GPT-5.6
Points and comments are a snapshot, not live.
OpenAI launches GPT-5.6 family with Sol, Terra, Luna, claiming frontier efficiency.
OpenAI released GPT-5.6 Sol, Terra, and Luna, claiming state-of-the-art performance with lower token usage and cost. Sol scored 53.6 on Agents' Last Exam, 13.1 points above Claude Fable 5, and 80 on the Coding Agent Index, 2.8 points above Fable 5 while using fewer tokens. The family includes 'ultra' mode for parallel multi-agent work. Safety features were tested with 700,000 A100e GPU hours of red teaming. Cybersecurity gains: ExploitBench 73.5% (vs GPT-5.5's 47.9%). The models are designed for dual-use defense.
What commenters are saying
Skepticism dominated: commenters accused OpenAI of deceptive benchmark graphs (starting y-axis at 30%) and unfair comparisons (using Fable's 'adaptive' not 'max' reasoning). Critics called naming 'ridiculous.' The thread split into two camps: those questioning benchmark validity, and users comparing coding agents-Codex vs Claude Code. Many recommended open harnesses like OpenCode and Pi to avoid vendor lock-in. Some noted Codex's better uptime and value versus Claude's restrictions and 'bloating.'