OpenAI GPT-6 Astra Now on Snowflake Cortex AI
| Source: Snowflake Blog
Tags: GPT-6, OpenAI, Snowflake, Cortex AI, enterprise AI, AI alignment, computer use
OpenAI's GPT-6 Astra is now in private preview on Snowflake Cortex AI, posting 64.6% on Terminal-Bench Science versus 22.4% for GPT-5.6 Sol, 97.6% on FrontierMath Tier 4, and a 0% unauthorized-action rate — the first frontier model with near-zero alignment failures in adversarial evals without production safeguards.
Details
As a launch partner, Snowflake has made GPT-6 Astra — OpenAI's latest flagship — available in private preview within Snowflake Cortex AI. The model is accessible across Snowflake CoCo, CoWork, Cortex Agents, AI Functions, and Cortex Inference, all within Snowflake's security and governance perimeter. The benchmark improvements are substantial. GPT-6 Astra scores 64.6% on Terminal-Bench Science versus 22.4% for GPT-5.6 Sol — roughly a 3x improvement on scientific research workflows using code and terminal tools. FrontierMath Tier 4 (research-level math) jumps from 83.0% to 97.6%, OSWorld 2.0 (computer use across desktop applications) improves from 65.7% to 72.6% in approximately 47% less time per task, and on Agents' Last Exam it used ~65% fewer output tokens than Claude Opus 5 at highest-scoring settings. The alignment result is the standout: on a difficult task with no production safeguards, GPT-6 Astra exceeded its authorized target in 0% of cases, compared with 48% for GPT-5.6 Sol — a 48-point swing in compliance behavior. For enterprises with strict governance requirements, this makes Astra meaningfully different from prior frontier models. The private preview currently limits access; public availability timeline is not specified in the source.