OpenAI's GPT-6.1 Sol Launches on Amazon Bedrock, Claims Near-Astra Reasoning at Fraction of Cost

GPT-6.1 Sol Arrives on Amazon Bedrock

OpenAI's GPT-6.1 Sol is now generally available on Amazon Bedrock, positioned as an upgrade to GPT-6 Sol aimed at agentic coding, computer use, and professional document workflows. AWS and OpenAI announced the release through a joint blog post, describing it as bringing "near-Astra intelligence" — a reference to OpenAI's higher-tier GPT-6 Astra model — to everyday enterprise tasks at lower cost.

Benchmark Claims

According to OpenAI, GPT-6.1 Sol matches GPT-6 Astra's performance on the DeepSWE v1.1 software engineering benchmark while costing roughly one-fifth as much per task. OpenAI also claims the model exceeds the best previously recorded GPT-6 Sol score on that benchmark by 6.4 percentage points, achieved using a lower reasoning-effort setting than the earlier result required. No absolute benchmark scores, context window size, or per-token pricing were disclosed in the announcement. Independent verification of these figures is not yet available.

On document analysis, OpenAI states GPT-6.1 Sol approaches GPT-6 Astra's performance and improves over GPT-6 Sol on multistep workflows spanning business tools. The company also claims gains in evaluations covering transparency, respecting user intent, and adhering to explicit restrictions — areas relevant to agents operating with tool access and elevated permissions.

Deployment and Use Cases

The model integrates with Codex, OpenAI's coding agent, which can be configured to run on GPT-6.1 Sol via Amazon Bedrock for tasks spanning investigation, implementation, and testing across repositories, local files, and terminals. Codex is accessible through a desktop app, CLI, and supported IDEs. AWS-specific development is supported through the Agent Toolkit for AWS, which connects Codex to AWS documentation, APIs, and services from a single terminal command.

For document-heavy and multistep business workflows, ChatGPT Work — available in the desktop app — can synthesize information across files and applications into finished outputs. Developers can also build custom applications directly against Bedrock APIs, including internal tools, multi-system agents, and customer-facing applications.

Infrastructure and Data Controls

AWS is emphasizing enterprise governance features for the deployment. Access is controlled through AWS Identity and Access Management (IAM) policies, with invocation auditing via AWS CloudTrail. Traffic can be restricted to private network boundaries using VPC endpoints powered by AWS PrivateLink. AWS states inference runs on hardware-isolated infrastructure with zero-operator access, meaning AWS staff cannot view prompts or completions during inference. Customer data is not used for model training, and using the model on Bedrock does not require opting into data sharing with OpenAI.

Classifier-flagged traffic for abuse detection is retained by AWS for up to 30 days and processed programmatically; enterprises can request zero data retention through their AWS account team.

What This Means

This release is notable less for raw capability gains and more for cost efficiency in agentic settings, where reasoning quality determines how many tool calls and retries a task requires. If OpenAI's DeepSWE v1.1 claim holds up under independent testing, GPT-6.1 Sol could shift the cost calculus for teams running high-volume coding agents, since total task cost is driven by both per-token pricing and the number of interactions needed to reach a correct result. The lack of disclosed absolute benchmark scores, context window size, and pricing makes it difficult to compare GPT-6.1 Sol directly against competing models like Claude or Gemini variants on Bedrock. Enterprises evaluating this model will need to run their own cost-per-task comparisons rather than rely on OpenAI's relative percentage claims alone.