GPT-6 Astra base instructions run 16% longer than GPT-5.6 Sol's, cutting permission loops
WolframRvnwlf · x · 2026-09-05
A hands-on comparison of the base instruction templates for GPT-5.6 Sol and GPT-6 Astra published in OpenAI's Codex repository (as of September 5, 2026):
- Scale: Astra's template runs 3,282 words vs Sol's 2,826 — about 16% longer.
- Key behavioral shift: Astra pushes harder to complete authorized work — fewer permission loops, better preservation of the user's objective, and explicit limits on unnecessary process.
- Other rewrites: substantial changes to personality, writing style, skill usage, and execution guidance.
The author notes these are base instructions only; additional instructions for collaboration modes, permissions, approvals, auto-review, and tool policies also shape final behavior.
More from coding & agent
- yacine tests frontier models on CAD/PCB design: big gains, still sneaks in unneeded fixes — yacineMTB · 2026-09-05
- Open-source pipeline classifies phone screenshots, then analyzes and searches the web — TheMoonMidas · 2026-09-05
- YouTube Tutorial Shares a Free MiniMax-H3 Workflow — solomars3 · 2026-09-05
- SemiAnalysis and vLLM launch AgentX, a benchmark for real multi-turn agentic traffic — AccBalanced · 2026-09-05
- Zilliz CTO on Notion's Vector Infra: 90% Cost Cuts and the Platform Hiding Under One AI Feature — J_Luan_ · 2026-09-05
- A Commerce Agent Needs Spending Boundaries at the Card Layer, Not the Prompt — NoFunnyMan · 2026-09-05