On June 27, 2026, OpenAI officially launched the GPT-5.6 model series, introducing a new astronomical naming convention—Sol (Sun)—to represent flagship, balanced, and cost-efficient tiers respectively. The new models deliver breakthrough performance in coding, biology, and cybersecurity, with GPT-5.6 Sol achieving a record 91.9% score on the Terminal-Bench 2.1 programming benchmark, surpassing Anthropic’s latest closed model, Claude Mythos 5 (88.0%). Currently, access is restricted to U.S. government-approved “trusted partners” only.
Astronomical Naming, Tiered Capabilities: A Sustainable Model Portfolio
For the first time, OpenAI has moved away from purely numerical versioning, adopting Latin celestial terms to establish a long-term product architecture:
Sol (Sun): Flagship model for high-complexity scientific, engineering, and security tasks;
Terra (Earth): General-purpose workhorse balancing performance and cost;
Luna (Moon): Lightweight, high-speed model optimized for low-latency inference.
“Numbers denote generations; Sol/Terra/Luna denote persistent capability tiers that can evolve independently,” OpenAI explained. This marks a strategic shift from “version upgrades” to “product-line management,” offering enterprises more granular AI service options.
Coding Supremacy: Ultra Mode Hits 91.9% on Terminal-Bench
GPT-5.6 Sol introduces a new Ultra reasoning mode, leveraging sub-agents to parallelize complex task chains. On the authoritative Terminal-Bench 2.1:
Standard mode scores 88.8%, already ahead of Claude Mythos 5 (88.0%);
Ultra mode pushes further to 91.9%, setting a new industry record.
In GeneBench v1 (biology), Sol achieves stronger genomic analysis with fewer tokens;
in ExploitBench (cybersecurity), it matches Mythos Preview’s exploit generation quality using just one-third the output tokens.
Five-Layer Security Architecture for High-Risk Scenarios
To counter escalating misuse risks, OpenAI has deployed a multi-layered safety framework across the GPT-5.6 family:
Built-in refusal mechanisms
Real-time content classifiers during generation
Account-level risk profiling
Tiered access controls
High-risk queries escalated to larger models for review
Violations are intercepted before user display, ensuring compliant outputs.
Commercial Strategy: Caching Optimization and Cerebras Acceleration
Pricing remains tiered:
Sol: $5 input / $30 output per million tokens
Terra: $2.5 / $15
Luna: $1 / $6
Prompt caching has been enhanced to reduce costs for repeated queries. Notably, OpenAI plans to deploy GPT-5.6 Sol on Cerebras chips in July, achieving up to 750 tokens/second—initially available to select enterprise clients.
Editor’s Note: The GPT-5.6 launch signals not just technical advancement, but OpenAI’s maturing product strategy. By combining capability tiering, domain specialization, and robust safety, OpenAI is evolving from a general-purpose AI provider into a trusted enterprise intelligence infrastructure. In an era where AI competition has entered deep waters, the winner will be whoever best balances performance, cost, and trust.