GPT-5.6 family launches across three tiers, Luna hits 234 tok/s at $2.25/M

GPT-5.6 Sol nearly matches Claude Fable 5 at 56% of the price

GPT-5.6 Sol (OpenAI) scores 58.9 on the quality index at $11.25/M input tokens. Claude Fable 5 (Anthropic) scores 59.9 at $20.00/M. That's a 1.0-point quality gap for $8.75/M in savings. For pipelines running millions of tokens daily, Sol is the stronger marginal deal.

Sol also generates at 85 tok/s versus Fable 5's 65 tok/s — 31% faster inference latency means tighter iteration loops in agentic workflows.

<!-- fact:claude-fable-5|quality=59.9|price=20.00|speed=65 --> <!-- fact:gpt-5-6-sol|quality=58.9|price=11.25|speed=85 --> <!-- fact:gpt-5-6-terra|quality=55.0|price=5.63|speed=166 --> <!-- fact:gpt-5-6-luna|quality=51.2|price=2.25|speed=234 -->

Luna is the throughput play

GPT-5.6 Luna (OpenAI) outputs 234 tok/s at $2.25/M — the fastest and cheapest model in the current field. Quality sits at 51.2, which is below GPT-5.4 at 51.4, but Luna is 45% faster and half the price. For batch classification, summarization, or any job where quality above 50 is sufficient, Luna is the obvious pick.

Terra splits the middle

GPT-5.6 Terra (OpenAI) lands at 55.0 quality, $5.63/M, and 166 tok/s. It matches GPT-5.4 on price but beats it by 3.6 quality points. Terra replaces GPT-5.4 as the default mid-tier recommendation. The xhigh variant pushes quality to 51.6 at 130 tok/s — not worth it. Stick with the base Terra.

Grok 4.5 holds the value floor

Grok 4.5 (xAI) stays relevant at 53.8 quality for $3.00/M with 110 tok/s. It outperforms Luna on quality by 2.6 points at only $0.75/M more. For workloads that need a bit more reasoning than Luna provides but can't justify Terra's $5.63/M, Grok 4.5 fills the gap.

What to watch

• Whether Anthropic responds to Sol's price-quality pressure on Fable 5 before the gap normalizes.
• If Luna's 234 tok/s holds under production load or degrades with concurrent requests.
• Whether the GPT-5.6 Sol reasoning variants (high, medium) get adopted or ignored given the flat $11.25/M pricing across all tiers.

For model selection by workload, use the LLM Selector. To compare full specs, Explore the leaderboard.