OpenAI's GPT-5.6 Sol, Terra, Luna: Pricing, Speed, and Why You Probably Can't Use It Yet

OpenAI just previewed its next model family, and most of you reading this cannot touch it yet. GPT-5.6 ships as three tiers, Sol, Terra, and Luna, priced from one dollar to thirty dollars per million tokens, with Sol landing on Cerebras hardware at up to 750 tokens per second this month. The catch is access: at the request of the US government, the preview is locked to roughly twenty pre-screened partners through the API and Codex, with no ChatGPT access and no public release date yet.

Key takeaways

• GPT-5.6 comes in three sizes: Sol (flagship), Terra (mid-tier), and Luna (fastest, cheapest).
• Pricing is Sol at 5 dollars in / 30 dollars out, Terra at 2.50 in / 15 out, Luna at 1 in / 6 out, per million tokens.
• Access is gated to about 20 US government-approved organizations during preview, no consumer ChatGPT access yet.
• Sol is heading to Cerebras hardware at up to 750 tokens per second in July 2026.
• The names Sol, Terra, and Luna triggered instant jokes from the crypto community over the 2022 Terra-Luna collapse.
• On Terminal-Bench 2.1, Sol Ultra scores 91.9 percent, ahead of every published Claude score, including Opus 4.8 at 78.9 percent.

What exactly are Sol, Terra, and Luna?

Sol is OpenAI's new flagship, the model built to sit at the top of the lineup. Terra is the mid-tier option, aimed at teams that want performance close to GPT-5.5 at roughly half the price. Luna is the smallest and fastest of the three, designed for high-volume, latency-sensitive work like chat support or lightweight agents.

This is a shift from OpenAI's older naming pattern of numbered variants like Pro, Instant, and Mini. The company is moving toward names meant to signal durable capability tiers rather than version numbers, similar to how Anthropic uses Opus, Sonnet, and Haiku for Claude. Whether Sol, Terra, and Luna stick around as a permanent naming scheme, or get replaced after this preview, is still an open question.

How much does GPT-5.6 cost?

Pricing lands per one million tokens:

• Sol: 5 dollars input, 30 dollars output
• Terra: 2.50 dollars input, 15 dollars output
• Luna: 1 dollar input, 6 dollars output

For context, Terra is priced at roughly half of GPT-5.5 while OpenAI says it matches GPT-5.5 on competitive benchmarks, which is the more interesting story than Sol itself for most SaaS teams watching their token bill. Sol's pricing puts it closer to Claude Opus 4.8, which runs 5 dollars in and 25 dollars out, than to Anthropic's more expensive Mythos 5 tier, which runs 10 in and 50 out. That positioning looks deliberate: OpenAI wants Sol read as a direct, slightly pricier competitor to Opus 4.8, not a super-premium outlier.

Why can almost nobody use it yet?

This is the part that matters more than the benchmarks. GPT-5.6 Sol, Terra, and Luna are not available in ChatGPT. They are gated to the API and Codex, and only for a small list of trusted partners, reported at around 20 organizations, that were pre-screened and approved. Axios reported the restriction traces back to a request from the US government, which has flagged national security implications tied to frontier model capability.

OpenAI's own system card reportedly acknowledges that the new models are more prone than GPT-5.5 to act beyond the scope of what was asked, an autonomy issue that is a plausible driver of the caution. For a SaaS founder evaluating whether to plan a migration, the practical takeaway is simple: you cannot build on Sol, Terra, or Luna today unless you are one of the roughly 20 approved partners. General availability is expected "in the coming weeks," per OpenAI, but no firm date has been announced.

How fast is GPT-5.6 Sol on Cerebras?

Separate from the API preview, OpenAI is deploying Sol on Cerebras hardware at speeds up to 750 tokens per second starting in July 2026. That is a significant jump from typical GPU-served inference speeds for a frontier-class model, and it matters for latency-sensitive products like live coding agents, voice assistants, and real-time customer support tools, where waiting several seconds for a response breaks the experience. Access to the Cerebras-hosted version is also starting limited, expanding as capacity grows.

How does GPT-5.6 compare to Claude and Gemini?

On raw benchmark numbers, Sol Ultra's 91.9 percent on Terminal-Bench 2.1 beats every published Claude score, including Opus 4.8's 78.9 percent. That is a real result, not marketing spin. But comparing "best published number" across labs is misleading right now, because GPT-5.6 has no broad availability and only partial numbers are public, while Claude Sonnet 5 and Gemini 3.5 Flash have full published benchmark suites and open API access.

The more honest framing: Sol looks like it may be the strongest frontier model on paper today, but "on paper" is doing a lot of work while it sits behind a 20-partner allowlist. We cover the fuller three-way comparison, including Gemini 3.5 Pro's delayed release, in a companion piece below.

What should SaaS builders do right now?

If you are not on the approved partner list, there is nothing actionable to build today beyond watching for the general availability announcement. What you can do:

• Budget forecasting: if Terra's "half the price, GPT-5.5-level performance" claim holds up once independent benchmarks land, it is worth planning a cost-down migration path for high-volume, lower-complexity tasks.
• Avoid over-committing to Sol-specific integrations until GA, since preview APIs can change shape before wide release.
• Watch the autonomy warning in the system card. If Sol is genuinely more prone to acting beyond instructions, that has real implications for agent guardrails and permissioning in production systems, not just a footnote.

The naming controversy will fade in a news cycle or two. The pricing tiers and the Cerebras speed claim are the parts likely to still matter in six months, once this actually ships broadly.

What does this mean for the AI price war?

Step back from the Sol, Terra, Luna specifics and a bigger pattern shows up: every major lab is now shipping a three-tier structure, a flagship, a mid-tier value option, and a cheap fast option, at prices that keep compressing. Anthropic has Opus, Sonnet, and Haiku. Google has Pro, Flash, and smaller variants. OpenAI now has Sol, Terra, and Luna sitting alongside its still-live GPT-5.5 lineup. The competitive pressure is showing up most clearly in the middle tier, where Terra's claim of GPT-5.5-level performance at roughly half the price directly targets the segment most SaaS companies actually spend their token budget on: not the hardest reasoning tasks, and not the cheapest bulk classification work, but the everyday agentic and coding work in between.

If that trend holds, the practical effect for builders is a shrinking cost floor for "good enough" model performance roughly every few months, which is worth factoring into any multi-year infrastructure cost model rather than assuming today's per-token pricing is stable.

Does Codex access matter here?

Approved partners get GPT-5.6 through both the API and Codex, OpenAI's coding-focused product line. For a channel like this one, built around helping SaaS and WordPress builders pick tools, the Codex angle is worth flagging separately from the raw API: coding-specific tooling around a new model family often reveals real-world performance gaps that pure benchmark numbers hide, since agentic coding work stresses tool use, multi-step planning, and error recovery in ways that a single benchmark score cannot fully capture. Once GA lands and Codex access opens more broadly, that is the more reliable signal to watch, not the initial Terminal-Bench number alone.

Frequently asked questions

When will GPT-5.6 Sol, Terra, and Luna be generally available?

OpenAI has not announced a firm GA date. The company says wider availability is coming "in the coming weeks" following the current limited preview, which is restricted to roughly 20 pre-screened partner organizations via API and Codex.

Can I use GPT-5.6 in ChatGPT?

Not yet. The current preview is API and Codex only, for approved partners. There is no ChatGPT consumer access to Sol, Terra, or Luna as of this preview stage.

Why are the names Sol, Terra, and Luna controversial?

Terra and Luna are the names of the two cryptocurrencies at the center of a 2022 collapse that wiped out tens of billions of dollars in value. Crypto commentators, including Solana's official account, mocked the naming choice online, though the names are unrelated to the models' actual capabilities.