GPT-5.6 Sol, Terra & Luna Explained + How to Access

What Launched

On June 26, 2026, OpenAI previewed GPT-5.6 as a family of three models: Sol, Terra, and Luna. The short version: Sol is the frontier tier for the hardest work, Terra is the balanced everyday tier, and Luna is the fast, cheap tier for high-volume tasks. Access is gated behind a government-vetted preview, so most developers cannot call it yet.

The naming is the first thing worth understanding. OpenAI split the generation number (5.6) from durable capability tiers (Sol, Terra, Luna) that can advance on their own cadence. In practice that means the tier name tells you what a model is tuned for, while the number tells you which generation of the underlying stack it is built on. It is a cleaner story than a single ever-rising version number, and it lets OpenAI ship a cheaper tier without implying it is a worse generation.

The Three Tiers

Here is the family in one table, using OpenAI's own positioning language and the launch pricing (per one million tokens). The price anchors are the useful part: Sol costs the same as GPT-5.5, while Terra delivers roughly GPT-5.5-level quality at about GPT-5.4 pricing.

A few honest caveats on the numbers above. Sol's 1M-token context is secondary-confirmed; Terra and Luna context sizes were not detailed at launch, and non-text modalities were not spelled out either. Also note the naming collision: the “Sol” voice in ChatGPT is a different thing entirely and has nothing to do with this model tier.

If you are price-shopping frontier coding models more broadly, it is worth seeing how a competing lineup splits its tiers. Our Claude Sonnet 5 vs Opus 4.8 comparison is a useful mental model for how a “balanced” tier trades off against a frontier one.

The Two New Modes: max and ultra

GPT-5.6 ships two new controls that matter a lot for how you use Sol in particular.

• max is a reasoning-effort setting. It gives Sol the most deliberation time on a single hard problem. Think of it as “spend more thinking budget on this one thing.”
• ultra is a subagent mode. Instead of one model grinding through a task, it spins up subagents to split and parallelize a complex project. OpenAI credits ultra with the top agentic scores, including the headline benchmark below.

The practical read: max is for depth on a single hard reasoning problem, and ultra is for breadth across a multi-part project you want handled in parallel. If you are already building subagent workflows, the model is only half the story; the tooling layer matters just as much. Our guide to the best MCP servers for Claude Code covers the kind of tool surface these agentic modes lean on.

GPT-5.6 also changes prompt caching in a way that is aimed squarely at agent loops. Developers set explicit cache breakpoints, backed by a guaranteed 30-minute minimum cache lifetime. Cache writes cost 1.25x the uncached input rate, but cache reads get a 90% discount. You pay a little more to establish the cache and far less on every reuse inside the window. For long-running agentic runs that re-read the same context repeatedly, that is a real cost lever, not a rounding error.

Benchmarks: What to Trust

The headline is Terminal-Bench 2.1, an agentic terminal benchmark where the model has to operate a shell to finish real tasks. Two Sol numbers are the only figures worth quoting as hard results:

A few other launch benchmarks are worth noting for shape rather than precision:

• Agent's Last Exam: Sol is the only model past the halfway mark, at 50.9% in what OpenAI calls code mode.
• ExploitBench: Sol roughly matches Claude Mythos Preview while using about one-third the output tokens.
• GeneBench (quant bio and genomics): Sol and Terra beat GPT-5.5 and GPT-5.4, with Sol getting there on fewer tokens.

One structural caveat: GPT-5.6 is not yet on the public SWE-bench Pro leaderboard, because access is gated. So the usual apples-to-apples coding comparison against purchasable models simply does not exist yet. Anyone showing you a clean SWE-bench Pro ranking that includes GPT-5.6 is extrapolating.

Safety Ratings

The safety story is more interesting than the usual boilerplate, and it directly explains the gating. All three tiers — not just Sol — are rated High for both cyber and bio/chem capability, and below High for AI self-improvement. That is unusual: normally the cheap tier is safe by virtue of being weak. Here, even Luna carries capability that may trigger new governance obligations.

For context on the effort behind those numbers: OpenAI reports spending roughly 700,000 A100e GPU-hours red-teaming jailbreaks, with monitoring recall of about 94.8% on bio and 81.6% on cyber. Sol crossed the internal CTF threshold most often at 96.7% but stayed below the “Cyber Critical” line. The takeaway is not that these models are dangerous in your hands; it is that the capability profile is high enough that the release process itself became a policy question.

Why Access Is Gated (and How to Actually Get It)

This is the part most coverage buries. GPT-5.6 opened as a limited preview to roughly 20 organizations only, chosen with details shared with the U.S. government. There is no public waitlist. You cannot buy, sign up for, or talk your way into early access unless you are one of the vetted partners.

The trigger was an Executive Order dated June 2, 2026(“Promoting Advanced AI Innovation and Security”) that ordered a 30-day federal benchmarking window, nominally due in early July. In OpenAI's words: at the government's request, it started with a limited preview for a small group of trusted partners. Notably, OpenAI did not sound thrilled about it and pushed back in its own post.

“We don't believe this kind of government access process should become the long-term default.”— OpenAI, in the GPT-5.6 preview announcement

There is a competitive backdrop worth stating plainly. The preview followed a U.S. export-control order aimed at Anthropic that pulled Claude Fable 5 and Mythos 5 from the market. If you are trying to make sense of that lineup, our Claude Opus 4.8 launch coverage is the reference point for the Anthropic side of this story.

On pricing, Sam Altman framed the value proposition around the tiers. As reported around the launch (the exact tweet ID is unverified, but the quote is corroborated by VentureBeat), Altman said Sol is a smart, efficient, significant step forward at the same price as GPT-5.5, and that Terra offers GPT-5.5-level performance at half the price.

Two more access details worth knowing: Codex was upgraded to GPT-5.6 for preview partners, and Sol plugs into OpenAI's Daybreakcyber-defense program. General availability is described as coming in the weeks after launch — so the honest answer to “when can I use it?” is “soon, but on OpenAI's timeline, not yours.”

Cerebras: Sol at 750 Tokens per Second

One infrastructure note that changes the usage math: OpenAI plans to launch Sol on Cerebras in July, reaching up to 750 tokens per second. For an agentic frontier model, throughput is not a vanity metric. When Sol is running in ultra mode with subagents, or grinding a long max-effort problem, the difference between a slow frontier model and a fast one is the difference between a workflow you can sit in front of and a batch job you kick off and walk away from. If the Cerebras numbers hold, Sol becomes a genuinely interactive frontier option, not just a powerful-but-sluggish one.

FAQ

Glossary

Sol / Terra / Luna: The three capability tiers of GPT-5.6. Sol is frontier, Terra is balanced, Luna is fast and cheap.

Generation vs tier: 5.6 is the generation number. Sol, Terra, and Luna are durable capability tiers that OpenAI can advance independently, so the naming decouples raw version from what a model is tuned to do.

max mode: A reasoning-effort setting that gives Sol maximum deliberation time on a single hard problem.

ultra mode: A subagent mode that spawns multiple subagents to split and parallelize a complex project. Credited with the top agentic scores.

Terminal-Bench 2.1: An agentic terminal benchmark: the model must operate a shell to complete real tasks. It is the headline number for this launch.

CTF crossing: The rate at which a model crosses an internal capture-the-flag threshold in cyber evaluations, used to gauge offensive-cyber capability.

Executive Order (June 2, 2026): Promoting Advanced AI Innovation and Security, which ordered a 30-day federal benchmarking window that triggered the gated preview.

Cerebras: A wafer-scale inference vendor. OpenAI plans to run Sol on Cerebras in July at up to 750 tokens per second.

Verdict

GPT-5.6 is a genuinely interesting release wrapped in a frustrating rollout. The tier split is smart: separating generation from capability lets OpenAI ship a $1/$6 Luna without pretending it is a lesser generation, and Terra at GPT-5.5-level quality for half the price is the tier most teams will actually want. The one hard benchmark we trust — Sol + ultra at 91.9% on Terminal-Bench 2.1 — is a real result, and the subagent-driven ultra mode is the part worth watching, because it is a bet on orchestration rather than raw single-shot intelligence.

The catch is that you probably cannot use it. A ~20-org preview with no public waitlist means this is, for now, an announcement more than a product. Treat the competitor benchmark tables with suspicion until the public leaderboards include GPT-5.6, watch for the Cerebras speed launch in July, and plan around a GA date that is measured in weeks, not days. When it opens up, Terra is the tier we would reach for first.

Sources

Figures here are drawn from OpenAI's first-party materials and one primary press report, with clearly-labeled caveats where sources conflict. Grouped by type:

Official (OpenAI)

• OpenAI: Previewing GPT-5.6 Sol (official launch post)
• OpenAI: GPT-5.6 preview system card

Reported

• VentureBeat: OpenAI unveils GPT-5.6 Sol, Terra, and Luna (pricing, benchmarks, Cerebras)

Social

• @OpenAI launch thread on X (June 26, 2026)