🚀 AWS Launches OpenAI-Compatible API for Bedrock (and I Did Some Tests!)

#bedrock #aws #openai

🚀 AWS OpenAI-Compatible API for Bedrock OSS GPT: Real-World Dev Tests & Insights

AWS just dropped a bombshell by launching open-weight GPT models (gpt-oss-120b, gpt-oss-20b) on Amazon Bedrock as serverless, pay-as-you-go option to self hosting — this caught a lot of headlines 👀.

But as someone obsessed with Developer Experience (DX), I was even more stoked that AWS finally launched an official OpenAI-compatible endpoint:

https://bedrock-runtime.us-west-2.amazonaws.com/openai/v1

This puts AWS right alongside Gemini/VertexAI and Anthropic as companies providing first-party OpenAI SDK compatibility, and finally brings native OpenAI "plug-and-play" infra to AWS 🤩

Before we dive in, here’s my previous deep-dive blog for context on using Amazon Bedrock models with OpenAI compatibility (without fixed costs):

👉 https://dev.to/aws-builders/use-amazon-bedrock-models-via-an-openai-api-compatible-serverless-endpoint-now-without-fixed-cost-5hf5

😎 Wait—Does This Make the Bedrock Proxy Gateway Projects Stale?

Nope! While simple OpenAI API calls still runs perfectly with these new official compatibile Amazon Bedrock Runtimeendpoints, there are critical compatibility gaps — so custom proxies/gateways like my bedrock-access-gateway-function-url project and the official bedrock-access-gateway project are nowhere near obsolete yet.

By the Way, AWS Credits Does Cover the `gpt-oss` models!

It has been quite a pity that typical AWS Credits do not cover other models like Claude or Mistral. It's great that this time with the official partnership between AWS and OpenAI, the usage for gpt-oss models seems covered finally:

🧑‍💻 I Vibe-Coded the Tests with Kiro AI IDE — Here’s What Happened

You can check my actual repo and test scripts here:

📦 https://github.com/gabrielkoo/test-official-amazon-bedrock-openai-compatible-endpoint

AWS Official Docs: AWS Bedrock Docs: OpenAI Chat Completions API

(Note: The docs didn’t mention support is limited to latest OpenAI gpt-oss models only, and didn't mention directly support for tool_call or response_format features/parameters, so do read on for my real-world dev findings!)

Summary Table:

As a control test, I also ran the same testing script with my local Ollama gpt-oss:20b model:

Test	Bedrock OpenAI Endpoint	Local Ollama + `gpt-oss`	Notes
Basic Chat Completion + Reasoning	✅ Works: Step-wise reasoning included	✅ Works, chain-of-thought	Bedrock wraps reasoning in `<reasoning>` tags, both are great for math/logic explanations
`reasoning_effort` parameter	✅ Supported, adjusts explanation depth	✅ Supported, works	Fine-tune output detail—rare outside OpenAI OSS models
🔧 Tool/Function Calling (`tools`/`tool_choice`)	❌ Only intent as JSON, not native `tool_calls`	✅ True OpenAI tool_call schema	Just like the initial state of `o1-mini`: AWS "gets" the tools, but you can’t delegate natively
`response_format` (JSON schema)	❌ Not supported, API error	✅ Works, emits JSON	Bedrock doesn’t do response schema yet
Token param (`max_tokens` vs `max_completion_tokens`)	❌ Must use `max_completion_tokens`	✅ Accepts either	Bedrock is stricter, adjust scripts accordingly
Non-OSS Model (`Nova`, `Titan`, etc.)	✅ "Model not found", as expected	✅ Same	Endpoint only exposes OSS GPT models (`gpt-oss-*`) for now

🌟 Key DX Takeaways

API Parity, But Not Quite Full Compatibility: Any OpenAI SDK code (and most OpenAI agent frameworks) can now point at AWS’s Bedrock Runtime /openai/v1 endpoint with just env var tweaks — no code rewriting. This UX is 🔥 for engineers and innovation teams!
Transparent, Adjustable Reasoning: Get full chain-of-thought out-of-the-box, plus parametric reasoning level control. Perfect for math, science, coding, and explainability applications!
Tool Calling Support—Almost There: AWS Bedrock OSS GPT models recognize function call intent (return JSON), but don’t yet emit proper tool_calls blocks in the response. (While the same test case passed for my local Ollama test on the same gpt-oss model). This feels just like the early days of the o1-mini models, so: expect better in future, but for now, your OpenAI tool-calling apps will need adapters or proxies. 🙂
Strict OpenAI Param Parsing: If your code uses max_tokens, it’ll break here — swap to max_completion_tokens for as gpt-oss is a reasoning model.
Model Registry/Open Weight Only: The official endpoint supports only gpt-oss models so far (gpt-oss-120b, gpt-oss-20b). Other Amazon Bedrock serverless models like Nova will get you a fast fail as of now.
Still Need Dev Proxy Power: bedrock-access-gateway-function-url is not obsolete! It remains highly useful for compatibility workarounds or for non-gpt-oss models.

As a disclaimer, if we refer to the model documentation of gpt-oss on AWS, https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-openai.html, tool calls were not supported directly for gpt-oss on Amazon Bedrock hosted version as of now, so it’s totally possible when AWS release more model support for the new OpenAI endpoint, these could no longer become a concern!

💡 DX Tips for Devs

Add backward/forward compatibility adapters if you’re building serious agentic stacks.
Watch the docs (Invoke a model with the OpenAI Chat Completions API), but always test yourself — AWS hasn't listed all the current limitations yet clearly in the documentation!
Port your scripts with care:
- Reasoning, logic, coding, and OpenAI-style chat: 👍
- Tool use, output schemas: ⚠️ adapter or proxy needed today!
OSS all the things! Making the full open-weight LLM stack API-portable is a big win for multi-cloud, local/cloud hybrid, and plain old sovereignty nerds. It feels like a new era for "bring your own infra, but keep your SDKs."

Closing

Thanks to Kiro AI IDE for the vibe-code collab and to AWS's team on making the OpenAI compatible endpoint finally launched!

AWS’s new official OpenAI compatible endpoint for gpt-oss models is a major leap for cloud/dev portability, even if a few “classic” OpenAI features like native function calling aren’t fully there (yet). If you need adapters, or want total control, proxies like mine remain essential. But for standard chat/reasoning patterns, Bedrock is now OpenAI SDK ready — from your own existing code based on OpenAI SDKs.

Happy building! 🚀🤖

P.S. This article was written on 2025-08-06, 1 day after the new endpoint it was launched, so it's totally possible that by the time you read this blog article, AWS could have already rolled out more support for other common OpenAI SDK features!