Does Bedrock Pricing Match OpenAI Direct?

The mistake

There are two, and this page used to commit one of them. The old mistake was assuming Bedrock prices OpenAI models differently from OpenAI direct — this page said so, citing OpenAI's pricing page, which then described Bedrock pricing as something that "may differ." The current mistake is the opposite overcorrection: assuming that because the prices now match, a Bedrock deployment behaves like OpenAI direct in every way that affects cost.

What OpenAI says now

OpenAI's Bedrock guide states that Bedrock pricing in commercial regions matches OpenAI direct pricing for equivalent services, and that using a region-specific Bedrock service is priced at the same rate as Regional processing on the OpenAI API — which carries a regional uplift. Billing still runs through AWS, and AWS's own pricing page and model cards are where Bedrock's service tiers and billing options are defined.

Why it still matters

Matching per-token prices doesn't make the deployments equivalent. On Bedrock, GPT-5.4 and GPT-5.5 get a smaller context window than on OpenAI direct, which moves where a long agentic session runs out of room. Several features that change cost are unavailable there: asynchronous tool calling, reasoning updates via configuration_update (the GPT-6 way to change effort without breaking the cache), mid-turn steering, hosted file search and remote MCP servers. A workload that leans on any of those will be rebuilt differently on Bedrock, and the rebuild has its own token cost.

The fix

Price Bedrock from OpenAI's direct tables for commercial-region, global use; add the regional uplift if you pin a region; and confirm service-tier availability — Batch, Flex, Fast — in AWS's own documentation rather than assuming it carries over. Check the platform table for the window and feature differences before committing an architecture.

Reconciling a Bedrock bill

Because commercial-region Bedrock pricing now matches OpenAI direct, a bill can be checked: take the token counts Bedrock reports for a period, price them at OpenAI's direct rates for the same model and tier, and compare with the AWS invoice. A gap then points at something specific — a regional endpoint billed with the uplift, a tier you didn't expect, or a smaller context window forcing more, shorter sessions — rather than at a vague platform premium.