OpenAI released GPT-5.6 to the public on July 9, splitting its flagship line into three tiers: Sol, Terra, and Luna. The launch ends a wait of roughly 12 days, during which the model sat behind a gate tied to a White House review of a voluntary AI safety framework, and it arrives already wired into GitHub Copilot. Sol, the flagship, is where OpenAI wants attention: the company calls it its most capable cybersecurity model yet and says it leads on coding, cybersecurity, and science.
The GitHub Copilot integration matters as much as the tier split. Microsoft’s coding assistant is one of the most widely deployed AI pair-programming tools in enterprise settings, and putting all three GPT-5.6 tiers there on day one gives OpenAI a distribution channel that a standalone API launch would not.
OpenAI says Sol shifts the performance-efficiency frontier for long-horizon security work, including vulnerability research and exploitation. The company reports that Sol leads its own benchmarks on Terminal-Bench, BrowseComp, OSWorld, and Agents Last Exam. On coding specifically, Sol scores 64.6 percent on SWE-Bench Pro. Anthropic’s Claude Fable 5 scores 80 percent on the same benchmark, a 15.4 point gap that OpenAI’s own launch materials do not address.
That gap complicates the framing. The Decoder, covering the release independently, described Sol as nearly matching Fable 5 on aggregated benchmarks at roughly one-third the cost, rather than surpassing it outright. That is the more accurate read: Sol looks strong across a broad benchmark suite and weak on the single benchmark developers weigh most heavily for production coding work.
Pricing splits the three tiers by what OpenAI calls cost of failure, its own framing for who should use each one:
- Sol: 5 dollars per million input tokens, 30 dollars output. For expensive, high-stakes work.
- Terra: 2.50 dollars input, 15 dollars output. For everyday production use.
- Luna: 1 dollar input, 6 dollars output. For high-volume pipelines such as classification and summarization.
All three tiers share a 1.05 million token context window and a 128,000 token maximum output. GPT-5.6 also introduces more predictable prompt caching: explicit cache breakpoints, a 30 minute minimum cache life, and cache writes billed at 1.25 times the uncached input rate, while cache reads keep the existing 90 percent discount.
Luna is the tier worth watching for cost-sensitive teams. OpenAI says it delivers roughly 85 percent of Sol’s quality at about one-fifth the price, a ratio aimed at high-volume, low-stakes workloads rather than frontier reasoning tasks.
On ARC-AGI-3, results stayed low across the industry: Sol scored 7.78 percent, well ahead of Anthropic’s Opus 4.8 at 1.5 percent. OpenAI says Sol became the first model to win a public ARC-AGI-3 game. A single win on a benchmark where every model still scores in single digits says more about the benchmark’s difficulty than about general capability. OpenAI also says GPT-5.6 improves multi-agent parallel processing and design judgment and ships with strengthened safety measures, without specifying which evaluations changed or by how much.
Teams choosing a coding model on cost alone should still benchmark against Fable 5 directly rather than trust the aggregate comparison. Sol’s one-third cost advantage does not close an 80 percent to 64.6 percent gap on SWE-Bench Pro, and that gap, not the pricing table, is the number that should decide production coding contracts over the next quarter.
Based on OpenAI’s July 9, 2026 announcement, “Introducing GPT-5.6: Sol, Terra, and Luna,” with independent benchmark framing from The Decoder.
