# Bedrock Provisioned Throughput on Older Model Generation

> Provisioned Throughput purchased against a deprecated Bedrock foundation model generation pays yesterday's per-Model-Unit rate. ZopNight looks up real no-commitment PT rates for both the old model and its successor, pricing savings as cost times (1 minus newRate/oldRate); 1-month and 6-month committed-term PTs abstain because per-term rates are unresolvable.

Source: https://zop.dev/integrations/aws/recommendations/bedrock-provisioned-throughput-on-older-model-generation
Updated: 2026-08-19

---

## Matching a PT to a deprecated base model

- resource.Type: bedrock-pt
- metadata: modelArn and foundationModelArn, both extracted via extractBedrockModelIDs / modelIDFromArn (handles foundation-model / and inference-profile trailing :); matching both means a custom-model PT on a deprecated base FM is still flagged.
- SKU rates: [the per-MU-hr provisioned throughput rate](https://aws.amazon.com/bedrock/pricing/) for the deprecated model id AND its successor id (via TargetSKUCandidates → SetSKURates).

## Pricing the rate delta to the successor model

Concrete or abstain (Category rightsizing, since migrating a PT to a cheaper in-family generation is a rate-delta cost lever). This is live: the producer landed 2026-07-13, in this PR. savings = cost × (1 − newRate/oldRate), both rates looked up directly via model.LookupSKURate (not via the fraction-fallback helpers, which would fabricate). The rule abstains when either rate is missing or the successor is not cheaper. How the producer closed the gap: the AWS Bulk Pricing PT SKU exposes the model only as a CamelCase marketing display token (usagetype …-TitanImageGeneratorV2-ProvisionedThroughput-1month-ModelUnits) / human display string (titanModelUnit/modelUnit "Titan Image Generator V2"), never the canonical foundation-model id the lookup keys on (anthropic.claude-v2, meta.llama2-70b-chat-v1, …), and there is no lossless display→id transform. So the aggregator captures the no-commitment PT rate as a structured bedrock-pt-raw|\<displayToken> sentinel (extractBedrockKey, internal/pricing/aws/pricing.go) and resolves the display token → foundation-model id in FetchRates against the discoverer's live bedrock:ListFoundationModels harvest (exact case/space-normalized match or abstain, never fuzzy), keying the one real per-MU-hour rate under every id that shares the display token. The rule already declared TargetSKUCandidates (successor) and prices the real delta via SetSKURates, so it fires concretely with no rule-logic change. Because the resolved key carries no commitment-term dimension, the rule guards explicitly with bedrockResourceIsCommittedTerm(resource.Metadata) (mirroring the aggregator's bedrockPTIsNoCommit in internal/cost/calc_bedrock.go; separate modules, kept in lock-step, with drift guards on both sides) and returns nil before the rate lookup for any 1-month/6-month/Custom-term PT. Those remain a tracked, documented abstain (the display token alone can't select a per-term rate). Never a $0/advisory rec or fabricated fraction. Concrete-fire + abstain matrix pinned in bedrock_old_model_generation_test.go; the aggregator sentinel-capture + committed-term abstain pinned in TestExtractResourceKey_BedrockPT_NotCaptured. See the aggregator docs/COST.md entry for the AWS Bedrock Provisioned Throughput per-MU-hour rate, now live.

## Recreating the PT on the newer model

1. Identify the latest generation in the same model family (e.g., Claude 3.5/4, Titan v2, Llama 3, Cohere Command R)
2. Benchmark output quality on representative prompts
3. Recreate the PT against the newer model identifier
4. Migrate traffic and delete the old PT

## Why metadata alone decides this rule

None (metadata-driven)