GPT‑6.1 Sol brings Astra’s performance to a fifth of the price

On September 29 2026, OpenAI officially announced GPT‑6.1 Sol, a week after the unveiling of GPT‑6 Sol. The new variant is not available in the general ChatGPT chat at launch, but it is included in the ChatGPT Work offering and in Codex for Plus, Pro, Business, Enterprise and Edu plans. OpenAI emphasizes that the decision to limit initial access responds to the need to validate performance in professional environments before opening it to the mass market. This early‑access restriction enables the company to gather usage data and feedback in controlled settings, ensuring the model’s reliability before a broader rollout.

Availability and platforms

GPT‑6.1 Sol is deployed exclusively within products aimed at enterprise and educational users. In ChatGPT Work, subscribers can enable the model from the advanced model settings section, while in Codex the integration is performed via a “premium model” option that replaces the default engine. The model’s absence from the general chat allows OpenAI to collect usage metrics and feedback in controlled environments, reducing the risk of exposure to unexpected errors before a possible expansion. By limiting availability, OpenAI can closely monitor system stability and adjust parameters based on feedback from businesses and educational institutions that are piloting the model.

Performance versus GPT‑6 Astra

OpenAI states that GPT‑6.1 Sol approaches the level of GPT‑6 Astra in agentic programming, computer use and professional tasks, maintaining a factual error rate of 4.1 % according to its own internal measurements. By comparison, Astra records a factual error rate of 4.0 % under the same tests. The company highlights that, although the difference is minimal, the cost reduction offsets the slight loss of precision, positioning Sol as an attractive option for workloads that prioritize cost‑benefit ratio. Thus, the 0.1‑percentage‑point gap in error rate is considered negligible relative to the substantial savings achieved with the cheaper model.

In terms of capabilities, GPT‑6.1 Sol exhibits advanced reasoning that enables it to generate agent code, manipulate virtual desktop environments and execute complex workflows without direct human intervention. Users report that the model can write automation scripts, debug code snippets and provide software architecture recommendations with quality comparable to Astra. Moreover, the “computer use” capability includes interaction with office applications, opening the door to virtual assistants capable of drafting documents, creating presentations and analyzing data in real time. These features allow developers and technical teams to automate repetitive tasks and obtain design suggestions while working directly within familiar software environments.

Pricing structure and API

OpenAI has published a pricing table for the GPT‑6.1 Sol API that reflects its goal of being five times cheaper than Astra. The entry cost is set at $2 per million input tokens, while cached storage is priced at $0.10 per million cached tokens. Output generation is billed at $10 per million produced tokens. These figures represent a significant reduction compared with the announced prices for Astra, which hover around $10 per million input tokens and $50 per million output tokens. This pricing policy aims to make cutting‑edge models accessible to a larger user base by dramatically lowering the per‑token cost.

• Input: $2 / 1 M tokens
• Cache: $0.10 / 1 M tokens
• Output: $10 / 1 M tokens

With this tariff structure, developers can design applications that consume large volumes of data without the cost becoming prohibitive. For example, a log‑analysis tool that sends 5 M input tokens and receives 2 M output tokens would incur a monthly expense of roughly $30, versus the $250 that using Astra at prior prices would entail. OpenAI has also enabled usage‑based billing plans that allow startups and research teams to test the model without long‑term commitments. These flexible billing options facilitate experimentation and prototyping for small teams that want to leverage the model’s capabilities without having to invest large sums up front.

Despite the advances, OpenAI acknowledges that GPT‑6.1 Sol maintains a slightly higher factual error rate than Astra and that, because it is not available in the general chat, its initial adoption is limited to users with enterprise or educational subscriptions. The company warns that the model remains susceptible to hallucinations in highly specialized domains and that accuracy can vary depending on prompt quality. These limitations are part of the iteration process that OpenAI has communicated as essential for improving future versions. In other words, the company continues to work on reducing errors and enhancing model robustness before expanding it to a broader audience.

The market implications are notable. Companies that previously deemed the use of state‑of‑the‑art models prohibitive due to cost can now access capabilities close to Astra’s at a fraction of the price, encouraging internal process automation, code‑assistant creation and AI integration into everyday workflows. Additionally, availability in educational plans opens the door to training the next generation of developers with advanced AI tools without institutions having to allocate large budgets. This democratization of cutting‑edge AI technology is expected to spur innovation across both the private sector and academia.

In summary, the launch of GPT‑6.1 Sol marks a turning point: professional users receive a model that combines near‑Astra performance with a pricing model five times lower. OpenAI has implemented a strategy that prioritizes adoption in enterprise and academic environments while continuing to collect data to close the precision gap. Going forward, the conversation around generative AI will focus not only on model power but also on the economic viability of large‑scale deployment. Thus, the emphasis shifts toward pairing high performance with controlled costs, paving the way for broader adoption of generative artificial intelligence.