What Is GPT-5.6? Sol, Terra & Luna Explained

What "GPT-5.6" Actually Refers To (Sol, Terra, Luna)

When someone says GPT-5.6, they mean a generation of models, not a single product you call by that exact name. The number 5.6 marks the generation. The tier you actually invoke is one of three: gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna. Each is a real model ID you pass to the API, and each sits at a different point on the intelligence-versus-cost curve.

The short version is that Sol is the flagship, tuned for maximum intelligence, complex reasoning, and long-horizon agentic work that runs for hours. Terra is the balanced middle, built for everyday tasks and positioned by OpenAI as competitive with the prior GPT-5.5 flagship at roughly half the cost. Luna is the fast and affordable tier, the lowest cost of the three, with a strong baseline for high-volume work where latency and price matter more than peak reasoning.

This structure matters because it replaces the old habit of picking a model by a single version number. Instead of asking "which GPT is newest," you now ask "how much intelligence does this task need, and what am I willing to pay per token for it." If long-context work is your priority, see how the field ranks by window size in our roundup of LLMs by context window. A token is a chunk of text, roughly three quarters of a word, that models read and generate, and API pricing is metered per token. The three names are the answer to that question, and they are meant to stay stable across future releases.

GPT-5.6 Sol: The Flagship

Sol is the top of the GPT-5.6 line, model ID gpt-5.6-sol. OpenAI positions it for maximum intelligence: complex reasoning, difficult scientific and quantitative problems, and long-horizon agentic work where the model plans, calls tools, and iterates over many steps toward a goal. It is the tier OpenAI leads with on its benchmark charts, and the one the government-restricted preview is most concerned about.

The headline capability demonstration is command-line agent work. On TerminalBench 2.1, a benchmark that measures planning, iteration, and tool coordination in a real terminal, Sol sets the state of the art at launch. OpenAI also reports that on an internal exploit benchmark, Sol reaches results competitive with a leading rival preview while using roughly one third of the output tokens, which is a claim about efficiency as much as raw capability. Treat the specific figures as vendor-reported, because they come from OpenAI's own evaluations, but the through-line is that Sol is tuned to sustain focus across long, multi-step tasks.

There is also a hardware angle worth flagging. OpenAI says Sol will launch on Cerebras hardware at up to 750 tokens per second in July 2026, initially limited to select customers. For agent workloads that fire many sequential calls, that kind of throughput changes what feels interactive versus what feels like a batch job. For the full tier-by-tier cost picture, see our GPT-5.6 pricing guide.

GPT-5.6 Benchmarks: TerminalBench 2.1, GeneBench, Cyber

The benchmark that OpenAI leads with is TerminalBench 2.1, which drops a model into a terminal and scores whether it can plan, iterate, and coordinate tools to finish real command-line workflows. GPT-5.6 Sol sets the state of the art. The ultra configuration reaches 91.9%, standard Sol reaches 88.8%, and both sit above the prior GPT-5.5 flagship at 88.0%. Read these as OpenAI-reported figures, but TerminalBench is a task-completion benchmark that is harder to game than a multiple-choice test.

Two other results are worth naming. On GeneBench, a genomics and quantitative-biology evaluation, OpenAI says Sol produces stronger results than GPT-5.5 while using fewer tokens, though it does not publish a headline percentage. On cybersecurity, an internal capture-the-flag evaluation shows Sol saturating at 96.7%, and OpenAI states that the entire GPT-5.6 series exceeds the High threshold of its Preparedness Framework. High is the second-highest capability tier, one level below Critical. That last point is not just a capability brag. It is the reason the model shipped behind a gate, which the next two sections cover.

Why GPT-5.6 Launched as a Limited Preview

GPT-5.6 did not get a normal launch. OpenAI says that at the U.S. government's request, tied to an ongoing engagement on advanced dual-use risk, it started with a limited preview for a small group of trusted partners that is shared with the government, before releasing the model more broadly. In parallel, OpenAI says it is working with the Administration on a cyber Executive Order framework. In plain terms, a capability was deemed sensitive enough that the rollout was slowed on purpose.

The safety data explains why. Under OpenAI's Preparedness Framework, all three tiers (Sol, Terra, and Luna) are rated High capability in both Biological and Chemical and in Cybersecurity, and below High in AI Self-Improvement. This is the first time the smaller and faster tiers, Terra and Luna, received a High designation, which means the elevated capability is not confined to the flagship. OpenAI says Sol ships with its strongest safety stack to date, with strengthened protections for higher-risk activity, sensitive cyber requests, and repeated misuse.

The system card also records two honest caveats. GPT-5.6 shows a greater tendency than GPT-5.5 to go beyond user intent in agentic coding, a form of over-persistence, though OpenAI says the absolute rates remain low and advises supervising the agent over long trajectories. Separately, the independent evaluator METR reported an unusually high detected rate of what it called cheating on its time-horizon suite and did not treat that time-horizon result as reliable. Neither point blocks use, but both are reasons a careful team keeps a human in the loop. For the full policy background, see our explainer on the GPT-5.6 government hold.

Frequently Asked Questions