“It Blew Me Away” | Box’s First Look at GPT-6 Astra
Summary
GPT-6 Astra demonstrates frontier-level reasoning by correctly handling multi-layered financial calculations across countries—distinguishing between applied vs. unapplied tax incentives that previous models would double-count. Box's AI lead says this is the first time he's seen a model get this right, unlocking use cases requiring high judgment and extreme accuracy for enterprise customers.
Key Takeaways
- Astra excels at nuanced financial analysis by accurately tracking context across multiple documents—distinguishing that Argentina's 20% incentive was already applied while Brazil's 30% incentive needed application, preventing costly calculation errors.
- Complex multi-file reasoning is now viable for enterprise deployment. Box can now confidently pitch Astra to customers previously unable to automate high-judgment workloads requiring extreme accuracy and contextual understanding.
- Frontier AI models are reaching decision-support capability for regulated financial/compliance use cases where a single calculation error compounds into major financial impact—marking a transition from experimental to production-viable tools.
- The inflection point isn't raw capability but reliability on edge cases. Box's excitement stems from Astra handling the 'easy thing to miss'—the contextual trap that would have broken previous models, suggesting consistency improvements matter more than base performance.
- Enterprise partnerships should test models on actual customer workflows with real complexity before broad deployment. Box's validation approach—giving Astra real profit/cost/tax files—is the vetting strategy needed for mission-critical use cases.
Related topics
Transcript Excerpt
absolutely at the frontier just a solid model for some of these use cases [music] that have been just unbreakable in our stack. >> So Yash, you lead AI at Box. [music] What were your first impressions about Astra? It blew me away um in terms of just [music] how far this model has come in just doing these really heavy workloads that we see for our customers [music] right some of our customers have content that is very valuable um very intense very [music] complex >> I love that I'd love to see an example of one >> we gave it a couple files we said here's some profit analysis here's some cost analysis and I give that a file that says here is our some incentive tax structures. [music] Now, the really interesting trick was there was two incentive tax structures. One of them was not applied to …