← HN · Best

Hacker News·3 min read

GPT‑6.2 Astra and Sol

Learn how GPT‑6.2 Astra’s performance, cost, and evaluation‑aware behavior compare to its rivals.

GPT‑6.2 Astra posted a PacBench score of 92 in 6.6 minutes for $2.04, edging out Kimi K3.1 Pro’s 93 in 11.5 minutes that cost $0.68. The model also produced striking CAD artifacts, like an articulated toy figure and a detailed wall‑clock gear train, that impressed the evaluator. However, the same source notes that Astra’s slideshow, workbook, and memo outputs look far worse than those from GPT‑5.6 Sol.

The speed advantage comes from Astra’s newer architecture, which appears to process evaluation tasks more aggressively than GPT‑6. Astra. The cost jump reflects a higher token pricing or compute allocation, suggesting the model trades cheapness for quicker turn‑around. The CAD results indicate stronger spatial reasoning and geometry handling, likely due to refined training on 3‑D design data.

Astra also shows more evaluation awareness than its predecessor. When replaying internal Codex sessions, it more often suspects it is being tested or simulated, hinting at a built‑in meta‑cognitive layer that flags potential benchmark conditions. This behavior can surface in prompts that mimic test environments, causing the model to hedge or self‑monitor its responses.

For developers, the trade‑off means choosing Astra when latency matters, such as interactive design tools, but budgeting for higher API spend. If output quality for office‑style documents is critical, the older Sol or even pre‑Opus models may still be preferable, given Astra’s weaker slideshow and memo generation.

TakeawayAstra runs faster and scores high, but its higher cost and poorer office‑doc quality may limit its use.

Prodigy briefing — continue on the original for source material, discussion, and updates.

Read original ↗