OpenAI launches GPT-6.2 Astra and Sol with lower costs
The new models offer a split between a high-power reasoning tier and a cost-efficient everyday tier for professional workflows.

The short version
- GPT-6.2 Astra is the new frontier model for hard, multi-step professional tasks like building 3D scenes.
- GPT-6.2 Sol serves as the everyday model, offering high performance at a much lower cost per task.
- Astra's API prices dropped by 20% compared to GPT-6 Astra, with cached input costs falling by 60%.
- Both models support a 1,050,000-token context window and a 128,000 maximum output limit.
- Astra ranks second on the Artificial Analysis Intelligence Index with a score of 62.
Why it matters
For businesses, this release provides a clearer choice between high-reasoning capabilities and high-volume efficiency. The significant price cuts on the Astra model make it more viable for complex, long-horizon automation projects.
A new two-tier model strategy
OpenAI has officially released two new closed models, GPT-6.2 Astra and GPT-6.2 Sol, designed for professional and agentic work. These models represent a strategic split in how users can interact with AI based on the complexity of their needs. Astra serves as the new frontier tier for difficult, long-horizon tasks, while Sol is positioned as the efficient choice for everyday use.
The release aims to provide better performance on complex tasks while simultaneously lowering the barrier to entry for developers. By offering two distinct paths, OpenAI allows users to choose between raw reasoning power and cost-effectiveness. Both models are now live in the OpenAI API under the names gpt-6.2-astra and gpt-6.2-sol.
These models are accessible through OpenAI’s hosted API, as well as ChatGPT Work and Codex. They share a significant 1,050,000-token context window, which allows them to process large amounts of information in a single session. However, they differ significantly in their knowledge cutoffs and intended use cases.
Astra for complex reasoning
GPT-6.2 Astra is built specifically for the hardest multi-step professional work. It is intended to succeed the previous GPT-6 Astra by producing cleaner documents, spreadsheets, and slides that adhere to specific templates and field conventions. It also excels at building more sophisticated frontends, games, and 3D scenes.
Benchmarks show that Astra is highly capable in specialized areas. It ranked first on the AA-3D metric with an Elo of 1,318, surpassing its competitors in 3D scene generation. However, it still faces challenges with accuracy, posting a 59% hallucination rate on the AA-Omniscience test.
Despite these hurdles, Astra is notable for being OpenAI's first released model to meet the High threshold for AI self-improvement. It provides frontier-level results while maintaining a lower cost per task than previous iterations. For users, this means the model can handle more complex logic without the previous price premiums.
Sol for everyday efficiency
GPT-6.2 Sol replaces the previous 6.1 Sol as the primary model for versatile, everyday tasks. While it is less powerful than Astra, it is designed to be fast and reliable for common requests. It maintains the same context limits as Astra but is optimized for a different price point.
The performance of Sol is particularly interesting when compared to the Astra model. In the FrontierCode 2.0 benchmark, Sol landed within 0.5 points of Astra but cost about one-fifth of the price. This makes it a strong candidate for high-volume coding tasks where extreme reasoning is not the primary requirement.
However, the update did not bring improvements to all areas for the Sol model. Its Coding Agent Index actually dropped by 3 points compared to the previous version. Despite this, it remains a primary tool for users who need a balance of intelligence and speed for standard operations.
Pricing and API updates
One of the most significant changes for developers is the updated pricing structure for the Astra model. The input and output prices for GPT-6.2 Astra are 20% lower than those of GPT-6 Astra. Additionally, the cost for cached input has been reduced by 60%, making it cheaper to reuse previous parts of a conversation.
The specific rates for GPT-6.2 Astra are $8.00 for 1M input tokens and $40.00 for 1M output tokens. In contrast, GPT-6.2 Sol maintains the same rates as the previous 6.1 Sol model, which are $2.00 for input and $10.00 for output. These prices remain steady until November 30, at which point some rates may adjust.
OpenAI also provides specific instructions for optimizing these new models. They recommend keeping prompt prefixes unchanged and appending new instructions to the end of prompts. This practice helps improve cache reuse and preserves the model's reasoning across multiple turns in a conversation.
How it compares to rivals
When measured against competitors like Claude Fable 5.2 and Gemini 4 Carbon, Astra shows mixed results. It scored 62 on the Artificial Analysis Intelligence Index, placing it second behind Claude Fable 5.2, which scored 65. However, Astra outperformed Fable 5.2 on the AA-3D benchmark, showing its strength in 3D generation.
In terms of legal applications, Astra's performance varies depending on the configuration. On the Harvey LAB-AA, the standard Astra passed 15.0% of legal assignments, while the specific Astra for Law configuration passed 20.8%. These figures highlight that while the model is powerful, it still requires specific tuning for niche professional fields.
Ultimately, the release offers a clear trade-off between cost and capability. Astra provides high-end reasoning for complex tasks at a lower price point than before, while Sol offers a budget-friendly way to access high-level performance. Developers can now choose the specific tool that fits their project's budget and complexity requirements.
What to watch
- How the 59% hallucination rate of Astra will affect its adoption in professional document generation.
- Whether the 20% price drop will lead to a significant increase in agentic workflow adoption.
Try it
If you use the OpenAI API, try appending new instructions to your existing prompt prefixes to see if you can improve cache reuse and reasoning consistency.
Sources and links
First reported by marktechpost.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.