OpenAI on AWS Bedrock — What GPT-5.5, Codex, and Managed Agents Change for Indian SMBs

OpenAI announced on June 1, 2026 that its frontier models — GPT-5.5, Codex, and managed AI agents — are now generally available on Amazon Bedrock. This isn't just another model launch. It's the first time OpenAI's primary models are available through a hyperscaler marketplace that Indian businesses already pay. No separate API key. No OpenAI billing dashboard. Just your existing AWS account, consolidated into one invoice with GST.

The practical implications for Indian SMBs and mid-market companies are more interesting than the headline.

What's actually on Bedrock

The GA launch on Bedrock includes three things:

1. GPT-5.5 — OpenAI's latest frontier model, available through the standard Bedrock InvokeModel and Converse APIs
2. Codex — OpenAI's coding agent model, accessible programmatically and through Bedrock's managed agent framework
3. Managed Agents — Pre-built agent scaffolding that connects Codex to your AWS resources (Lambda, S3, RDS) without writing MCP servers yourself

The pricing is per-token, same as OpenAI's direct API, but billed through AWS. For Indian companies, this means one invoice, rupee-converted pricing visible in AWS Cost Explorer, and — critically — a GST-compliant invoice you can actually claim input credit on.

Why this matters for Indian companies

Most Indian SMBs building on AWS have wrestled with the dual-billing problem: your compute is on AWS, your LLM calls go to OpenAI's API, and finance gets two separate invoices — one with GST, one without. OpenAI doesn't provide Indian GST invoices. AWS does.

For a mid-market company spending ₹2-5 lakh/month on inference, the GST input credit difference alone is ₹36,000-90,000/month. That's not theoretical. We've seen this exact problem at studios building AI features for Indian manufacturing and textile clients.

Beyond billing, there are three practical advantages:

IAM-based access control. You don't share API keys. Bedrock models are accessed through IAM roles — the same mechanism you use for S3, RDS, and Lambda. One less secret to leak, one less key rotation to automate.

VPC endpoint support. Inference calls stay inside your VPC if you configure a Bedrock VPC endpoint. For regulated industries (BFSI, healthcare, government contractors), data never transits the public internet. OpenAI's direct API can't offer this.

AWS CloudTrail logging. Every model invocation is logged. For SOC 2, ISO 27001, or just internal audit — you have a trail. OpenAI's API dashboard gives usage totals, not per-request audit logs.

What doesn't change

The model is the same. GPT-5.5 on Bedrock is not a "special AWS edition" with different capabilities or pricing. It's the same frontier model, same tokenizer, same context window. If your prompts work against api.openai.com, they work against Bedrock.

Latency is comparable. Early benchmarks from the HN discussion on June 1 show Bedrock latency within 5-10% of OpenAI's direct API for GPT-5.5 inference, mostly attributable to the IAM auth handshake on first request.

The Codex angle

Codex on Bedrock is the sleeper announcement here. Codex is OpenAI's coding agent — it writes, debugs, and iterates on code autonomously. Making it available through Bedrock means Indian dev teams can integrate coding agents into their existing AWS workflows without building custom infrastructure.

We see two immediate patterns for Indian SMBs:

1. CI/CD code review agents — A Lambda that invokes Codex on every PR to review for security issues specific to Indian compliance (data localization, Aadhaar masking, UPI transaction validation)
2. Internal tool builders — Non-technical operations staff describing a report they need in Hindi or Marathi, with Codex generating the SQL and Python against their RDS instance

Neither pattern requires new infrastructure. Both work inside the VPC.

The competitive landscape shift

This move reshapes the Indian cloud AI market in two ways:

AWS vs Azure vs GCP. Microsoft has been OpenAI's exclusive cloud partner since 2019. Bedrock availability breaks that exclusivity. Azure still has the deepest OpenAI integration (they host the infrastructure), but AWS now has feature parity for the models that matter. Indian companies that standardized on AWS for everything else no longer need an Azure subscription just for GPT-5.5 access.

Open-source models lose a pricing advantage. Self-hosting Llama 4 or DeepSeek-R1 on EC2 was attractive partly because it consolidated billing. If Bedrock gives you GPT-5.5 on the same invoice, the "simpler procurement" argument for self-hosting weakens. The quality gap between GPT-5.5 and the best open-source models remains real in mid-2026.

What to do tomorrow

If you're an Indian company already on AWS and paying OpenAI separately, the migration path is straightforward: swap your openai Python SDK calls for boto3 Bedrock converse() calls. The prompt format is slightly different but the model behavior is identical.

If you're evaluating LLM providers for a new project, the decision matrix just got simpler. Bedrock gives you GPT-5.5, Claude 4, and Llama 4 behind one IAM role and one invoice. That covers the frontier closed-source, safety-focused, and open-source categories in one billing relationship.

We've been saying for months that Indian SMB adoption of frontier AI would accelerate when procurement stopped being a separate workflow. June 1, 2026 might be that inflection point.