Backend · Programmer · Artificial Intelligence

OpenAI Lifts GPT-5.6 Preview Restrictions Ahead of Thursday's Full Launch

By cxuanAI · Jul 9, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The phased rollout signals that GPT-5.6's cybersecurity and autonomous-agent capabilities are strong enough to require controlled release. Developers get a concrete price-performance ladder: Sol for hard reasoning tasks, Terra as a cheaper GPT-5.5 equivalent, and Luna for high-volume, low-cost workloads.

Summary

The three-model GPT-5.6 family — Sol (peak intelligence), Terra (balanced cost-effectiveness), and Luna (cheap and fast throughput) — is moving from limited organizational preview to wider availability. Sol already topped TerminalBench 2.1 for command-line coding workflows, beating Mythos 5, while the cheaper Terra edged out Fable 5. Two new capability modes, `max reasoning effort` and `ultra mode`, let Sol spend more time on deep reasoning or spawn sub-agents for complex tasks. On the security side, Sol approached Mythos Preview's exploit performance using roughly one-third the tokens, though OpenAI's System Card confirms it did not produce autonomous attack chains under test conditions and remains below the Cyber Critical threshold. Full evaluations will drop alongside the broader release.

Takeaways

— GPT-5.6 Sol, Terra, and Luna expand to wider preview access globally, with a full public launch on Thursday.

— Sol is the top-intelligence model, Terra matches GPT-5.5 performance at half the cost, and Luna prioritizes speed and low price over raw capability.

— Sol beat Mythos 5 on TerminalBench 2.1, a benchmark for command-line planning, iteration, and tool use.

— Terra, the mid-tier model, outperformed Fable 5 on the same benchmark.

— Two new modes — `max reasoning effort` for deeper single-model reasoning and `ultra mode` for spawning sub-agents — extend Sol beyond single-agent limits.

— On ExploitBench, Sol reached near-Mythos Preview exploit performance while consuming roughly one-third the tokens.

— OpenAI's System Card states Sol did not produce autonomous attack chains in Chromium and Firefox tests and has not crossed the Cyber Critical threshold.

— Full benchmark evaluations will be published alongside the broader release.

Conclusions

OpenAI is segmenting its model line by a clear cost-capability trade-off rather than a single flagship, mirroring how cloud infrastructure tiers compute. Sol, Terra, and Luna map to premium compute, standard instances, and spot instances.

The `ultra mode` sub-agent architecture suggests OpenAI is baking multi-agent orchestration directly into the model interface rather than leaving it to external frameworks like AutoGen or CrewAI.

Sol's security performance — near state-of-the-art exploit capability at one-third the token cost — implies efficiency gains that lower the compute barrier for both red-team and blue-team workflows.

Phased release tied to a safety threshold (Cyber Critical) is becoming the standard playbook for frontier models, and it means capability announcements will routinely precede actual availability.

Concepts & terms

TerminalBench 2.1

A benchmark that evaluates an AI model's ability to handle command-line workflows requiring multi-step planning, iterative problem-solving, and coordination of external tools.

GeneBench v1

A benchmark for assessing model performance on long-cycle genomics and quantitative biological analysis tasks.

ExploitBench / ExploitGym

Benchmarks that measure a model's capability in vulnerability research and exploit development, used to evaluate cybersecurity reasoning over extended tasks.

Cyber Critical threshold

An internal OpenAI safety classification indicating a model can autonomously execute complete attack chains; GPT-5.6 Sol has not crossed this threshold under tested conditions.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗