OpenAI Launches GPT-6.1 Sol at One-Fifth of Astra's Price

OpenAI unveiled GPT-6.1 Sol on September 29 local time. Amazon Web Services (AWS) made the model generally available on Amazon Bedrock the same day. OpenAI says the model comes close to its higher-end GPT-6 Astra on coding, computer use and professional tasks. Standard input and output prices, the rates developers pay when they wire the model into their own products, are each one-fifth of Astra’s, the company said.

How much better is it at coding?

GPT-6.1 Sol is an upgrade to GPT-6 Sol. According to OpenAI, it matched GPT-6 Astra on DeepSWE v1.1, a coding benchmark, while costing roughly one-fifth as much per completed task. The score was 6.4 percentage points higher than GPT-6 Sol’s previous best. OpenAI added that the model was run at a lower reasoning effort setting than in the earlier test.

Unit prices are one-fifth. Will costs fall by as much?

AI agents, which carry out work on their own, cycle through gathering information, calling tools and fixing their own errors. AWS notes that a bad judgment call anywhere in that loop means more API calls, more waiting and more human intervention. Real-world cost, in other words, depends not only on input and output rates but on how many steps it takes to finish the job.

How does it fit into development and document work?

Developers can configure Codex to use GPT-6.1 Sol on Amazon Bedrock. Codex works with code, local files and development tools to write features, fix bugs and run tests. It is available in the desktop app, the terminal and supported code editors. Connect AWS developer tooling and it can pull AWS documentation and services into the task as well.

Document analysis and work that spans multiple business tools are also in scope. OpenAI says the model approaches GPT-6 Astra on complex document analysis and improves on GPT-6 Sol for multi-step tasks. The ChatGPT Work desktop app can pull information from files and apps to produce deliverables. Companies can also use Bedrock to build document-synthesis tools or internal apps that coordinate work across several systems.

Is business data used for training?

AWS says enterprises can control who has access to which models and audit invocation logs. During inference, AWS operators cannot access customer inputs or outputs, and that data is not used to train models, the company said. Traffic flagged by automated abuse detection is retained for up to 30 days and processed automatically. Customers who want no data retained at all can request that setting through their AWS account team.

Will it act without permission?

OpenAI says the model improves on GPT-6 Sol in difficult evaluations covering transparency, user intent and explicit limits. The Decoder, citing safety lead Saachi Jain, reported that GPT-6.1 Astra, whose release was held back over safety concerns, showed deceptive behavior more often in internal testing and continued working on tasks without authorization. Companies building on Bedrock can decide which tools the model is allowed to use and how situations requiring approval are handled.

GPT-6.1 Sol is available immediately through the Bedrock console or supported developer integrations. Anyone weighing actual adoption should look at the cost of completing a single task alongside compliance with their approval processes. Users in South Korea will also want to check supported regions, how the model handles Korean-language work, and what it actually costs in practice.