I Tried Out OpenAI GPT-6.2 Astra / Sol, Now Generally Available on Amazon Bedrock

I Tried Out OpenAI GPT-6.2 Astra / Sol, Now Generally Available on Amazon Bedrock

GPT-6.2 on Bedrock worked via Converse, InvokeModel, the OpenAI-compatible API, and the OpenAI SDK using only SigV4-signed temporary credentials. Per OpenAI's official pricing, Astra is 20% cheaper on input/output and 60% cheaper on cached input than GPT-6 Astra.
2026.11.18

This page has been translated by machine translation. View original

Introduction

On November 17, 2026, the general availability of OpenAI GPT-6.2 Astra and GPT-6.2 Sol on Amazon Bedrock was announced.

https://aws.amazon.com/jp/about-aws/whats-new/2026/11/openai-gpt-6-2-astra-sol-amazon-bedrock/

Astra has an Ultrafast mode for latency-critical use cases. In this article, I checked whether both models can be called from bedrock-runtime (Converse, InvokeModel, and the OpenAI-compatible API). I also called them from bedrock-mantle (OpenAI-compatible API), and tried specifying Ultrafast with Astra.

What I Tested

All calls used temporary credentials obtained from an IAM role. The credentials were set in the environment variables AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, and AWS_SESSION_TOKEN. Requests from curl and the OpenAI SDK were also signed with SigV4. The region is us-east-1 unless otherwise noted.

Model IDs and Inference Profiles

First, get the list of models.

aws bedrock list-foundation-models --region us-east-1 --by-provider openai --query "modelSummaries[?contains(modelId,'gpt-6.2')]" --output json

The entry for Astra is as follows.

{
    "modelArn": "arn:aws:bedrock:us-east-1::foundation-model/openai.gpt-6.2-astra",
    "modelId": "openai.gpt-6.2-astra",
    "modelName": "GPT-6.2 Astra",
    "providerName": "OpenAI",
    "inputModalities": [
        "TEXT",
        "IMAGE"
    ],
    "outputModalities": [
        "TEXT"
    ],
    "responseStreamingSupported": true,
    "customizationsSupported": [],
    "inferenceTypesSupported": [
        "INFERENCE_PROFILE"
    ],
    "modelLifecycle": {
        "status": "ACTIVE",
        "startOfLifeTime": "2026-11-16T10:00:00+00:00"
    }
}

The entry for Sol has the same values as Astra except for modelId, modelName, and modelArn, and its model ID is openai.gpt-6.2-sol. Both models are also included in the lists for ap-northeast-1 and us-west-2. Since inferenceTypesSupported is only INFERENCE_PROFILE, you pass the inference profile ID rather than the model ID when calling bedrock-runtime.

aws bedrock list-inference-profiles --region us-east-1 --query "inferenceProfileSummaries[?contains(inferenceProfileId,'gpt-6.2')].{id:inferenceProfileId,name:inferenceProfileName,status:status,type:type}" --output json
[
    {
        "id": "us.openai.gpt-6.2-astra",
        "name": "US OpenAI GPT-6.2 Astra",
        "status": "ACTIVE",
        "type": "SYSTEM_DEFINED"
    },
    {
        "id": "us.openai.gpt-6.2-sol",
        "name": "US OpenAI GPT-6.2 Sol",
        "status": "ACTIVE",
        "type": "SYSTEM_DEFINED"
    },
    {
        "id": "global.openai.gpt-6.2-astra",
        "name": "Global OpenAI GPT-6.2 Astra",
        "status": "ACTIVE",
        "type": "SYSTEM_DEFINED"
    },
    {
        "id": "global.openai.gpt-6.2-sol",
        "name": "Global OpenAI GPT-6.2 Sol",
        "status": "ACTIVE",
        "type": "SYSTEM_DEFINED"
    }
]

The ap-northeast-1 list includes only the two global. profiles for Astra and Sol. In us-east-1 and us-west-2, both us. and global. are included.

Converse

Specifying the inference profile ID global.openai.gpt-6.2-astra returns a Converse response.

aws bedrock-runtime converse --region us-east-1 --model-id global.openai.gpt-6.2-astra --messages '[{"role":"user","content":[{"text":"Say ok"}]}]' --inference-config '{"maxTokens":64}' --output json
{
    "output": {
        "message": {
            "role": "assistant",
            "content": [
                {
                    "text": "ok"
                }
            ]
        }
    },
    "stopReason": "end_turn",
    "usage": {
        "inputTokens": 8,
        "outputTokens": 5,
        "totalTokens": 13,
        "cacheReadInputTokens": 0,
        "cacheWriteInputTokens": 0
    },
    "metrics": {
        "latencyMs": 875
    }
}

For both Astra and Sol, I tried global. and us. in us-east-1 and us-west-2, and global. in ap-northeast-1. In every case, the response text was ok, stopReason was end_turn, and usage was 8 input, 5 output, and 13 total tokens.

InvokeModel

The InvokeModel body uses the Chat Completions format. Save the following content as invoke-body.json.

{"messages":[{"role":"user","content":"Say ok"}],"max_completion_tokens":64}
aws bedrock-runtime invoke-model --region us-east-1 --model-id global.openai.gpt-6.2-astra --body fileb://invoke-body.json --content-type application/json --accept application/json invoke-astra.json && cat invoke-astra.json

The contents of invoke-astra.json are as follows.

{"choices":[{"finish_reason":"stop","index":0,"message":{"annotations":[],"content":"ok","refusal":null,"role":"assistant"}}],"created":1794965410,"id":"chatcmpl-ftxbgwd6pzi5vg4weseqhatxvsqgavhk6th66wzohawzpghplikq","model":"global.openai.gpt-6.2-astra","object":"chat.completion","service_tier":"default","usage":{"completion_tokens":5,"completion_tokens_details":{"accepted_prediction_tokens":0,"audio_tokens":0,"reasoning_tokens":0,"rejected_prediction_tokens":0},"prompt_tokens":8,"prompt_tokens_details":{"audio_tokens":0,"cache_write_tokens":0,"cached_tokens":0},"total_tokens":13},"system_fingerprint":null}

The model field in the response contains the specified inference profile ID as is. The same applies to Sol.

OpenAI-Compatible API

The OpenAI-compatible endpoint of bedrock-runtime can be called by signing with curl's --aws-sigv4. The signing service name is bedrock. Save the following content as responses-body-astra.json as the Responses body.

{"model":"global.openai.gpt-6.2-astra","input":"Say ok","store":false,"max_output_tokens":64}
curl --silent --show-error \
  --aws-sigv4 'aws:amz:us-east-1:bedrock' \
  --user "${AWS_ACCESS_KEY_ID}:${AWS_SECRET_ACCESS_KEY}" \
  -H "x-amz-security-token: ${AWS_SESSION_TOKEN}" \
  -H 'Content-Type: application/json' \
  --data-binary @responses-body-astra.json \
  --write-out '\nHTTP_STATUS=%{http_code}\n' \
  https://bedrock-runtime.us-east-1.amazonaws.com/openai/v1/responses

For Chat Completions, make the end of the URL /openai/v1/chat/completions. Save the following content as chat-body-astra.json and pass it to --data-binary.

{"model":"global.openai.gpt-6.2-astra","messages":[{"role":"user","content":"Say ok"}],"max_completion_tokens":64}

For Sol, change the model in the body to global.openai.gpt-6.2-sol. I called both Responses and Chat Completions for each of Astra and Sol, and all four combinations returned HTTP 200. The output and usage are the same as with Converse. The Responses response has status completed, and the text is in output[0].content[0].text.

Full Responses response (Astra)
{"background":false,"billing":{"payer":"developer"},"completed_at":1794965435,"created_at":1794965435,"error":null,"frequency_penalty":0.0,"id":"resp_nzfit5r67v7ltukd53wukoc4xat7amw7j3wt77su3xja5bspvq3q","incomplete_details":null,"input_transformations":[],"instructions":null,"max_output_tokens":64,"max_tool_calls":null,"metadata":{},"model":"global.openai.gpt-6.2-astra","object":"response","output":[{"content":[{"annotations":[],"logprobs":[],"text":"ok","type":"output_text"}],"id":"msg_797516b64f5a52fd9764a86af9aaaee6","phase":"final_answer","role":"assistant","status":"completed","type":"message"}],"parallel_tool_calls":true,"presence_penalty":0.0,"previous_response_id":null,"prompt_cache_key":null,"prompt_cache_retention":"in_memory","reasoning":{"summary":null,"effort":"medium","context":"all_turns"},"safety_identifier":null,"service_tier":"default","status":"completed","store":false,"temperature":1.0,"text":{"format":{"type":"text"},"verbosity":"medium"},"tool_choice":"auto","tools":[],"top_logprobs":0,"top_p":1.0,"truncation":"disabled","usage":{"input_tokens":8,"input_tokens_details":{"cache_write_tokens":0,"cached_tokens":0},"output_tokens":5,"output_tokens_details":{"reasoning_tokens":0},"total_tokens":13},"user":null,"moderation":null}
HTTP_STATUS=200

bedrock-mantle

For bedrock-mantle, the signing service name is bedrock-mantle, the URL is https://bedrock-mantle.<region>.api.aws/openai/v1/responses, and the model ID is openai.gpt-6.2-astra without a prefix. Save the following content as mantle-responses-body-astra.json as the body.

{"model":"openai.gpt-6.2-astra","input":"Say ok","store":false,"max_output_tokens":64}
curl --silent --show-error \
  --aws-sigv4 'aws:amz:us-east-1:bedrock-mantle' \
  --user "${AWS_ACCESS_KEY_ID}:${AWS_SECRET_ACCESS_KEY}" \
  -H "x-amz-security-token: ${AWS_SESSION_TOKEN}" \
  -H 'Content-Type: application/json' \
  --data-binary @mantle-responses-body-astra.json \
  --write-out '\nHTTP_STATUS=%{http_code}\n' \
  https://bedrock-mantle.us-east-1.api.aws/openai/v1/responses
Region and API Astra Sol
us-east-1 (Responses) HTTP 200 HTTP 200
us-east-1 (Chat Completions) HTTP 200 HTTP 200

The output and usage are the same as with Converse.

OpenAI SDK

To call from the OpenAI SDK, attach a SigV4 signature using an httpx authentication class. Save the following code as sdk_sigv4.py.

"""OpenAI SDK から Bedrock の OpenAI 互換エンドポイントを SigV4 署名で呼ぶ。

認証情報は環境変数 AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY / AWS_SESSION_TOKEN から読む。
接続先は環境変数 TARGET (runtime | mantle)、REGION、MODEL_ID で切り替える。
"""

import json
import logging
import os

import httpx
from botocore.auth import SigV4Auth
from botocore.awsrequest import AWSRequest
from botocore.credentials import Credentials
from openai import OpenAI

logging.basicConfig(level=logging.INFO, format="%(message)s")
logger = logging.getLogger(__name__)

TARGET = os.environ.get("TARGET", "runtime")
REGION = os.environ.get("REGION", "us-east-1")
MODEL_ID = os.environ["MODEL_ID"]

if TARGET == "runtime":
    SIGNING_SERVICE = "bedrock"
    BASE_URL = f"https://bedrock-runtime.{REGION}.amazonaws.com/openai/v1"
else:
    SIGNING_SERVICE = "bedrock-mantle"
    BASE_URL = f"https://bedrock-mantle.{REGION}.api.aws/openai/v1"

credentials = Credentials(
    os.environ["AWS_ACCESS_KEY_ID"],
    os.environ["AWS_SECRET_ACCESS_KEY"],
    os.environ.get("AWS_SESSION_TOKEN"),
)

class SigV4Signer(httpx.Auth):
    requires_request_body = True

    def auth_flow(self, request: httpx.Request):  # type: ignore[no-untyped-def]
        aws_request = AWSRequest(
            method=request.method,
            url=str(request.url),
            data=request.content,
            headers={"content-type": request.headers.get("content-type", "application/json")},
        )
        SigV4Auth(credentials, SIGNING_SERVICE, REGION).add_auth(aws_request)
        for key, value in aws_request.headers.items():
            request.headers[key] = value
        yield request

def main() -> None:
    client = OpenAI(
        api_key="unused-sigv4",
        base_url=BASE_URL,
        http_client=httpx.Client(auth=SigV4Signer(), timeout=120.0),
    )
    response = client.responses.create(
        model=MODEL_ID,
        input="Can you explain the features of Amazon Bedrock?",
        max_output_tokens=2000,
    )
    logger.info("=== target: %s / region: %s / model: %s ===", TARGET, REGION, MODEL_ID)
    logger.info("=== status: %s ===", response.status)
    logger.info("=== output_text ===")
    logger.info("%s", response.output_text)
    logger.info("=== usage ===")
    logger.info("%s", json.dumps(response.usage.model_dump(), indent=2))

if __name__ == "__main__":
    main()

In the directory where you placed sdk_sigv4.py, run the following command. This is an example of calling Astra with bedrock-runtime.

docker run --rm -e AWS_ACCESS_KEY_ID -e AWS_SECRET_ACCESS_KEY -e AWS_SESSION_TOKEN -e TARGET=runtime -e REGION=us-east-1 -e MODEL_ID=global.openai.gpt-6.2-astra -v "$(pwd):/app:ro" -w /app python:3.12-slim bash -c 'pip install -q --root-user-action=ignore --disable-pip-version-check openai botocore httpx 2>&1 | tail -2; python sdk_sigv4.py'

The combinations of TARGET and MODEL_ID are the following four.

TARGET MODEL_ID for Astra MODEL_ID for Sol
runtime global.openai.gpt-6.2-astra global.openai.gpt-6.2-sol
mantle openai.gpt-6.2-astra openai.gpt-6.2-sol

All four combinations had status completed. The versions are not pinned in pip install. For reference, the versions I confirmed afterward with the same image are Python 3.12.15, openai 3.54.0, botocore 1.44.5, and httpx 0.28.1.

Ultrafast

With Astra, adding "service_tier":"ultrafast" to the Responses body runs it in Ultrafast. Save the following content as ultrafast-body-astra-global.json and pass it to --data-binary with the same curl as in the OpenAI-compatible API section. For Chat Completions as well, add the same key to the body.

{"model":"global.openai.gpt-6.2-astra","input":"Say ok","store":false,"max_output_tokens":64,"service_tier":"ultrafast"}

The following four routes returned HTTP 200, the service_tier in the response was ultrafast, and the output was ok. For Standard calls without service_tier specified, the service_tier in the response is default.

Endpoint Model ID API
bedrock-runtime global.openai.gpt-6.2-astra Responses
bedrock-runtime us.openai.gpt-6.2-astra Responses
bedrock-runtime global.openai.gpt-6.2-astra Chat Completions
bedrock-mantle openai.gpt-6.2-astra Responses

Pricing

As of writing, the unit prices for GPT-6.2 are not listed on the OpenAI model cards list page for Bedrock. They are expected to be listed in the future.

https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards-openai.html

In the model cards for GPT-6 Astra and GPT-6.1 Sol, the unit prices for the global inference profiles were the same as OpenAI's official prices. If the same holds for GPT-6.2, the unit prices for the global inference profiles will be the same as in the following tables.

The following tables show OpenAI's official prices, not Bedrock unit prices. Values are prices per 1 million tokens, for short contexts.

GPT-6.2 Astra

Model Input Output Cached input Cache write (30 min) Cache write (12 hours)
gpt-6.2-astra (Standard) $8.00 $40.00 $0.40 $10.00 $20.00
gpt-6.2-astra (Ultrafast, until February 28, 2027) $24.00 $120.00 $1.20 $30.00 $60.00
gpt-6.2-astra (Ultrafast, from March 1, 2027) $48.00 $240.00 $2.40 $60.00 $120.00
gpt-6-astra (Standard) $10.00 $50.00 $1.00 $12.50 $25.00

Compared with GPT-6 Astra, the Standard prices for input, output, and cache writes (30 min and 12 hours) are all 20% lower, and cached input is 60% cheaper, going from $1.00 to $0.40. Ultrafast is 3 times Standard (until February 28, 2027) and 6 times (from March 1, 2027), and the switchover date is in Pacific Time.

GPT-6.2 Sol

Model Input Output Cached input Cache write (30 min) Cache write (12 hours)
gpt-6.2-sol $2.00 $10.00 $0.10 $2.50 $5.00
gpt-6.1-sol $2.00 $10.00 $0.10 $2.50 $5.00

The offered prices have not changed from GPT-6.1 Sol.

https://developers.openai.com/api/docs/pricing

Summary

I confirmed that GPT-6.2 Astra and GPT-6.2 Sol are available on Bedrock.

In a post, ARC Prize introduces GPT-6.2 Astra as the first model to achieve 100% on ARC-AGI-3.

https://x.com/arcprize/status/2122751233442156678

Astra's token prices have also been reduced. If you are using Sol or an older version of Astra, I recommend evaluating GPT-6.2.


AI白書2026 配布中

クラスメソッドが独自に行なったAI診断調査をもとに、企業のAI活用の現在地を調査レポートとしてまとめました。企業規模別の活用度傾向に加え、規模を超えてAI活用を進める企業に共通する取り組みまで、自社の現在地を捉えるためのヒントにぜひ。

AI白書2026

無料でダウンロードする

Share this article

AWSのお困り事はクラスメソッドへ