Bedrock Capacity

Bedrock Capacity Planning

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 proceeding. Present versions to user, let them pick. Never guess.
• ALWAYS get real limits from get_tier_limits_for_model(). Never assume or hardcode quota values.
• NEVER compute tokens manually. Always use calculate_compounded_tokens_for_agent() and build_capacity_profile_from_tokens() to derive the capacity profile.
• NEVER recommend quota increase before walking through the 3-Step Optimization Framework.
• All values in code examples are illustrative only. Always use user-specified values when provided. Prices and quotas must come from the cache, never from examples in this document.
• If user asks for detailed explanation, read the report file at result["report_file"]. Present the information as-is, then explain as needed. Do NOT recompute or manually derive calculations.

Capacity Facts (always apply)

• Claude 3.7 and ALL later Claude models: output_burndown_rate=5. All others use 1. This is the #1 capacity miscalculation error.
• Prompt cache reads don't count toward TPM — the biggest TPM saver.
• max_tokens is reserved upfront — if set to 4096 but output is ~100, you're wasting 3,996 TPM/request. The gap between max_tokens and actual output wastes quota on every request.

Key Concepts

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

# 2. Create session directory
session_dir = create_report_session(model_name="Claude Sonnet 4.6", volume=1000000)

# 3. Get token profile
#    PATH A — after pricing already ran in conversation:
capacity_profile = cost_result["capacity_profile"]["main_agent"]
#    PATH B — standalone (user asked capacity directly):
token_result = calculate_compounded_tokens_for_agent(
    questions_per_agent_session=5, input_tokens=100, output_tokens=150,
    system_prompt_tokens=2000, tools_passed_to_agent=10, tool_spec_tokens=100,
    tools_invoked=5, tool_call_tokens=100, tool_result_tokens=100,
    detail_level="full",  # REQUIRED for build_capacity_profile_from_tokens
)
capacity_profile = build_capacity_profile_from_tokens(
    token_result, sessions_per_month=1000000, model_name="Claude Sonnet 4.6"
)

# 4. Get limits
tier_limits = get_tier_limits_for_model(home, "Claude Sonnet 4.6", "us-west-2")

# 5. Check fit
result = check_capacity_fit(
    capacity_profile=capacity_profile,
    questions_per_month=5000000,
    output_burndown_rate=5,  # 5 for Claude 3.7+, 1 for others
    tier_limits=tier_limits,
    output_dir=session_dir,
)

Workflow

1. Resolve Model

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

• Present the list to user, ask which version.
• User provides exact model name upfront → skip to Step 2.

2. Present Assumptions

Before computing, confirm parameters with the user:

• Sessions per month
• Questions per session
• Tools passed / invoked per question
• System prompt size (tokens)
• Input/output tokens per question

Only proceed after user confirms or adjusts values.

3. Obtain Token Profile

Path A — pricing already ran in this conversation (preferred, avoids recomputation):

capacity_profile = cost_result["capacity_profile"]["main_agent"]

Path B — standalone capacity check (user asked about capacity directly):

token_result = calculate_compounded_tokens_for_agent(
    questions_per_agent_session=5, input_tokens=100, output_tokens=150,
    system_prompt_tokens=2000, tools_passed_to_agent=10, tool_spec_tokens=100,
    tools_invoked=5, tool_call_tokens=100, tool_result_tokens=100,
    detail_level="full",
)
capacity_profile = build_capacity_profile_from_tokens(
    token_result, sessions_per_month=1000000, model_name="Claude Sonnet 4.6"
)

For multi-agent: use aggregate_capacity_by_model(cost_result["capacity_profile"]) to group load by model.

4. Get Real Limits

tier_limits = get_tier_limits_for_model(home, "Claude Sonnet 4.6", "us-west-2")
# → {"rpm_high": 10000, "tpm_high": 6000000, "tpd_high": 8640000000}

If returns None → tell user to run --refresh. Do NOT proceed without limits.

5. Check Capacity Fit

session_dir = create_report_session(model_name="Claude Sonnet 4.6", volume=1000000)

result = check_capacity_fit(
    capacity_profile=capacity_profile,
    questions_per_month=5000000,
    output_burndown_rate=5,     # 5 for Claude 3.7+, 1 for others
    max_tokens_setting=4096,
    peak_to_avg_ratio=3.0,
    active_hours_per_day=12,
    active_days_per_month=22,
    tier_limits=tier_limits,
    output_dir=session_dir,
)

Returns compact summary:

{
    "fits": True,
    "rpm_utilization_pct": 5.7,
    "tpm_utilization_pct": 96.6,
    "tpd_utilization_pct": 9.8,
    "recommendations": ["..."],
    "report_file": "~/bedrock_reports/.../capacity-claude-sonnet-4-6.md",
}

Full calculations (RPM/TPM/TPD derivation, optimization checklist) are in the report file.

6. Present Results

Per-model comparison table:

• If fits: highlight metrics close to limit (>80%)
• If doesn't fit: walk through the 3-Step Framework below

7. Completeness Check (MANDATORY)

When Workload Doesn't Fit: 3-Step Framework

Walk through in order. Do NOT skip to quota increase.

Step 1: Optimize

Step 2: Traffic Profile

Ask user to validate:

• Peak RPM (P99 over 1-min window)?
• Active hours/days? Off-peak = lower effective RPM
• Peak-to-average ratio (default 3×)?

Step 3: Ramp Plan

• Volume needed now or in 3-6 months?
• Current monthly volume (baseline)?
• Target at 1 / 3 / 6 months?
• Increases granted after demonstrating usage.

Related Skills