Summary & Key Takeaways ​

• OpenAI's new GPT-5.6 models are named Luna, Terra, and Sol, varying in size and cost.
• The models show significant improvements in long-running agentic performance, outperforming Claude Fable 5.
• Pricing is detailed per 1M input/output tokens for each model.
• OpenAI audited SWE-Bench Pro, finding ~30% of tasks broken, which may explain benchmark discrepancies.
• Simon Willison shares early access impressions, noting competence but not necessarily superiority for complex coding tasks.
• The release includes new API features for developers to explore.

Our Commentary ​

Simon Willison's breakdown of GPT-5.6 is exactly what we needed. Official announcements are one thing, but an independent, critical look at pricing, benchmarks, and real-world performance is invaluable. The agentic performance claims are huge, and the context around SWE-Bench Pro is a fascinating peek behind the curtain of AI evaluation. I'm still processing the implications of these models.