OpenAI previews GPT-5.6 Sol, its highest-capability model to date, alongside Terra and Luna tiers. The limited preview began June 27, 2026 OpenAI official GPT-5.6 preview, with stricter safety guardrails and new reasoning modes aimed at agentic coding, biology, and cybersecurity workloads.
GPT-5.6 Sol: New Reasoning Modes for Agentic Workflows
The flagship Sol model introduces two new reasoning configurations not available on prior GPT-5-series models. The max effort mode allocates additional compute time for deep, step-by-step reasoning on high-stakes tasks requiring extended planning and tool use. The ultra mode coordinates specialized subagents to accelerate complex, multi-step workflows. Both modes are designed to boost performance on tasks requiring extended planning and tool use OpenAI official GPT-5.6 preview.
Benchmark Gains Across Coding, Biology, and Cybersecurity
In standardized benchmark testing, GPT-5.6 Sol achieves the highest recorded score to date on Terminal-Bench 2.1. This coding benchmark evaluates end-to-end command-line workflow performance across planning, iterative debugging, and tool coordination tasks OpenAI official GPT-5.6 preview.
For biology use cases, the model outperforms GPT-5.5 on GeneBench v1, a genomics and quantitative biology evaluation. It also consumes fewer input tokens per query than its predecessor when running GeneBench v1 tasks OpenAI official GPT-5.6 preview.
On cybersecurity benchmarks, GPT-5.6 Sol matches the performance of Mythos Preview on ExploitBench. It uses roughly one-third of the output tokens of the Mythos Preview model to hit that matching score OpenAI official GPT-5.6 preview. It also posts improved results on ExploitGym, with performance scaling as reasoning effort increases OpenAI official GPT-5.6 preview.
Stricter Safety Guardrails Tailored to Model Tiers
GPT-5.6 Sol launches with OpenAI’s most robust safety stack to date, with safeguards tailored to each model’s capability tier OpenAI official GPT-5.6 preview. The limited initial preview is restricted to a small group of vetted trusted partners, per the company’s published Preparedness Framework.
In tests run on Chromium and Firefox browsers, the model identified software bugs and exploitation primitives. It did not autonomously produce a functional full-chain exploit under the tested conditions OpenAI official GPT-5.6 preview. It also does not cross the Cyber Critical threshold outlined in the framework. OpenAI noted the model is better suited to helping users find and patch vulnerabilities than carrying out end-to-end attacks OpenAI official GPT-5.6 preview.
Tiered Pricing and Use Case Alignment
The accompanying Terra model delivers performance competitive with GPT-5.5 at a lower cost point than prior OpenAI flagship models. It is built for routine enterprise and developer workloads that do not require the full capability of the Sol variant OpenAI official GPT-5.6 preview.
Luna offers strong baseline capability at OpenAI’s lowest public price point to date. It targets high-volume, low-latency use cases that require fast response times OpenAI official GPT-5.6 preview. The tiered structure lets teams match model capability directly to task requirements without paying for unnecessary capability overhead OpenAI official GPT-5.6 preview.
