prices(home, "us-west-2", "Claude Sonnet 4.6")
rag_prices = get_model_prices(home, "us-west-2", "Nova 2.0 Lite")

result = calculate_agent_session_compounded_cost(
    main_agent_config={
        **main_prices,
        "agent_sessions_per_month": 10000,
        "questions_per_agent_session": 5,
        "tools_passed_to_agent": 10,
        "tools_invoked": 5,
    },
    subagents=[
        {
            "type": "rag",
            "token_params": {"rag_n_chunks": 10, "rag_chunk_size": 300, "output_tokens": 300},
            "model_prices": rag_prices,
            "questions_invoked": 3,
        },
        {
            "type": "research",
            "token_params": {"n_research_iterations": 4, "output_tokens": 1000},
            "model_prices": research_prices,
            "questions_invoked": 0,  # 0 = pre-session (output cached in system prompt)
        },
    ],
    output_dir=session_dir,
)

Key rules for sub-agents:

• questions_invoked=0 → pre-session (output added to system_prompt, cached)
• questions_invoked=N → invoked as tool in the first N questions
• tools_invoked must be >= number of sub-agents with questions_invoked > 0
• Sub-agent models that support caching (non-None cache_read_price) get caching automatically

5. Detect Prompt Caching Support

A model supports caching if cache_read_price and cache_write_price are non-None in get_model_prices() output. Some models (Nova) have cache_write_price=0.0 — caching is free for writes.

6. Present Results

The function writes a full report and returns a compact summary:

{
    "file_path": "~/bedrock_reports/...",
    "monthly_total": 3247.48,
    "annual_total": 38969.70,
    "session_total": 0.324748,
    "savings_pct": 60.9,
    "recommended_ttl": "5min",
    "capacity_profile": {...},  # pass to check_capacity_fit()
}

• Present as markdown table. Include file_path for the user.
• If _file_write_failed: True → full result is inline; render token_table verbatim.

7. Capacity Fit Check (MANDATORY — run for every pricing result)

Run check_capacity_fit() for each model in the workload. Do not skip this step.

tier_limits = get_tier_limits_for_model(home, "Claude Sonnet 4.6", "us-west-2")
cap_result = check_capacity_fit(
    capacity_profile=result["capacity_profile"]["main_agent"],
    questions_per_month=50000,
    tier_limits=tier_limits,
    output_dir=session_dir,
)
# Returns compact: fits, utilization %, recommendations, report_file

• If tier_limits is None → warn user quota data is missing, suggest --refresh
• If workload doesn't fit → present recommendations from the result
• For multi-agent: check each distinct model separately

8. Completeness Check (MANDATORY)

Token-Only Calculations

If user asks about token usage (not pricing), no cache files needed:

• calculate_compounded_tokens_for_agent() — main agent session tokens
• calculate_rag_subagent_tokens() — RAG sub-agent tokens
• calculate_research_subagent_tokens() — research sub-agent tokens

Related Skills