GPT-5.5 on Bedrock vs OpenAI Direct API: Pricing & Tradeoffs (May 2026)

On May 1, 2026, AWS and OpenAI announced GPT-5.5 and GPT-5.4 on Amazon Bedrock in limited preview — ending Microsoft’s seven-year exclusive on OpenAI inference. Here’s the practical comparison: when Bedrock makes sense, when direct API wins, and the real cost numbers for both paths in May 2026.

Last verified: May 6, 2026

The decision in 30 seconds

Factor	OpenAI Direct API	GPT-5.5 on Bedrock
Raw price	$5/$30 per 1M in/out tokens	TBD at launch (assume parity to +30%)
Availability	GA, all regions OpenAI supports	Limited preview, request access
Data plane	OpenAI infrastructure	Stays in your AWS VPC
Auth/IAM	OpenAI API keys	AWS IAM, PrivateLink
Audit	OpenAI dashboard	CloudTrail logs, Bedrock guardrails
EDP / spend commits	No	Counts toward AWS EDP credits
Latest model versions	First (T+0)	Lags by days to weeks
Fine-tuning	Available	Not yet (May 2026)
Batch API	Available	Not yet (May 2026)
Best for	Startups, frontier-first, multi-cloud	AWS-heavy enterprises with compliance needs

Default answer for May 2026:

• AWS-native enterprise with compliance requirements → Bedrock.
• Anyone else → OpenAI direct API is cheaper, faster to access, and more feature-complete.

What was actually announced (May 1, 2026)

At the “What’s Next with AWS” 2026 event, AWS announced three OpenAI-related capabilities:

1. OpenAI models on Amazon Bedrock (limited preview). GPT-5.5 and GPT-5.4 available through standard Bedrock APIs.
2. Codex on Amazon Bedrock (limited preview). OpenAI’s coding agent runtime hosted in AWS, with the same harness as the OpenAI-hosted version.
3. Amazon Bedrock Managed Agents, powered by OpenAI (limited preview). A managed agent runtime built on the OpenAI agent harness with bundled inference, memory, skills, and AWS-native security.

All three require waitlist access. Broader availability is expected through Q3 2026.

Pricing comparison

OpenAI direct API (confirmed, May 2026)

• GPT-5.5: $5/1M input, $30/1M output.
• GPT-5.4: $2.50/1M input, $15/1M output.
• Cached input: ~50% discount.
• Batch API: ~50% discount on async workloads.

GPT-5.5 on Bedrock (May 2026)

AWS has not yet published GPT-5.5 Bedrock pricing publicly. Two reference data points to anchor expectations:

1. Claude on Bedrock is at parity with Anthropic’s direct API. AWS does not mark up Claude. This is the precedent OpenAI may follow.
2. Most third-party Bedrock models carry a 30-60% premium over their direct providers (per Mindstudio and TUN reporting). This is the historical baseline.

Practical assumption until AWS publishes: parity to +30%. Plan capacity assuming up to ~$6.50/$39 per 1M tokens, then re-tune when AWS confirms.

The hidden cost variables

The headline price isn’t where the decision usually gets made. Three other variables matter more for AWS-heavy customers:

• AWS EDP credits. Enterprise Discount Programs let you commit annual AWS spend in exchange for 5-15%+ discounts. Bedrock GPT-5.5 spend counts toward EDP commits and earns those discounts. Direct OpenAI API spend does not.
• Egress fees. AWS charges ~$0.09/GB for data transfer to the public internet. A heavy GPT-5.5 workload moving prompts and responses out to OpenAI’s API can cost $4,000-5,000/month in egress alone for high-volume customers. Keeping inference inside the AWS VPC eliminates this.
• Compliance overhead. SOC 2, HIPAA, FedRAMP, and PCI audits require documenting every data plane. Adding “OpenAI direct” as a separate auditable boundary costs real engineering and audit time. Keeping it inside the AWS VPC means it’s covered by your existing AWS audit.

When Bedrock wins decisively

Five scenarios where Bedrock is clearly the right pick:

1. Regulated enterprise with VPC-only requirements. No outbound traffic to OpenAI’s API allowed. Bedrock keeps the data plane inside your AWS network.
2. Heavy AWS commit ($1M+/year). EDP discounts and existing savings plans make Bedrock effectively cheaper despite headline markup.
3. Already deep on Bedrock (Claude, Llama, Nova). Adding GPT-5.5 to the same APIs, observability, and IAM model is operationally simpler than adding a second AI vendor.
4. Building with Managed Agents. If you want OpenAI’s harness with AWS-native operations, Bedrock Managed Agents is the only path.
5. Multi-region failover requirements. Bedrock spans many AWS regions with consistent IAM and billing; OpenAI direct is single-tenant infrastructure.

When direct OpenAI wins decisively

Five scenarios where direct API is clearly the right pick:

1. Startups under $1M annual AI spend. EDP discounts are too small to matter. Headline price wins.
2. Frontier-first product builders. New OpenAI models (4o → 5 → 5.4 → 5.5 → next) ship to direct API days to weeks before Bedrock. If you compete on capability, direct API wins.
3. Need fine-tuning or batch. Both unavailable on Bedrock for OpenAI models as of May 2026.
4. Multi-cloud or non-AWS shops. Routing through Bedrock from GCP or on-prem adds latency and complexity for no benefit.
5. Solo developers and small teams. Simpler billing, faster onboarding, no IAM ceremony.

Latency comparison

Reported numbers from early-access testing (May 2026):

• OpenAI direct API: P50 first-token latency ~600ms, P99 ~1.8s.
• GPT-5.5 on Bedrock: P50 first-token latency ~750ms, P99 ~2.1s — roughly 100-150ms higher due to AWS routing.

For interactive chat UX, the difference is negligible. For agent loops with 10-30 model calls per task, it adds 1.5-4.5 seconds total. Not a deal-breaker, but worth measuring on your workload.

Security and governance comparison

Capability	OpenAI Direct	Bedrock
PrivateLink	No	Yes
VPC-only inference	No	Yes
AWS IAM auth	No (API keys)	Yes
CloudTrail logging	No	Yes
Encryption at rest (KMS)	OpenAI-managed	Customer-managed
Bedrock Guardrails	No	Yes
Data residency controls	OpenAI regions	AWS regions
Compliance certifications	SOC 2, HIPAA available	SOC 2, HIPAA, FedRAMP, PCI included

For regulated industries (finance, healthcare, defense), Bedrock’s governance story is materially better.

What this means strategically

Three structural shifts triggered by this launch:

1. Microsoft Azure exclusivity is functionally over. Azure OpenAI Service will keep its current customers, but new enterprise wins are now contested. Expect Microsoft to respond with deeper Copilot integration and competitive pricing.
2. AWS now has the full frontier model menu. Claude (Anthropic), Nova (Amazon), Llama (Meta), and now GPT-5.5 (OpenAI). For multi-model strategies, Bedrock is the most complete platform in May 2026.
3. OpenAI’s go-to-market just doubled. Access to the AWS sales motion alongside Microsoft’s roughly doubles the enterprise reach. Expect OpenAI revenue acceleration through 2026-2027.

Bottom line

In May 2026, OpenAI’s direct API is the cheaper, faster-to-latest, and more feature-complete option for most use cases. GPT-5.5 on Bedrock wins for AWS-heavy regulated enterprises that need VPC-only data planes, EDP credit alignment, or are building with Bedrock Managed Agents. Both will exist long-term — they’re not substitutes, they’re channels for different buyers. Pick the channel that matches your existing infrastructure, compliance posture, and contract surface, not the headline price per token.

Sources: AWS Bedrock OpenAI page (May 2026), AWS What’s Next with AWS 2026 announcements, OpenAI pricing page (May 2026), Mindstudio and TUN coverage (May 2026), Stratechery Altman/Garman interview (April 2026).