As someone who works at big tech and spends countless hours in meetings hoping to get some small feature coordinated for deployment across two teams, I can't imagine the amount of meetings and 6-pagers that were involved in running these models on bedrock's hardware.
Remember that models on different inference platforms might not necessarily give exactly the same results, adding another axis of non-determinism to development. Things like quantization, custom model serving silicon, batching, or other inference optimizations might mean a model from the original provider performs differently from the hosted one :/
This paper isn't the exact same scenario, since it's an auditable open weight llama model, but shows the symptoms of this: https://arxiv.org/pdf/2410.20247
It's a shame people love to use hostile language (something I am also sometimes to blame), but I think redsocksfan45 misconception is good to address. The comment is however (rightfully) dead. I'll address it anyways.
Model performance consistency is important not because you want inference determinism (which you can actually get by setting tempetature to zero and applying a static seed). The `another axis of non-determinism` can be illustrated by the question "if I move from openrouter to bedrock, will gpt-5.5 perform the same?", to which the answer is no, at least not necessarily.
This is important because workflows that used to work on one platform might degrade or outright not work on another, even using the same model, which you have to account when deciding which provider to use.
Claude got a looooot more buy in with a lot of privacy-concerned orgs I work with because they could access it through their "trusted" intermediate Amazon. OpenAI has been banned and is not trusted. I'm not sure that I agree with these orgs' legal teams' assessments, but they definitely read the terms of service far closer than I did.
We will see if this changes the equation, but it feels like OpenAI is pretty far behind and playing catch up on all fronts. Though to be honest, "pretty far behind" is like 2-8 weeks in the AI world, so it may not matter a ton, it's mostly perception. And for me and my information bubble, perception of OpenAI is rock-bottom due to Sam Altman. From appearing unethical to appearing unhinged with demands from fabs and everything else, I'm not a fan.
It just not about AWS being some "trusted intermediary"... it's that the model runs inside the customer own AWS account under a different contract. AWS explicitly states inputs/outputs are not shared with model providers and are not used to train base models [1]
And for OpenAI, there is a May 2025 preservation order in NYT v. OpenAI. The court is forcing OpenAI to retain ChatGPT output logs indefinitely, including chats users have deleted that would normally be purged within 30 days [2].
That makes it a non starter for HIPAA/GDPR bound orgs.
You can sign ZDR agreements with any of the major LLM providers. Using AWS alone is also not sufficient. Even though AWS is running the model, you need to contact them for proper ZDR.[0]
While Anthopic has the best model and a focussed (no disturbance, lawsuits) leadership, they got a lot of enterprise access due to AWS. It is mutual no doubt, with both sides benefitting. The culture of feedback loop of AWS customers would have helped them in getting to enterprise-ready faster. Just my hypothesis.
A lot of companies already have data processing agreements and compliance sign-off for using AWS. Many are hesitant to send their data to AI startups with an incentive to train their models and a history of being.... loose with how they intake training data. Even when they do give assurances otherwise. AWS is more trusted in this aspect.
If this ends up similar to Claude on Bedrock, it's the same price.
No, you definitely don't have access to the weights. The raw weights are secret enough that when a model hits a certain level of capability, their guidelines are that they need enough procedures in place to try to keep nation states from being able to hack in and steal the weights.
I think cost is fairly similar. As for who wants it, enterprise stuff is a big thing, but it also seems very reliable. Have never had any rate restrictions or had it just be down, which it sounds like a lot of people have issues with for Anthropic's servers.
Great, I can now buy openAI through AWS with an interface that is totally incompatible with all my tools (unless AWS have finally given up and just made bedrock useful by adopting openAPI finally)
Availability through Bedrock has been a major driver in use of Anthropic in my org. And I am betting there is actual margin in it as well.
I wonder if this is directly linked to the split up with Microsoft. Just from my anecdata, OpenAI is getting completely ignored in serious enterprise deployments because what they offer on Azure sucks and there is no other corporate friendly way to get it. They probably saw themselves getting destroyed in enterprise and realised it was existential to be able to compete with Anthropic on AWS.
OpenAI just gave up Azure exclusivity, killed the AGI clause, and stopped paying Microsoft revenue share to get on AWS. Anthropic figured out 18 m ago that enterprises buy from their cloud, not from the best model. OpenAI is just catching up.
Just waiting for Gemma 4, DeepSeek 4 now. Then the only thing I'll be able to complain about is the completely different API to interface with (unless they FINALLY move to full OpenAI support).
Amazon Bedrock Mantle provides OpenAI compatible API endpoints for model inference, powered by Mantle, a distributed inference engine for large-scale machine learning model serving. These endpoints allow you to use familiar OpenAI SDKs and tools with Amazon Bedrock models, enabling you to migrate existing applications with minimal code changes—simply update your base URL and API key.
OpenAI is starting to scramble after missing revenue targets, and this move is likely 24 months too late. There’s almost nobody that will suddenly start using AI because of this decision and those already using it have already broadly made their model decisions to use things other than OpenAI.
It’s good to see more model diversity on Bedrock, but this smells more a rushed play by OpenAI to stop the market share bleeding vs a concerted strategy.
I'm not so sure about that. Like we're using Claude Code with Bedrock and have most things on AWS with SOC2 compliance and all that. Normally switching to Codex would have a ton of friction in terms of separate contract and billing, worrying about data retention, etc.
But with Codex and GPT5.5 having generally good reviews, when it goes on Bedrock, it would be very simple for us to just try it out. Obviously there's the general friction of knowledge and comfort of Claude Code vs Codex, but that seems possible to overcome depending on the differential of the models and also of the features. Like if Codex gets remote control for Bedrock, that'd be a big advantage over Claude Code on it.
38 comments
[ 4.1 ms ] story [ 59.5 ms ] threadThis HN post itself has 4 simultaneous announcement links; not a coincidence.
There are billions of investor money on the line if the wrong thing is said at the wrong time, it needs to be carefully crafted and staged.
This paper isn't the exact same scenario, since it's an auditable open weight llama model, but shows the symptoms of this: https://arxiv.org/pdf/2410.20247
Model performance consistency is important not because you want inference determinism (which you can actually get by setting tempetature to zero and applying a static seed). The `another axis of non-determinism` can be illustrated by the question "if I move from openrouter to bedrock, will gpt-5.5 perform the same?", to which the answer is no, at least not necessarily.
This is important because workflows that used to work on one platform might degrade or outright not work on another, even using the same model, which you have to account when deciding which provider to use.
We will see if this changes the equation, but it feels like OpenAI is pretty far behind and playing catch up on all fronts. Though to be honest, "pretty far behind" is like 2-8 weeks in the AI world, so it may not matter a ton, it's mostly perception. And for me and my information bubble, perception of OpenAI is rock-bottom due to Sam Altman. From appearing unethical to appearing unhinged with demands from fabs and everything else, I'm not a fan.
And for OpenAI, there is a May 2025 preservation order in NYT v. OpenAI. The court is forcing OpenAI to retain ChatGPT output logs indefinitely, including chats users have deleted that would normally be purged within 30 days [2]. That makes it a non starter for HIPAA/GDPR bound orgs.
[1] https://aws.amazon.com/bedrock/faqs/
[2] https://openai.com/index/response-to-nyt-data-demands/
[0]: https://platform.claude.com/docs/en/build-with-claude/claude...
You can just run "air gapped" inference?
Is this only of interest to enterprise customers already on AWS (who want "air gapped" behavior)? Is there any other use case for this?
This will be more expensive than calling OpenAI directly, right?
If this ends up similar to Claude on Bedrock, it's the same price.
I think cost is fairly similar. As for who wants it, enterprise stuff is a big thing, but it also seems very reliable. Have never had any rate restrictions or had it just be down, which it sounds like a lot of people have issues with for Anthropic's servers.
AI is kind of like the ultimate corporation drug. They are all on it. And can't get rid of it - ever again.
I wonder if this is directly linked to the split up with Microsoft. Just from my anecdata, OpenAI is getting completely ignored in serious enterprise deployments because what they offer on Azure sucks and there is no other corporate friendly way to get it. They probably saw themselves getting destroyed in enterprise and realised it was existential to be able to compete with Anthropic on AWS.
Microsoft Azure has been the worst interms of maintaining a highly available service and also managing predictable latency.
Their azure customer support is bad. Not ready for any real enterprise cloud offering. They behave like Comcast customer support.
It was absolutely idiotic to lock it down to Azure. It wasn't meant to be an iphone+at&t combo where the phone is an end all be all.
A cloud product depends on a lot of services and nobody would switch cloud providers for a candy.
https://docs.aws.amazon.com/bedrock/latest/userguide/bedrock...
It’s good to see more model diversity on Bedrock, but this smells more a rushed play by OpenAI to stop the market share bleeding vs a concerted strategy.
But with Codex and GPT5.5 having generally good reviews, when it goes on Bedrock, it would be very simple for us to just try it out. Obviously there's the general friction of knowledge and comfort of Claude Code vs Codex, but that seems possible to overcome depending on the differential of the models and also of the features. Like if Codex gets remote control for Bedrock, that'd be a big advantage over Claude Code on it.