OpenAI releases GPT-6.2 Astra and Sol models
OpenAI's new models offer a choice between high-reasoning capabilities for complex workflows and a faster, more efficient version for everyday tasks.

The short version
- GPT-6.2 Astra achieved 61% accuracy on the BigFinanceBench, outperforming Claude Opus 5.5.
- GPT-6.2 Sol is 15% faster than its predecessor while using 13% fewer input tokens.
- Astra produced financial models with one-third the output tokens of Claude Opus 5.5.
- Astra correctly routed 98% of events in a simulated workday involving over 50 specialized agents.
Why it matters
The release highlights a shift toward specialized model selection, where users can choose between high-power reasoning for complex deals and high-speed efficiency for routine data tasks. It also signals a move toward multi-agent systems that can manage long-running, context-heavy professional workflows.
The new GPT-6.2 lineup
OpenAI has officially released two new models within its flagship family: GPT-6.2 Astra and GPT-6.2 Sol. These models are designed to handle different types of demands, ranging from high-level reasoning to high-speed, everyday queries. The release aims to provide users with more specific tools tailored to the complexity of their specific needs.
The Astra model is positioned as the primary choice for complex workflows that require deep reasoning. It is built to handle intricate tasks where accuracy and the ability to generate structured artifacts, like financial models, are paramount. By offering a more specialized approach, OpenAI allows for more precise control over how the AI handles difficult information.
In contrast, the Sol model focuses on efficiency and speed for lighter weight tasks. It is intended for users who need quick responses without the overhead of a larger reasoning engine. This dual-model approach allows for a more nuanced deployment of AI resources depending on the specific goal of the user.
Performance in financial tasks
Rogo conducted evaluations using their proprietary BigFinanceBench to see how these models handle real-world financial work. The results showed that GPT-6.2 Astra achieved 61% accuracy on the benchmark, which is higher than the 53% accuracy reported for Claude Opus 5.5. This suggests that Astra is currently a leader in precise financial reasoning and data interpretation.
When it comes to generating spreadsheets, Astra produced financial models that were comparable in quality to Claude Opus 5.5. However, it did so while generating roughly one-third as many output tokens. This reduction in tokens is significant because it directly translates to lower costs and reduced latency for the end user.
The Sol model showed different strengths, specifically in its ability to maintain existing accuracy while improving performance metrics. It achieved 58% accuracy on the BigFinanceBench, matching the performance of the previous GPT-6.1 Sol. For users who prioritize speed, Sol returned answers 15% faster than its predecessor while consuming 13% fewer input tokens.
Handling complex workflows
A major focus of the new evaluations is how models handle long-running workflows rather than isolated questions. In professional settings, work is rarely a single prompt; it involves a stream of information like emails, news events, or rumors. The models must be able to identify which new pieces of information are relevant to a specific, ongoing project.
To test this, researchers simulated a financial professional's workday using more than 50 specialized agents managing concurrent tasks. The goal was to see if the models could correctly route new events to the specific agent already responsible for that work. This type of coordination is essential for maintaining context across a busy, multi-tasking environment.
GPT-6.2 Astra performed well in these preliminary tests, correctly routing 98% of the events. This outperformed Claude Opus 5.5, which routed 91% of the events. These results suggest that Astra is particularly well-suited for environments where an AI must manage multiple moving parts and maintain a consistent understanding of a live transaction.
Where other models still lead
Despite the strengths of the new GPT-6.2 models, certain tasks still favor other existing AI systems. For example, presentation generation remains a challenge for the new models. Blind evaluators preferred the decks created by Claude Opus 5.5 in 94.8% of the comparisons, indicating it remains the leader for that specific output.
The data suggests that model performance is highly dependent on the specific task at hand. While Astra leads on precise financial questions and complex workflows, Claude Opus 5.5 remains the stronger choice for in-depth research and creating visual presentations. This highlights the importance of selecting the right tool for the right job rather than relying on a single universal model.
For users who do not want to manage these distinctions manually, platforms like Rogo are beginning to abstract the decision-making process. They use a bespoke fusion of models that are continuously fine-tuned to match a specific workflow with the best-suited model. This ensures that the most efficient and accurate tool is used automatically based on the user's current needs.
What to watch
- How the multi-agent routing accuracy scales as the number of agents increases beyond 50.
- Whether future updates to Astra will improve its performance in presentation generation to match Claude Opus 5.5.
Try it
Compare the output of a complex financial reasoning prompt across different models to see if the 'token reduction' in Astra actually simplifies the final result.
Sources and links
- Reportingrogo.com: the article we sourced
First reported by rogo.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.