OpenAI made GPT-5.6 generally available on July 9 after a limited preview that began on June 26 for a small group of trusted partners.
The company said the rollout will happen in phases over about 24 hours. OpenAI also said it coordinated the broader release with the U.S. government.
The family ships in three tiers rather than one: Sol as the flagship, Terra as a balanced everyday option, and Luna as the fastest, cheapest tier. All three are now live across ChatGPT, ChatGPT Work, Codex, and the API, rolling out globally over roughly 24 hours.
What OpenAI is Claiming, and Where the Numbers Back it Up
OpenAI’s central pitch is efficiency, not just raw capability. On Agents’ Last Exam, a test of long-running professional workflows, Sol scored 53.6, beating Anthropic’s Claude Fable 5 by 13.1 points.
On the Artificial Analysis Coding Agent Index, Sol set a new high of 80, edging out Fable 5’s 77.2 while using less than half the output tokens and costing roughly a third less. Cybersecurity gains were sharper still, with Sol nearly doubling GPT-5.5’s exploit-generation success rate.
Pricing lines up with that efficiency framing: Sol runs $5 input and $30 output per million tokens, Terra $2.50 and $15, Luna $1 and $6, all cheaper on paper than Claude’s current top-tier pricing.
The Benchmarks OpenAI’s Own Tables Don’t Win
The published results are more mixed than the launch framing suggests. On SWE-Bench Pro, a widely watched real-world coding benchmark, Sol scored 64.6% against Fable 5’s 80% and Claude Mythos 5’s 80.3%, a roughly 15-point gap in Anthropic’s favor.
Fable 5 also edges out Sol on the Artificial Analysis Intelligence Index (59.9 to 58.9), GDPval-AA’s Elo ranking, and HealthBench Professional.
Notably, OpenAI published a separate post the day before this launch arguing that roughly 30% of SWE-Bench Pro’s test tasks are flawed, a claim independent observers noted arrived right as OpenAI’s own model was trailing on that exact benchmark.
Early hands-on impressions have echoed the split: developer and independent researcher Simon Willison, who had preview access, said Sol felt competent but not clearly better than Fable 5 on complex coding work he’d already been running.
The launch highlights a broader trend across leading AI labs. No single model now dominates every benchmark or use case. Instead, developers are choosing models based on workload, cost, latency, and reliability. That shift matters more than simply picking whichever flagship model scores highest on a benchmark.
Source: Official OpenAI announcement, "GPT-5.6: Frontier Intelligence That Scales With Your Ambition"
