Amazon Bedrock vs Vertex AI: API cost and speed

On GPT-OSS 120B, Vertex AI lowers the request example to $0.00126 from Bedrock’s $0.002163, while Bedrock has the higher Endpoint Accuracy Index: 101% versus 72%.123

Updated · 3 sources · Methodology

DimensionWinnerMarginEvidence
Estimated request costVertex AI$0.000903Amazon Bedrock $0.002163 / Vertex AI $0.0012623
Output speed (tokens/s)Vertex AI79.5 tokens/sAmazon Bedrock 101.2 tokens/s / Vertex AI 180.7 tokens/s
First-chunk median (s)Vertex AI0.18 sAmazon Bedrock 1.02 s / Vertex AI 0.84 s
Endpoint Accuracy Index v1.0Amazon Bedrock29 ppAmazon Bedrock 101% / Vertex AI 72%1

Performance

ProviderOutput speed (tokens/s)First-chunk median (s)Endpoint Accuracy Index v1.0
Amazon Bedrock1101.21.02101%
Vertex AI1180.70.8472%

Prices and request cost

ProviderInput / 1M tokensOutput / 1M tokens
Amazon Bedrock2$0.1545$0.618
Vertex AI3$0.09$0.36
Product documentationAmazon Bedrock documentationVertex AI documentation
API documentationNovita AI · Chat API
Cost example: 10,000 input + 1,000 billed output tokens
Amazon Bedrock$0.002163
Vertex AI$0.00126

Capabilities and serving conditions

DimensionAmazon BedrockVertex AI
Compared modelGPT-OSS 120B · GPT-OSS 120B
Billing modelUsage-based input and billed output token pricing; contract capacity is separate.Usage-based input and billed output token pricing; contract capacity is separate.
Serving and capacityThe price row uses standard Sydney on-demand GPT-OSS 120B. IAM, regional availability and service quotas apply; Batch and other regions have different prices.2The GPT-OSS 120B managed API lists $0.09/$0.36 per 1M input/output tokens. Google also lists Batch at $0.045/$0.18; managed APIs and self-deployed Model Garden instances have separate billing.3

When each option fits

Amazon Bedrock

  • Consider Amazon Bedrock for 101.2 tokens/s and a $0.002163 request example; its first-chunk wait is 1.02 s.21

Vertex AI

  • Consider Vertex AI for 180.7 tokens/s and a $0.00126 request example; its first-chunk wait is 0.84 s.31

Methodology and sources

For GPT-OSS 120B, AA performance uses 10,000 input tokens and high reasoning: output speed and first reasoning-chunk latency are medians. Endpoint Accuracy Index comes from a separate AA evaluation relative to a reference endpoint; it is not a pass rate. The uncached cost example bills 10,000 input and 1,000 output tokens at the named sellers’ standard tariffs. Rolling third-party performance values are the batch snapshot identified by the page’s Updated line, not live guarantees.

  1. Artificial Analysis · GPT-OSS 120B provider table
  2. Amazon Bedrock · API pricing
  3. Vertex AI · API pricing

Build with Novita Model APIs

Create an API key and connect your application.

Start BuildingExplore Models