Trying Out the Ultrafast Mode Supported by GPT-6.1 Sol on Amazon Bedrock
This page has been translated by machine translation. View original
Introduction
On October 8, 2026, GPT-6.1 Sol on Amazon Bedrock added support for Ultrafast mode.
In this article, I call GPT-6.1 Sol from the OpenAI SDK with ultrafast specified, and check the results and the evidence.
What is Ultrafast mode
It is a service tier for low latency that was made available earlier for GPT-6 Astra on September 30, 2026.
The model card states that the unit price is 6 times that of Standard.
Ultrafast costs six times as much as Standard.
These are the unit prices for Global CRIS (bedrock-runtime) when the input is 272K tokens or less. The unit is USD / 1 million tokens.
| Standard | Ultrafast | |
|---|---|---|
| Input | 2.00 | 12.00 |
| Output | 10.00 | 60.00 |
How to use it
Ultrafast is used by specifying ultrafast in service_tier of Bedrock's Responses API (/openai/v1/responses). There are two ways to call it: use the OpenAI SDK, or send an HTTP request signed with SigV4 yourself. In this article, I use the OpenAI SDK.
Calling with the OpenAI SDK
Using the OpenAI SDK (Python), I call the Responses API of Global CRIS. For authentication, I create a short-term token from temporary credentials with aws-bedrock-token-generator and pass it as the API key. To specify Ultrafast, you just add service_tier="ultrafast" to the arguments of responses.create.
The execution environment is Python 3.12.15, openai 3.26.1, and aws-bedrock-token-generator 1.1.0.
"""GPT-6.1 Sol を Bedrock (bedrock-runtime, Global CRIS) の Responses API で 1 回呼ぶ。
使い方: python call_sol.py <default|ultrafast>
"""
import sys
import time
from datetime import timedelta
from aws_bedrock_token_generator import provide_token
from openai import OpenAI
REGION = "us-east-1"
MODEL = "global.openai.gpt-6.1-sol"
PROMPT = "Amazon Bedrock を 20 文字以内で説明してください。"
MAX_OUTPUT_TOKENS = 300
tier = sys.argv[1]
assert tier in ("default", "ultrafast")
token = provide_token(region=REGION, expiry=timedelta(minutes=15))
client = OpenAI(
api_key=token,
base_url=f"https://bedrock-runtime.{REGION}.amazonaws.com/openai/v1",
)
kwargs = dict(model=MODEL, input=PROMPT, max_output_tokens=MAX_OUTPUT_TOKENS)
if tier == "ultrafast":
kwargs["service_tier"] = "ultrafast"
t0 = time.perf_counter()
resp = client.responses.create(**kwargs)
print(f"elapsed_s={time.perf_counter() - t0:.3f}")
print(resp.model_dump_json(indent=2))
python call_sol.py ultrafast
Results
I ran the same prompt once without service_tier and once with ultrafast. Reasoning effort was not specified (it was medium in the response).
| Specified | service_tier in response | Elapsed time | Input tokens | Output tokens (of which reasoning) | Output text |
|---|---|---|---|---|---|
| None | default | 3.018 sec | 19 | 117 (104) | AWS's generative AI foundation service |
| ultrafast | ultrafast | 1.710 sec | 19 | 197 (183) | AWS service that lets you use generative AI models |
The elapsed time is the time until the full response was returned in non-streaming mode, measured with time.perf_counter(), and each is a single measurement.
Here is an excerpt of the response JSON. ultrafast is returned in service_tier.
{
"model": "global.openai.gpt-6.1-sol",
"service_tier": "ultrafast",
"status": "completed",
"reasoning": {
"effort": "medium"
},
"usage": {
"input_tokens": 19,
"output_tokens": 197,
"output_tokens_details": {
"reasoning_tokens": 183
},
"total_tokens": 216
}
}
Recording in CloudTrail
Calls to the Responses API are recorded as CloudTrail management events (eventName: Responses). For calls made with ultrafast specified, serviceTier in additionalEventData contains ultrafast. For calls made without specifying it, serviceTier itself is absent.
{
"eventTime": "2026-10-09T06:05:27Z",
"eventSource": "bedrock.amazonaws.com",
"eventName": "Responses",
"awsRegion": "us-east-1",
"userAgent": "OpenAI/Python 3.26.1",
"requestParameters": {
"modelId": "global.openai.gpt-6.1-sol"
},
"additionalEventData": {
"callWithBearerToken": true,
"stream": false,
"inferenceRegion": "us-east-1",
"serviceTier": "ultrafast",
"inputTokens": 19,
"outputTokens": 197
}
}
Summary
At the time of the previous article on October 1, GPT-6.1 Sol on Amazon Bedrock could not use the ultrafast service tier, but this time I was able to confirm that it can be used from the OpenAI SDK via the Responses API.
The measurements were taken once each, and the question was a simple one that doesn't really require Sol, but the elapsed time was 1.7 seconds with ultrafast compared to 3.0 seconds with default, confirming a reduction.
For tasks that tend to take a long time, where you need to get responses from Sol on Bedrock even a little faster and can accept a token unit price 6 times that of Standard, please try the newly supported ultrafast.
