GPT-6.1 Sol Hits Amazon Bedrock With a 10% Regional Premium

OpenAI's GPT-6.1 Sol went generally available on Amazon Bedrock on September 29, 2026, and the price depends on which endpoint you call. The global inference profile matches OpenAI's first-party price of $2 per million input tokens and $10 per million output tokens. Anything else costs 10% more. This GPT-6.1 Sol Bedrock breakdown covers the fine print to read before wiring it in.

GPT-6.1 Sol Bedrock access: three paths, two prices

The model ID is openai.gpt-6.1-sol, with roughly a 1M-token context window and text and image input. The AWS model card lists two endpoints.

| Path | Model ID | Reach | Input / output per 1M tokens (up to 272K input) | | --- | --- | --- | --- | | bedrock-mantle endpoint | openai.gpt-6.1-sol | us-east-1 only | $2.20 / $11.00 | | bedrock-runtime, US cross-Region | us.openai.gpt-6.1-sol | US geography | $2.20 / $11.00 | | bedrock-runtime, global cross-Region | global.openai.gpt-6.1-sol | Any commercial Region | $2.00 / $10.00 |

Mantle serves the Responses and Chat Completions APIs under /openai/v1, not /v1. Runtime adds Converse and Invoke, and does not allow direct single-Region invocation.

What the 10% premium costs on a real workload

Say you push 100M input and 10M output tokens a month. Global costs $300. US cross-Region or Mantle costs $330. Thirty dollars is noise, but at billions of tokens it is a line item.

The premium buys placement, not speed. Per AWS's cross-Region inference docs, geographic profiles keep processing inside the geography, while global profiles can route to any commercial Region. Pay it only if compliance demands US-only processing. Global profiles also need your SCPs to allow "aws:RequestedRegion": "unspecified".

The 272K long-context cliff

Above 272,000 input tokens, long-context rates apply to the whole request, not only the overflow. On global that is $4 input and $15 output per million; on the other paths, $4.40 and $16.50. OpenAI's model page states the same rule: 2x input, 1.5x output.

A 272,000-token prompt costs about $0.54 in input on global. A 272,001-token prompt costs about $1.09. Cap retrieval and chat history below the line.

What GPT-6.1 Sol on Bedrock doesn't have

• Service tiers. Standard only. Priority, Flex and Reserved are unsupported, so you can't buy lower latency or reserved capacity.
• Explicit prompt caching. The model card lists it as unsupported, AWS's launch announcement says it is supported, and the prompt caching page names only GPT-5.6 Sol, Terra and Luna among OpenAI models for explicit caching. The model card still prints cache prices, so the pricing table proves nothing. Treat caching as unavailable until your own test shows cached_tokens above zero.
• Token counting and server-side tools. The Count Tokens API, server-side system tools and intelligent prompt routing are unsupported. Budget from the usage object in responses instead.

The 10x output-token quota burndown

The model card says each output token consumes 10 tokens of quota. A request with 3,000 input and 1,000 output tokens uses 13,000 tokens of your per-minute quota but bills for 4,000.

Bedrock's token counting page adds a trap: at request start it deducts input tokens plus your maximum output setting, then refunds the unused part. A generous max_output_tokens can throttle you before the model writes a word. That page covers bedrock-runtime; the model card doesn't say how Mantle's quota treats it, so check Service Quotas for both.

Call GPT-6.1 Sol on Bedrock and read the usage

from openai import OpenAI

# Reads OPENAI_API_KEY (a Bedrock API key) from the environment.
client = OpenAI(
    base_url="https://bedrock-runtime.us-east-1.amazonaws.com/openai/v1",
)

response = client.responses.create(
    model="global.openai.gpt-6.1-sol",
    input="Summarize this incident report in five bullets: <paste report>",
    max_output_tokens=800,
)

print(response.output_text)
print(response.usage)

This uses the global profile with a tight output cap. For Mantle, set base_url to https://bedrock-mantle.us-east-1.api.aws/openai/v1 and the model to openai.gpt-6.1-sol. To test caching, send the same long prefix twice and check input_tokens_details.cached_tokens; zero means no caching.

What to do Monday morning

1. Default to global.openai.gpt-6.1-sol. Require a written residency reason for any premium path.
2. If you need US-only processing, choose between US cross-Region and Mantle by API surface, since the price is identical.
3. Set max_output_tokens from observed output lengths, and watch CloudWatch OutputTokenCount against your quota.
4. Add a guard that keeps prompts under 272K tokens.
5. Don't design around explicit caching, Priority or Flex.

The sticker price is the easy part: endpoint choice, quota burndown and the 272K line decide what you actually pay and how often production gets throttled.