Your Own ChatGPT on AWS with Bedrock: The Per-Seat Math

Team AI chat tools are priced per seat, every month, whether a seat is used heavily or sits idle. If you have an AWS account, there’s a different model: run a private, multi-user AI chat workspace on Amazon Bedrock, in your own account, and pay only for the tokens you actually use. Here’s the math, and the one setup detail that makes it keyless.

Pricing verified June 2026 against openai.com, anthropic.com, and AWS us-east-1 rates.

The per-seat model

Option	Price	Billing	Data lives in
ChatGPT Business	$25/user/mo ($20 annual)	Per user	OpenAI
ChatGPT Enterprise	~$60/user/mo (150-seat min)	Per user	OpenAI
Claude Team	$25/seat/mo ($20 annual, 5-seat min)	Per seat	Anthropic

At 25 users, that’s $500–625/month, and it climbs with every hire whether or not those people use it.

The self-hosted alternative

A self-hosted chat workspace (the open-source LibreChat is excellent) on AWS costs:

Software: ~$43/month if you use a managed appliance (or free if you assemble it yourself)
Compute: a t3.medium is about $30/month. No GPU required, because inference runs on Amazon Bedrock, not on the box.
Inference: Amazon Bedrock usage, billed per token through your own AWS account

So the fixed cost is roughly $75/month flat for unlimited users, plus metered Bedrock usage. You never pay for an idle seat.

Keyless Bedrock: the setup that matters

The detail that trips people up: you don’t paste an API key. You attach an IAM instance role with bedrock:InvokeModel to the instance at launch, and the app picks up credentials from the AWS SDK’s default chain. No keys to store, rotate, or leak. Inference is billed straight to your AWS account.

Two one-time, account-side steps to know about:

Enable model access. In the Bedrock console, request access to the Anthropic Claude models, per region. This is a toggle, not a purchase.
Use inference-profile model IDs. Current Claude models on Bedrock require cross-region inference profiles, so the model ID looks like us.anthropic.claude-sonnet-4-6 rather than the bare name. Raw IDs reject on-demand throughput.

Cost scenarios

Small team, light usage. A per-seat plan can be simpler and even cheaper for a handful of heavy daily users. Start there if ops isn’t your thing.

Growing team, 25 users. Per-seat runs $500–625/month. The appliance is ~$75/month flat plus your Bedrock token usage, typically $4,000–6,000/year saved, depending on how much you actually generate.

Data-sensitive team. Conversations stay in your AWS account and region. Nothing is shared with a third-party AI vendor.

When a per-seat plan wins

Pick a hosted plan if you want the vendor’s exact frontier models and polished features out of the box, minimal administration, and you’d rather not manage an instance or Bedrock model access. The self-hosted route wins on cost at team scale, on uneven usage, and any time your chat data has to stay in your own account.