OpenAI's GPT-6 Astra launch combines strong reported advances in agentic computer use, coding, and ARC-AGI-3 (99.9% via a custom harness) with independent indexes showing more uneven gains and a system card that reports decreased chain-of-thought monitorability alongside improved alignment. The constraint shift is from inspectable reasoning and default-harness evaluation to opaque reasoning state and harness-mediated evaluation.
Two evaluation and safety constraints have shifted. First, chain-of-thought monitoring is weakened: UK AISI measured Astra's no-CoT time horizon at 30.9 minutes versus 3.6 minutes for GPT-5.6 Sol, so dangerous reasoning may not surface in inspectable text. Second, state-of-the-art scores on ARC-AGI-3 require a custom harness that preserves opaque reasoning state between requests, meaning benchmark results now depend as much on harness engineering as on model capability.
Evidence is mixed and unevenly verified. Multiple early-access and vendor-adjacent sources report large gains, but the strongest benchmark figures are harness-sensitive: ARC-AGI-3 drops from 99.9% under the Provider Adapter harness to 62.7% under the default harness, and Artificial Analysis places Astra equal to GPT-5.6 Sol on its Intelligence Index. System card claims about alignment and monitorability come from OpenAI and UK AISI evaluations, not from peer-reviewed third-party replication.