GPT-5.6 (Sol, Terra, Luna): Release, Pricing & What's New (2026)

GPT-5.6 is OpenAI's newest model family, and it ships as three models rather than one: Sol (the new flagship), Terra (a capable lower-cost tier) and Luna (the fastest and most cost-efficient tier). It is genuinely brand-new — OpenAI published the preview system card on June 26, 2026, and independent coverage reports general availability rolling out from July 9, 2026 after a U.S. Department of Commerce review. This guide sticks to what OpenAI has officially confirmed: the three tiers, each model's context window and published API pricing, the safety designations from the system card, and how to access it. Where a number is not yet officially published, we say so instead of guessing. To be clear up front on the honest part: we do not serve GPT-5.6 on the DataLLM Lab gateway today — new OpenAI models arrive on OpenAI-compatible gateways as they roll out, and this page is a factual primer, not an availability claim.

What GPT-5.6 is

GPT-5.6 is OpenAI's newest model generation, and it is not a single model — it is a family of three: Sol, Terra and Luna. Per OpenAI's preview system card, Sol is the new flagship, Terra is a capable lower-cost option, and Luna is the fastest and most cost-efficient model. The naming is deliberate: the number 5.6 identifies the generation, while Sol, Terra and Luna identify durable capability tiers that OpenAI says can each advance on their own cadence — so a future point release could improve one tier without renaming the whole family.

In plain terms: pick the tier by how hard your task is and how much latency and cost you can absorb, and pick the generation (5.6) to get the current-generation version of that tier. If you are weighing this against the previous generation, see our GPT-5.6 vs GPT-5.5 comparison; if you are choosing between OpenAI and Anthropic more broadly, Claude vs ChatGPT covers that trade.

Is GPT-5.6 available yet?

Yes — but it is days old, so treat every detail as fresh and fast-moving. The confirmed timeline, sourced to primary and independent reporting:

The preview date is stated directly on OpenAI's GPT-5.6 preview system card ("Published June 26, 2026"), which also confirms the initial limited preview with trusted partners and broad availability "in the coming weeks." The general-availability date of July 9, 2026 comes from news coverage (Engadget), not an OpenAI primary announcement page, so we mark it at lower confidence and date it explicitly. Because this is a just-launched model, we have not first-party tested GPT-5.6; nothing below is a benchmark we ran.

Sol vs Terra vs Luna, compared

All three share the same 1,050,000-token context window and 128,000-token max output — they differ on capability tier and price. The table below is our synthesis of OpenAI's official API model docs; every number is sourced, and the "Best for" column is guidance, not a spec.

Two things worth noting. First, context and output are identical across the three — you do not trade context to move down a tier, only capability and price. Second, every tier prices cached input at exactly 10% of its uncached input rate (a 90% cache discount), so prompt caching is a meaningful lever on all three. OpenAI positions Sol as the flagship, Terra as the capable lower-cost option and Luna as the fastest and cheapest; anything more specific about relative benchmark scores is not yet officially published in a form we can cite, so we are not putting a number on it.

GPT-5.6 pricing

Pricing is published in OpenAI's developer docs, and it steps down cleanly by tier. Per 1M tokens, from the official Sol, Terra and Luna model pages:

# GPT-5.6 API pricing — USD per 1M tokens (OpenAI docs, July 2026)
Sol    input $5.00   cached $0.50   output $30.00
Terra  input $2.50   cached $0.25   output $15.00
Luna   input $1.00   cached $0.10   output $6.00

# Long-context surcharge (all three models):
# requests over 272K input tokens are billed at
#   2x the input rate and 1.5x the output rate

Two practical caveats. The 272K long-context surcharge matters if you routinely stuff the full million-token window — a Sol request above that threshold is billed at $10 input / $45 output per 1M, not $5 / $30. And the 90% cached-input discount is the single biggest lever on your bill for any workload with a stable system prompt or repeated context. For where GPT-5.6 lands against other providers on raw token cost, see our cheapest LLM API guide; for the broader "which API should I standardize on" question, best LLM API 2026.

Which GPT-5.6 model to use

Start at Terra, drop to Luna when cost or latency bites, and reach for Sol only when the task is genuinely hard. A simple decision path:

• Default to Terra. It is the balanced tier — most production traffic (chat, extraction, summarization, routine tool use) belongs here at half of Sol's price.
• Drop to Luna when you are running high volume, need the lowest latency, or have a hard cost ceiling — classification, routing, autocomplete, cheap first-pass drafting. It is the fastest and most cost-efficient tier.
• Reach for Sol when the task is the flagship's job: the hardest multi-step reasoning, long-horizon agentic workflows, or security-sensitive work where you want the top tier. Pay $30/1M output only where it changes the outcome.

A common production pattern is to route by difficulty: cheap tier for the bulk, escalate to Sol on the hard minority. That routing logic is exactly what an OpenAI-compatible API layer makes portable across models.

Safety designations

Per the preview system card, all three GPT-5.6 models carry "High" capability designations in two domains — but none reach the framework's top "Critical" level. Specifically, the system card reports that Sol, Terra and Luna all received a High designation for biological/chemical and for cybersecurity, while remaining below High for AI self-improvement. A "High" designation triggers additional deployment safeguards under OpenAI's risk framework; it is not the same as the highest "Critical" classification. This is also why the rollout was tied to a U.S. Department of Commerce review before general availability. If you are building in a regulated or security-sensitive context, read the system card directly rather than relying on this summary.

How to access GPT-5.6

Access is through OpenAI directly — the API model IDs and consumer surfaces roll out as availability expands. As of July 2026, the practical routes are:

• OpenAI API — the gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna model pages are live in OpenAI's developer docs; call them via the standard Chat Completions / Responses endpoints once your account has access.
• ChatGPT — consumer access ("ChatGPT 5.6") typically follows API availability; check your plan's model picker, as the rollout is staged.

A minimal call looks like any other OpenAI-compatible request — only the model ID changes:

from openai import OpenAI

client = OpenAI()  # OpenAI's own endpoint

resp = client.chat.completions.create(
    model="gpt-5.6-terra",          # or gpt-5.6-sol / gpt-5.6-luna
    messages=[{"role": "user", "content": "Summarize this contract."}],
)
print(resp.choices[0].message.content)

Because it uses the OpenAI-compatible request shape, the same code moves to any compatible endpoint by changing the base_url and key — which is the point of the next section.

GPT-5.6 and OpenAI-compatible gateways

To be straight with you: DataLLM Lab does not serve GPT-5.6 today. We are an OpenAI-compatible gateway, and new OpenAI models arrive on compatible gateways as they roll out — but a model being days old is exactly when we would rather tell you "not yet" than imply coverage we do not have. For GPT-5.6 specifically, call OpenAI directly once you have access.

What the gateway does give you is everything around that decision: one key across 300+ models you can route today (Claude, Grok, GLM, open-weight options), the same OpenAI-compatible request shape shown above, and automatic failover so no single provider's outage is your app's outage. When GPT-5.6 does land on OpenAI-compatible gateways, the only change on your side is the model ID — the routing, caching and failover you already wired stay put. Browse what is callable now on the pricing and models page.

FAQ