Bedrock Tier Advisor

Overview

Advises which Bedrock service tier and cross-region inference variant to use for a given workload. Combines static guidance (from bedrock-tier-guidance.md) with dynamic discovery (from the pricing cache) to produce a recommendation.

Part of the skill family: bedrock-pricing, agentcore-pricing, bedrock-capacity, agent-business-value, bedrock-tier-advisor (this).

Guidance File

The tier and variant guidance is stored in bedrock-tier-guidance.md in the same directory as this skill file.

Workflow

Step 1: Load Guidance

Read bedrock-tier-guidance.md for the decision framework.

Step 2: Discover Available Tiers for the Model

Use extract_bedrock_model_prices(results, all_tiers=True) from the bedrock-pricing script to see which tiers and variants actually exist for the user's model in their region.

import sys, os
sys.argv = ['bedrock_pricing.py']
_p = os.path.join(os.getcwd(), "skills/bedrock-pricing/scripts/bedrock_pricing.py")
if not os.path.exists(_p):
    _p = os.path.expanduser("~/.quickwork/skills/bedrock-pricing/scripts/bedrock_pricing.py")
exec(open(_p).read())

home = os.path.expanduser("~/bedrock_cache")
results = query_model_pricing(home, region_filter="us-east-1", model_filter="Claude Sonnet 4.6")
all_prices = extract_bedrock_model_prices(results, all_tiers=True)
# Returns: {"Standard Global": {...}, "Standard Regional": {...}, "Batch Global": {...}, ...}

Step 3: Apply Decision Framework

Match the user's workload characteristics against the guidance:

Step 4: Present Recommendation

Show the recommended tier + variant with reasoning, and list alternatives with trade-offs.

Example output:

Recommended: Standard Global ($3.00/$15.00 per M tokens)
- Prompt caching available (saves ~60% on input costs)
- Global variant is ~10% cheaper than Regional
- Good for production workloads with moderate traffic

Alternatives:
- Batch Global ($1.50/$7.50) — 50% cheaper, but async (24h window)
- Standard Regional ($3.30/$16.50) — use if data must stay in us-east-1

Keeping Guidance Current

The bedrock-tier-guidance.md file should be refreshed when AWS updates their tier structure or pricing guidance. This typically happens 1-2 times per year.

After running --refresh on the pricing cache, update the guidance by opening Quick Desktop and asking:

"Read the current guidance file at ~/.quickwork/skills/bedrock-tier-advisor/bedrock-tier-guidance.md. Query AWS documentation using the aws-documentation-mcp-server to extract the latest tier and variant guidance from these pages:

• https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html
• https://docs.aws.amazon.com/bedrock/latest/userguide/capacity-limits-cost-optimization.html
• https://docs.aws.amazon.com/bedrock/latest/userguide/cross-region-inference.html

Update bedrock-tier-guidance.md with any changes. Preserve the file structure and source citations."

Configuration

Model preferences and tier defaults are managed by the YAML configuration system in bedrock_pricing.py. Override any default via ~/.bedrock_skills/config.yaml (user-level) or ./.bedrock_skills.yaml (project-level).

Run python3 bedrock_pricing.py --init-config to generate a commented template showing all available settings with their current defaults.

Precedence: function parameter > environment variable > project config > user config > hardcoded default

Config values are defaults only. If the user specifies a value in their prompt, always use the user's value. Config defaults apply only to parameters the user has not mentioned.

Relevant config sections:

• model_preferences — default model selections by agent role (router, general, rag, research) and version
• defaults.tier_preference — preferred service tier when user doesn't specify

Lessons Learned

Do

• Always discover available tiers from cache before recommending — not all models support all tiers
• Default to Standard Global when the user hasn't expressed a preference
• Prefer tiers with prompt caching support — it's the biggest cost lever
• Show the price difference between recommended and alternative tiers
• Cite the guidance source URLs so users can verify

Don't

• Don't recommend Flex for production workloads — no SLA, high throttling risk
• Don't recommend Priority unless the user has a clear latency/reliability need — it's 3.5× the cost
• Don't assume all models have Global variants — Llama, Qwen, and GPT-OSS are Regional only
• Don't hardcode tier names in recommendations — always discover from cache