Amazon Bedrock Splits GPT-6 Astra Endpoints: Choose Caching or Audit Logging
Amazon Bedrock launched openai.gpt-6-astra across two distinct endpoints: bedrock-runtime and bedrock-mantle. Selecting an endpoint forces an architectural trade-off between explicit prompt caching and centralized invocation audit logging.
Why it matters
You can cut prompt costs by 86% using explicit caching on Bedrock Mantle, but you must architect around the total absence of invocation logs and Guardrails.
TL;DR
- 01Bedrock Mantle offers prompt caching and server-side tools, but completely omits invocation audit logging.
- 02Bedrock Runtime provides Guardrails and CloudWatch logging, but lacks prompt caching and InvokeModel support.
- 03Exceeding 272,000 input tokens doubles the price for the entire request across both input and output tiers.
- 04Output tokens deplete your TPM quota at a 10x multiplier relative to inputs.
Key facts
- Context Window
- 1,050,000 input tokens / 128,000 output tokens
- Global Pricing (<= 272K)
- $10.00/1M input, $50.00/1M output
- Global Pricing (> 272K)
- $20.00/1M input, $75.00/1M output
- Prompt Cache Discount
- Write 1.25x uncached, Read 0.10x uncached (86% net saving)
- Quota Burndown
- 1 output token consumes 10 TPM units
The Endpoint Split: Runtime vs Mantle
Amazon Bedrock serves openai.gpt-6-astra across two different endpoints with mutually exclusive guarantees. If your architecture relies on compliance monitoring, bedrock-runtime provides S3 and CloudWatch invocation logging alongside Bedrock Guardrails, but only through the Converse API. It does not support explicit prompt caching.
Switching to bedrock-mantle exposes the OpenAI-compatible /openai/v1/responses and /openai/v1/chat/completions routes. It provides explicit prompt caching and server-side tools, but execution completely bypasses invocation logs. Furthermore, Mantle is restricted strictly to us-west-2 with no regional or global profiles.
The 272K Token Cliff and 10x Quota Burndown
Pricing jumps significantly past 272,000 input tokens. On global profiles, inputs below 272K cost $10.00/1M (and $50.00/1M output), but crossing 272,000 reprices the *entire* request to $20.00/1M input and $75.00/1M output. Because CountTokens is unsupported on bedrock-runtime, client-side token counting with safety buffers is mandatory before dispatching large contexts.
Rate limits also penalize generation: 1 output token burns 10 TPM units. An agent exhausting the full 128,000-token output limit burns 1,280,000 quota units, making output limits the primary capacity constraint.
Explicit Caching Mechanics
Prompt caching requires explicit configuration. You define prompt_cache_breakpoint: {"mode": "explicit"} with a 1,024-token minimum prefix and up to 4 breakpoints. Writes cost 1.25x uncached input, while reads cost 0.10x. Cached tokens do not burn against your TPM quota.
Try it in 2 minutes
{
"model": "openai.gpt-6-astra",
"prompt_cache_options": {
"mode": "explicit",
"ttl": "30m"
},
"prompt_cache_breakpoint": {
"mode": "explicit"
}
}json
✓ When to use
- Use bedrock-mantle when maintaining static prompt prefixes over 1,024 tokens across frequent agent tool calls.
- Use bedrock-runtime when integrating Bedrock Guardrails or requiring comprehensive CloudTrail and CloudWatch telemetry.
✕ When NOT to use
- Do not use bedrock-mantle if compliance policies mandate S3/CloudWatch invocation logging for every model transaction.
- Avoid unbounded agent prompts near 272,000 tokens without client-side token counting safety guards.
What to do today
- Implement client-side token counting before sending prompts approaching 272,000 tokens to avoid the 2x pricing cliff.
- Configure bedrock-mantle with explicit breakpoints if prompt caching discounts outweigh audit logging requirements.
- Account for the 10x TPM consumption rate on output tokens when calculating agent throughput limits.