GPT-5.6 Sol Revealed: Three-Model Lineup, Terminal-Bench 2.1 SOTA, and Ultra Mode With Parallel Subagents
OpenAI has unveiled the GPT-5.6 series in limited preview: Sol, the flagship; Terra, a balanced model positioned against GPT-5.5-class performance at 2x lower cost; and Luna, a faster, lower-cost option designed for high-throughput use. The preview is restricted to a small set of government-cleared partners while OpenAI works toward general availability in the coming weeks.
The model family introduces two new capability modes. Max reasoning effort gives Sol more thinking time on complex tasks. Ultra mode goes further, spinning up parallel subagents to decompose and accelerate work that would otherwise be bottlenecked by a single-thread reasoning chain.
What Sol Actually Does
On Terminal-Bench 2.1, GPT-5.6 Sol posts a new state of the art. The previous published leader was GPT-5.5 at 84.7%. OpenAI has not released Sol’s exact score ahead of general availability, but the claim is a clear top position on a benchmark that tests command-line planning, iteration, and tool coordination across multi-step workflows — the class of task where coding agents live or die.
On GeneBench v1, a long-horizon genomics and quantitative-biology benchmark, Sol outperforms GPT-5.5 while consuming fewer output tokens. The efficiency story matters here: as biology agents move from research use cases toward production drug-discovery pipelines, cost per workflow is a real constraint.
The cybersecurity benchmark result is the most operationally significant. On ExploitBench, GPT-5.6 Sol is described as competitive with Claude Mythos Preview at roughly one-third of the output tokens. Mythos is the highest-capability published model on offensive security tasks. Closing that gap at substantially lower inference cost changes the economics of AI-assisted vulnerability research.
OpenAI is explicit that Sol does not cross the Cyber Critical threshold under its Preparedness Framework. In evaluations involving Chromium and Firefox, Sol identified bugs and exploitation primitives but did not autonomously produce a functional full-chain exploit under the tested conditions.
The Three-Model Structure
The GPT-5.6 lineup maps onto a clear tier strategy:
| Model | Role | Cost vs. GPT-5.5 |
|---|---|---|
| Sol | Flagship, max capability | Not yet disclosed |
| Terra | Balanced, everyday work | 2x cheaper |
| Luna | Fast, lowest cost | Not yet disclosed |
Terra is the commercial anchor. Matching GPT-5.5 performance at half the price, it targets the enterprise API workload that currently runs on GPT-5.5 standard: document processing, agentic workflows, coding assistance, and query answering at scale. Luna is positioned below GPT-4.1 mini in the speed tier, aimed at latency-sensitive, high-volume inference.
Ultra Mode
Ultra mode is the architectural step that matters for agentic workflows. Rather than extending a single reasoning chain further, ultra mode runs multiple subagents in parallel, aggregates their outputs, and synthesises a response. The claimed benefit is faster execution on complex, branching tasks that a sequential agent handles poorly.
The parallel subagent pattern is not new in AI infrastructure — OpenAI’s own Codex system, Cursor’s Composer, and Anthropic’s Managed Agents all implement variants — but embedding it natively in the model interface rather than requiring custom orchestration is a real simplification for developers.
Safety Stack
GPT-5.6 deploys a layered safeguard architecture: model-level refusals for prohibited cyber assistance, real-time misuse classifiers that can pause generation mid-output for a larger reasoning model to review, and account-level review triggered by flagged activity patterns across conversations. OpenAI describes spending multiple weeks pressure-testing the stack against real-world adversarial attacks.
The framing on the government preview arrangement is notable. OpenAI states it does not believe “this kind of government access process should become the long-term default” and frames the limited rollout as a short-term coordination step while an executive order cyber framework is negotiated. The subtext is that OpenAI is complying under commercial and political pressure while reserving the right to push back on the arrangement becoming permanent.
Context
The government gating mechanism for GPT-5.6 was announced Wednesday, before the model itself was revealed. Sol’s capabilities — particularly the ExploitBench result against Mythos and the Terminal-Bench SOTA claim — explain why the government moved first. A model that approaches Mythos on offensive cyber benchmarks at one-third the token cost is a different class of infrastructure risk than a model that doesn’t.
The broader competitive picture: Claude Fable 5 currently leads the Intelligence Index at 64.9 with an 80.3% SWE-Bench Pro score. GPT-5.6 Sol has not published SWE-Bench results yet. Terminal-Bench 2.1 SOTA and a GeneBench lead are its announced benchmarks for now. Full evaluation results are expected at general availability, which OpenAI targets in the coming weeks.