OpenAI lands on AWS Bedrock: GPT-5.5 and Codex accessible in the Amazon ecosystem

OpenAI lands on AWS Bedrock: GPT-5.5 and Codex accessible in the Amazon ecosystem

🔎 The biggest shift in AI cloud since the launch of ChatGPT

OpenAI on AWS. The phrase alone would have raised eyebrows six months ago. Yet, since June 1, 2026, it is the reality: GPT-5.5, GPT-5.4, and Codex are generally available on Amazon Bedrock.

This is the culmination of a historic $50 billion deal between the two giants. A deal that redraws the map of the AI cloud and poses an awkward question for Microsoft: what happens to Azure when its exclusive partner opens the door to the competition?

The impact for enterprises is immediate. No more need to manage two separate cloud providers, no more dual security validation circuits, no more parallel contract negotiations. Everything remains in the AWS ecosystem, with the same safeguards, the same governance tools, the same team habits.

This article breaks down what this availability concretely changes, what it implies for Azure's positioning, and how technical teams must prepare for it.

The essentials

• Generally available since June 1, 2026: GPT-5.5, GPT-5.4 and Codex are accessible via Amazon Bedrock in Commercial and GovCloud regions, with no special prerequisites.
• Pricing identical to the first-party OpenAI API: usage on Bedrock counts toward existing AWS commitments, allowing for discounts to be negotiated via Enterprise Discount Program contracts.
• $50B agreement: $15 billion initial + $35 billion conditional, signed between AWS and OpenAI, which consolidates AWS as the leading multi-model platform.
• Native AWS governance: IAM, VPC via PrivateLink, CloudTrail, Guardrails for Bedrock — OpenAI models benefit from the same security perimeter as Claude or Amazon models.
• Direct alternative to Claude on Bedrock: companies that were choosing Claude by default on Bedrock for lack of an OpenAI alternative now have a choice, according to l'analyse de MindStudio.

Recommended tools

What is actually available — and what isn't

The three OpenAI models deployed on Bedrock cover the main enterprise use cases. But we need to be precise about what "general availability" means here.

GPT-5.5 is OpenAI's frontier model, leading the agentic leaderboard with a score of 98.2. It is designed for complex reasoning tasks, multi-step planning, and advanced agentic workflows.

GPT-5.4 and GPT-5.4 Pro (score 91.8) offer a different value-for-money ratio. Less expensive per token, they are suited for high-volume workloads such as classification, data extraction, and enterprise chatbots.

Codex (GPT-5.3 Codex, score 80) is OpenAI's coding agent, already positioned as a leader alongside Cursor and GitHub Copilot. Its deployment on Bedrock paves the way for development pipelines fully integrated into the AWS infrastructure.

What is not available, on the other hand: OpenAI's recent open weight models, which are deployed separately on SageMaker. And certain niche or preview models remain exclusive to the direct API. The OpenAI documentation for Bedrock details the exact functional differences between the two access methods.

Why this partnership breaks the Azure-OpenAI dynamic

Since 2019, Microsoft was OpenAI's exclusive cloud partner. Azure OpenAI Service was the only "enterprise" path to access GPT with SLA, compliance, and contract negotiation guarantees.

This monopoly is over.

The AboutAmazon announcement says it bluntly: customers access GPT-5.5, GPT-5.4 and Codex on Bedrock with the same AWS security and governance. The implicit message is clear — you no longer need Azure for enterprise OpenAI.

For Microsoft, the calculation is delicate. The $13 billion investment in OpenAI remains massive, but the $50 billion AWS-OpenAI deal surpasses it in scale. According to PYMNTS, this pivot follows a new agreement with Microsoft that loosens cloud exclusivity.

The business reality: OpenAI is becoming a multi-cloud model provider, much like Anthropic which is already present on AWS and GCP. The platform strategy wins out over the exclusive partnership.

For enterprises, this means downward pressure on prices. When two hyperscalers compete for the same models, bargaining power shifts to the customer side.

What this concretely changes for technical teams

Unified calls via the Bedrock API

The first advantage is architectural. Instead of maintaining an OpenAI SDK on one side and a Bedrock SDK on the other, teams use a single API. The same code that calls Claude Opus 4.7 or Gemini 3.1 Pro can call GPT-5.5 by changing a model identifier.

This is particularly relevant for multi-model architectures. A company can route coding tasks to Codex, complex reasoning to GPT-5.5, and low-cost tasks to GPT-5.4 — all while maintaining a unified observability and billing layer.

Governance and security without compromise

This is the point that will interest CISOs and compliance teams the most. OpenAI models on Bedrock inherit the entire AWS security perimeter:

• IAM for granular access control (who can call which model, with which permissions).
• PrivateLink so that traffic never leaves the company's VPC.
• CloudTrail for complete auditing of every call (who, when, which model, how many tokens).
• Guardrails for Bedrock to filter content, restrict sensitive topics, and apply security policies at the model level.

For regulated sectors (banking, healthcare, defense), availability in GovCloud is a strong signal. It means that US government agencies can use GPT-5.5 without data passing through the standard commercial infrastructure.

Cost optimization through AWS commitments

Pricing is identical to OpenAI first-party rates, according to the AWS What's New announcement. But the difference lies in the billing structure.

Usage of GPT-5.5 and GPT-5.4 on Bedrock counts toward existing AWS commitments. If a company has negotiated an Enterprise Discount Program or Savings Plans, OpenAI calls benefit from the same discounts. This is a concrete advantage over the direct API, where negotiation is done separately.

Open weight models on SageMaker, not Bedrock

A distinction must be made between two separate deployments. Frontier models (GPT-5.5, GPT-5.4, Codex) are on Bedrock. OpenAI's open weight models are deployed on SageMaker, as specified by AboutAmazon.

These open weight models show remarkable cost performance: 10x more cost-efficient than a comparable Gemini model, 18x more than DeepSeek-R1, and 7x more than o4. For very high-volume workloads where budget is the limiting factor, this is a serious option to consider via SageMaker.

Codex on-premise vs Codex on Bedrock: two distinct strategies

There is an important nuance to understand. The arrival of Codex on Bedrock does not replace OpenAI's on-premise strategy. These are two complementary offerings for two different segments.

Codex on Bedrock is aimed at companies that want managed access without managing the underlying infrastructure. They pay per use, benefit from AWS's automatic scalability, and delegate operations.

Codex on-premise, pushed in partnership with Dell, is aimed at organizations that absolutely cannot send their source code to a third-party cloud — defense, intelligence, investment banks.

The two coexist. And the fact that OpenAI offers both shows strategic maturity: they are no longer forcing anyone into a single consumption model.

For teams evaluating the meilleurs LLM pour coder, a dimension now needs to be added: is the model available in my cloud of choice, and with what level of control?

The competitive landscape on Bedrock: GPT-5.5 vs Claude Opus 4.7

OpenAI's arrival on Bedrock transforms the platform's internal dynamics. Until now, Claude was the de facto premium model on Bedrock. This is no longer the case.

The comparison is interesting. GPT-5.5 dominates the agentic leaderboard with 98.2, compared to 94.3 for Claude Opus 4.7 (Adaptive). But raw scores don't tell the whole story.

Claude remains superior on certain fronts: handling very long contexts, adherence to complex instructions, and caution in sensitive responses. GPT-5.5 excels in mathematical reasoning, multi-step planning, and the richness of its tool ecosystem (notably the three GPT-Realtime-2 real-time voice models which have no equivalent at Anthropic).

In practice, companies on Bedrock will likely adopt a multi-model strategy: Claude for writing and document analysis tasks, GPT-5.5 for reasoning and agents, GPT-5.4 for volume.

The real winner in this story is AWS. More models means more potential lock-in at the orchestration layer, not at the model level.

Implications for enterprise multi-cloud strategy

The myth of model-level vendor lock-in collapses

For years, cloud architecture consultants have recommended abstracting the model layer to avoid lock-in. OpenAI's arrival on Bedrock confirms that this abstraction is not only possible, but is becoming the norm.

A company can now run GPT-5.5 on AWS, GPT-5.5 on Azure, and GPT-5.5 via the direct API — with the same model, different prices, and different levels of control. Lock-in is no longer at the model level, but at the orchestration platform level (Bedrock, Azure AI, Vertex AI).

AWS strengthens its position as a "model supermarket"

Bedrock was already the richest platform in terms of model diversity. With the addition of OpenAI, AWS widens the gap even further. A company can now compare, within a single console, GPT-5.5, Claude Opus 4.7, Gemini 3 Pro Deep Think, and open weight models — with the same API, the same billing, and the same governance.

This is a massive structural advantage. Azure primarily offers OpenAI models (along with a few open-source models). GCP primarily offers Gemini. Bedrock offers everything.

Beware of remaining functional differences

Not everything is identical between the direct OpenAI API and Bedrock. The OpenAI documentation lists concrete differences: certain advanced features (such as specific fine-tunings or certain structured output modes) may arrive with a delay on Bedrock.

Teams migrating existing workflows from the direct API to Bedrock must validate each integration point by point. It is not a simple endpoint URL change.

❌ Common mistakes

Mistake 1: Thinking Bedrock is just a "proxy" to the OpenAI API

Bedrock integrates OpenAI models into its own control plane. This means that AWS guardrails (Guardrails, IAM, CloudTrail) apply before the request reaches the model. The architecture is fundamentally different from a simple reverse proxy. Ignoring this distinction leads to incorrect assumptions about latency, logging, and error handling.

Mistake 2: Negotiating OpenAI prices separately from your AWS contract

This is the most expensive mistake. Since Bedrock usage counts toward AWS commitments, the optimal negotiation is to include OpenAI tokens in the global discussion with your AWS account. Separating the two negotiations means leaving money on the table.

Mistake 3: Migrating all OpenAI calls to Bedrock all at once

Even in general availability, a production migration deserves a gradual plan. Start with non-critical workloads, validate latency (Bedrock adds a layer of abstraction), verify that CloudTrail logs meet your audit needs, and then migrate gradually. A big-bang approach is a guaranteed incident.

Mistake 4: Ignoring open weight models on SageMaker

Focusing solely on GPT-5.5 and GPT-5.4 on Bedrock means ignoring the most cost-aggressive option. OpenAI open weight models on SageMaker, with their 10x to 18x cost advantage over competitors, can absorb 70-80% of your volume for a fraction of the budget. The optimal strategy combines both: Bedrock for premium, SageMaker for volume.

❓ Frequently Asked Questions

Is GPT-5.5 on Bedrock more expensive than on the direct OpenAI API?

No. The pricing is strictly identical according to AWS What's New (June 2026). But on Bedrock, usage can be covered by your existing AWS commitments, which can make it cheaper in practice.

Is Codex on Bedrock the same agent as standalone Codex?

Yes, it is the same GPT-5.3 Codex model (score 80 on the agentic leaderboard). The Bedrock integration adds the AWS governance layer but does not modify the capabilities of the model itself.

Are the data sent to GPT-5.5 on Bedrock used for training?

No. As with the OpenAI enterprise API, Bedrock customers' data is not used to train OpenAI models. This is specified in the service's terms of use.

Can GPT-5.5 on Bedrock be used from a European region?

General availability covers the Commercial regions (including several European regions such as eu-west-1, eu-central-1) and GovCloud. Check the exact availability for your region in the AWS console, as regional deployment may be gradual.

How to choose between GPT-5.5 on Bedrock and Claude Opus 4.7 on Bedrock?

It depends on your use case. GPT-5.5 (98.2) dominates in reasoning and agentic workflows. Claude Opus 4.7 (94.3) excels in long context and instruction fidelity. The most robust strategy is to benchmark both on your specific workload, not on generic scores.

✅ Conclusion

OpenAI's arrival on Amazon Bedrock is not just another model added to a catalog — it's the end of the era of exclusive AI cloud. With GPT-5.5, GPT-5.4 and Codex available behind IAM, PrivateLink and CloudTrail, companies gain game-changing flexibility, both technically and contractually. If your infrastructure runs on AWS, there is no longer any reason to keep one foot in Azure just for OpenAI.
```