How to Use GPT-5.6: Sol, Terra, and Luna Tiers Compared

OpenAI began a limited preview of GPT-5.6 on June 26, 2026, and the strangest thing about that launch was that almost nobody could use it. Access went to roughly twenty vetted partners through the API and Codex, in a phased rollout OpenAI says the U.S. government requested^[1].

That gate has since lifted. Knowing how to use GPT-5.6 is now a practical question rather than a waiting game, and the answer is mostly about picking one of three tiers: Sol the flagship, Terra the balanced workhorse, and Luna the budget tier. Each is a durable capability tier under one generation number, not a temporary variant.

This guide covers the three-tier design, the one benchmark OpenAI chose to lead with and what its four asterisks mean, the new max and ultra reasoning modes, the pricing that makes Terra the story, and the cybersecurity review that shaped the rollout.

TL;DR

• Three durable tiers. Sol (flagship), Terra (balanced, roughly 2x cheaper than GPT-5.5), Luna (fast and low cost)^[1].
• Sol tops Terminal-Bench 2.1 at 88.8% single-agent, or 91.9% in ultra mode. The single-agent lead over Claude Mythos 5 is 0.8 points, which is inside the noise band^[1].
• ultra is not one agent. It orchestrates subagents in parallel, so its score is not a like-for-like comparison against single-agent baselines^[1].
• Terra is the commercial story. GPT-5.5-class coding at half the price^[1].
• On reAPI all three run at 80% of OpenAI's published rates: Sol $4.00 / $24.00, Terra $2.00 / $12.00, Luna $0.80 / $4.80 per million tokens.
• The evaluation window was deliberately narrow. OpenAI published coding, biology, and cybersecurity only, holding back SWE-bench, GDPval, math, and hallucination numbers^[1].

What GPT-5.6 is

GPT-5.6 succeeds GPT-5.5, the fully retrained agentic model from April 2026. Where GPT-5.5 was one model with reasoning-effort levels, GPT-5.6 is a family of three tiers under one generation number. OpenAI's framing is explicit: the number identifies a generation, while Sol, Terra, and Luna identify durable capability tiers that can advance on their own cadence^[1].

That sentence is the most consequential design decision in the release. It decouples "how smart" from "how new." A future generation could ship a new Luna without touching Sol, and a Sol upgrade does not force everyone on Terra to re-test their pipelines.

Multiple third-party reports put Sol's context window at roughly 1.5 million tokens, about 43% larger than the GPT-5.5 generation, though OpenAI's preview post focused on safety and evaluations rather than the spec sheet^[1]. Treat that figure as reported rather than officially confirmed.

Terminal-Bench 2.1, and its four asterisks

OpenAI published only three benchmark areas in the preview, coding, biology, and cybersecurity, explicitly holding back the rest for general availability^[1]. The one it wants you to remember is Terminal-Bench 2.1, which tests command-line agentic workflows requiring planning, iteration, and tool coordination.

Read carefully, this table is more honest than the "new state of the art" headline suggests, in four ways^[1].

The 91.9% record uses ultra, which is not a single agent. It comes from Sol orchestrating subagents. Against single-agent baselines it is not apples to apples. The fair comparison is Sol's 88.8%.

Single-agent Sol beats Claude Mythos 5 by 0.8 points. Two runs of the same model can disagree by more than that. OpenAI holds a real lead at the top of the coding curve, but on this benchmark it is narrow, not a blowout.

Terra ties Claude Fable 5 and barely edges GPT-5.5. 84.3% matches Fable 5 exactly and sits less than a point above the model it replaces. Paired with the price cut, that is the actual commercial story: same coding ability for half the money, not a capability leap in the mid-tier.

Gemini 3.1 Pro Preview trails by about 21 points. That gap is real but benchmark-specific. Gemini leads elsewhere on long-context retrieval and multimodal work not shown here.

On biology, OpenAI reports Sol achieving stronger GeneBench v1 results than GPT-5.5 while using fewer tokens, without publishing raw numbers, so that is a directional claim rather than a verifiable score^[1].

max and ultra

The most important addition is not in any benchmark cell. GPT-5.6 introduces two reasoning settings, both exclusive to Sol^[1].

max is the familiar lever taken one notch further: a single agent given the most time to reason deeply on one hard problem.

ultra is the genuinely new primitive. It goes beyond a single agent by spawning subagents that split complex work into parallel pieces and recombine the results. This is OpenAI productizing multi-agent orchestration inside a single model endpoint, the pattern developers have been hand-rolling with frameworks for the past year.

The practical implication is real: for a multi-file refactor or an end-to-end audit, you no longer build the orchestration layer yourself. The catch is cost and latency, because subagents mean more total tokens and more wall-clock time. Reserve ultra for the high-stakes task where a few points justify the spend.

It also means benchmark comparisons need care. A score produced by an agent that quietly spawns subagents is a different kind of result from a single forward pass.

Pricing

Two things stand out in OpenAI's own numbers. Sol holds GPT-5.5's exact $5 / $30 pricing while adding capability and the new modes, which is a rare case of more for the same money. And Terra delivers GPT-5.5-class coding at half the price, which is the change most teams will actually feel^[1]. Layer Sol's token-efficiency gains on top, since it reaches comparable results with fewer output tokens, and the effective drop in cost per task is larger than the headline rates imply.

Every tier runs at 80% of the list rate on reAPI, so the relative economics between tiers are unchanged. Terra is still the default and Luna is still the volume play.

GPT-5.6 also overhauls prompt caching, adding explicit cache breakpoints and a 30-minute minimum cache life. Cache writes bill at 1.25x the uncached input rate while cache reads keep the 90% discount^[1]. For agents that re-send a large system prompt on every step, predictable caching can dominate the bill.

The review that shaped the rollout

This part is worth understanding even now that access has opened, because it is a precedent rather than a footnote.

OpenAI restricted the initial release to about twenty vetted partners at the U.S. government's request, after previewing the models' capabilities ahead of launch^[1]. The trigger was cybersecurity capability. Sol meaningfully advances vulnerability research, and in evaluations involving Chromium and Firefox it identified bugs and exploitation primitives, the building blocks of an exploit, without autonomously producing a functional full-chain exploit under the conditions tested. OpenAI states Sol does not cross the Cyber Critical threshold in its Preparedness Framework, and that distinction is the entire basis for releasing it^[1].

Notably, OpenAI did not endorse the arrangement as permanent, saying this kind of government access process should not become the long-term default because it keeps the best tools from the defenders who need them^[1].

The safety stack shipped alongside it is worth knowing because it previews how high-capability models will ship generally: trained-in refusals, real-time cyber and biology classifiers that can pause generation for review by a larger model, account-level review of flagged activity, and differentiated access to the most sensitive capabilities. OpenAI dedicated over 700,000 A100-equivalent GPU hours to automated red-teaming for universal jailbreaks^[1].

OpenAI is candid that these safeguards misfire: legitimate security researchers may hit refusals on dual-use requests.

Where the case is strong and where it is thin

Strong. Agentic command-line coding, where single-agent Sol tops the benchmark and ultra extends it. Token efficiency, which compounds for high-volume agents. Mid-tier value, where Terra's price cut is the most broadly useful change in the release. And native multi-agent orchestration, which removes a layer developers used to build by hand^[1].

Thin. The headline is a 0.8-point win on one benchmark. The evaluation window was deliberately small, with no SWE-bench, GDPval, math, or hallucination numbers published, so whether GPT-5.6 fixed the confident-hallucination problem that dogged GPT-5.5 is still unknown. And everything is vendor-reported under OpenAI's own harness^[1].

Until an expanded suite and independent scoring land, judgment on general capability should stay reserved.

How to use GPT-5.6: which tier to pick

Terra for most production workloads. It matches GPT-5.5-class capability at roughly half the cost, and the benchmark says its coding is a tie with Claude Fable 5.

Sol for long-horizon coding, security work, and anything justifying ultra. The subagent orchestration is worth real money on tasks where a few points change the outcome, and worth nothing on routine calls.

Luna for high-volume, latency-sensitive, routine traffic. At 82.5% on Terminal-Bench it is not far behind GPT-5.5's 83.4%, at a fraction of the cost.

Calling GPT-5.6 on reAPI

All three tiers sit behind one OpenAI-compatible /v1/chat/completions endpoint, so switching tiers is a model-string change rather than an integration. Note that the wire ids keep OpenAI's dots:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_REAPI_KEY",
    base_url="https://api.reapi.ai/v1",
)

resp = client.chat.completions.create(
    model="gpt-5.6-terra",
    messages=[{"role": "user", "content": "Refactor this module and add tests."}],
    max_tokens=16000,
    stream=True,
)

Swap gpt-5.6-terra for gpt-5.6-sol or gpt-5.6-luna to move tiers. Because the same key and client also reach Claude and Gemini, the tier decision and the vendor decision become the same kind of decision, which is the point when a benchmark table has your default model losing one row and winning another.

Current rates are on reapi.ai/models/gpt-5-6-sol, reapi.ai/models/gpt-5-6-terra, and reapi.ai/models/gpt-5-6-luna.

FAQ

What is the difference between max and ultra mode?

max is a single agent reasoning longer and deeper on one problem. ultra orchestrates multiple subagents that split complex work in parallel and recombine it. Both are exclusive to Sol, and ultra produced the 91.9% Terminal-Bench headline^[1].

Is GPT-5.6 Sol better than Claude for coding?

On Terminal-Bench 2.1, single-agent Sol at 88.8% edges Claude Mythos 5 at 88.0% by 0.8 points, a real but narrow lead. With ultra Sol reaches 91.9%, but that uses subagents and is not a like-for-like single-agent comparison^[1].

Which GPT-5.6 tier should I use?

Terra for most production workloads, since it matches GPT-5.5-class capability at about half the cost. Sol for long-horizon coding, security work, or tasks that justify ultra. Luna for high-volume routine calls^[1].

Why was the U.S. government involved in the release?

Sol meaningfully advances cybersecurity capability, including vulnerability research, without crossing OpenAI's Cyber Critical threshold. At the government's request OpenAI ran a phased preview to about twenty vetted partners while a cyber framework was developed^[1].

Does GPT-5.6 fix GPT-5.5's hallucination problem?

Unknown. OpenAI published only coding, biology, and cybersecurity evaluations in the preview and held back hallucination and knowledge-work numbers^[1]. Pair the model with source-grounded verification for high-stakes output.

How big is GPT-5.6 Sol's context window?

Third-party reports put it around 1.5 million tokens, roughly 43% larger than the GPT-5.5 generation, but OpenAI's preview post did not confirm a spec sheet^[1].

What changed in prompt caching?

Explicit cache breakpoints and a 30-minute minimum cache life. Cache writes bill at 1.25x the uncached input rate; cache reads keep the 90% discount^[1].

How do I call GPT-5.6 with the OpenAI SDK?

Point the client at https://api.reapi.ai/v1 and set model to gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna, keeping the dots in the id.

Reading a three-tier release

GPT-5.6 is two releases wearing one announcement. One is an ordinary, well-executed family update: a sensible naming reset, a mid-tier that halves cost without losing capability, and an ultra mode that bakes multi-agent orchestration into the endpoint. The other is a frontier model whose initial availability was set in coordination with a government, shipped behind a 700,000-GPU-hour safety stack.

For anyone actually building, the useful summary is smaller than either story. The benchmark lead at the top is 0.8 points and vendor-reported. The change you will feel is Terra's price. That is how to use GPT-5.6 sensibly: default to the middle tier, reserve Sol for the tasks that pay for ultra, and keep source-grounded verification in place until the expanded evaluation suite tells you whether the hallucination problem moved.

References

1. OpenAI. Models — the GPT-5.6 family, reasoning modes, and specifications. Retrieved July 2026 from platform.openai.com/docs/models
2. OpenAI. Pricing — per-token rates by model and tier. Retrieved July 2026 from platform.openai.com/docs/pricing

Further reading

• reAPI. How to use Claude Opus 5. reapi.ai/blog/how-to-use-claude-opus-5
• reAPI. How to use Claude Fable 5. reapi.ai/blog/how-to-use-claude-fable-5
• reAPI. Model catalog. reapi.ai/models