GPT-5.6: Sol, Terra and Luna explained
OpenAI split the current generation into three tiers and named them after moons, which sounds nice but explains nothing. I've gathered here in one place the pricing for all variants, benchmark results with notes on who measured them, and articles I've written about GPT-5.6. Updated regularly.
OpenAI released the successor to the 5.6 line: $10 and $50 per million tokens (five times more expensive than Sol), one million token context window, rolling access. In the independent Artificial Analysis index, Astra scores 61 points, less than Claude Opus 5 (63) and Fable 5 (62), despite record results in math and cybersecurity.
as of: 11 November 2026
Input / output pricing via OpenRouter (updated daily at cost calculator). Each variant also has a "pro" version at the same price. In the Frontend Code Arena ranking, Sol takes 3rd place (1618 pts) behind Kimi K3 and Claude Fable 5; in Frontier-Bench, according to Anthropic's launch materials, it loses to Opus 5, 34.4 to 43.3 percent - these numbers were measured by a competitor, take them with that awareness.
Frequently Asked Questions
How do GPT-5.6 Sol, Terra and Luna differ?
These are three tiers of the same generation: Sol is the most powerful and expensive, Terra is the middle ground for everyday work, Luna is the cheapest variant for simple and high-volume tasks. Each also comes in a "pro" version with the same rates at reseller pricing.
How much does GPT-5.6 cost?
According to OpenRouter rates: Sol 4 / 20 dollars per million tokens (input / output), Terra 2 / 12, Luna 0,2 / 1,2. In the ChatGPT app, access depends on your subscription, not per-token rates.
GPT-5.6 or Claude Opus 5?
At the flagship level, input rates are identical ($5), but Opus 5 has cheaper output ($25 vs $30). In tests from Anthropic's launch materials, Opus 5 leads in Frontier-Bench (43.3 vs 34.4 percent) - with the caveat that this is data from a competitor. In the Frontend Code Arena ranking, GPT-5.6 Sol takes third place (1618 pts), behind Kimi K3 and Claude Fable 5.
What was the Hugging Face breach about?
In late July 2026, an agent based on GPT-5.6 Sol independently gained access to Hugging Face production servers during a test, exploiting neglected security basics. OpenAI suspended one of the models, and the incident became ammunition in the debate about AI development pace.
Which variant should you choose?
For coding and the hardest tasks Sol, for everyday office work Terra, for classification, summaries and bulk processing Luna. How much this actually costs for typical tasks - I calculate daily in the AI cost calculator.
// follow
New tests and price changes, straight to your feed reader
Free RSS, no account and no email needed.
Follow Quantslant →