GPT-5.6 Sol, Terra & Luna Explained: Tiers, Pricing & Benchmarks (2026)

On June 26, 2026, OpenAI previewed GPT-5.6 — not a single model, but a family of three: Sol, Terra, and Luna. As of August 2026 it's OpenAI's newest model family. After a short, government-reviewed limited preview, it reached general availability on July 9, 2026, rolling out across ChatGPT, the API, Codex, and GitHub Copilot. The headline is the tiering: instead of one frontier model and a couple of "mini" spin-offs, OpenAI ships three models tuned to three jobs, plus new controls for how hard the model thinks.

This guide covers what each tier does, confirmed pricing, independent benchmarks, and how to choose — with vendor claims and independently-measured numbers clearly separated.

Want the full picture on the previous generation? Read our GPT-5.5 complete guide — benchmarks, pricing, and the agentic-coding patterns GPT-5.6 builds on.

What is GPT-5.6?

GPT-5.6 is the successor to GPT-5.5, released as three distinct models under one version number. OpenAI's framing is that most teams don't need one model for everything — they need the right model for each task. So the family splits along the classic capability-versus-cost curve:

• Sol — the flagship, built for "frontier reasoning and long-horizon agentic work." This is the model for the hardest problems: complex coding across large codebases, multi-step agents, scientific reasoning, and defensive security research.
• Terra — the balanced, production workhorse. OpenAI describes it as matching GPT-5.5's performance at roughly 2x lower cost at launch (~2.5x after the July 30 cut) — aimed at high-volume business tasks like customer support, internal tools, and document analysis.
• Luna — the cheapest, latency-optimized member, for high-volume everyday work: summarization, drafting, classification, and routine automation where latency and price matter more than raw reasoning depth.

The three carry the API names gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna (the bare gpt-5.6 alias routes to Sol). All three are reasoning models with roughly a 1M-token context window.

What's the difference between Sol, Terra, and Luna?

Here's the family at a glance — what each model is for and where it sits on the cost curve.

How much does GPT-5.6 cost?

OpenAI's published API pricing per 1 million tokens, current as of August 20, 2026. On July 30, 2026 OpenAI cut Luna by 80% and Terra by 20%; Sol was left at launch pricing. Cached input gets a large discount across all three.

The standout is Terra: if OpenAI's claim that it matches GPT-5.5's quality holds up in practice, you get the previous flagship's capability for less than half the price — $2/$12 against GPT-5.5's $5/$30. For most production workloads — the bulk of real token spend — that's the headline, not Sol's benchmark records.

What are the max, ultra, and pro modes?

GPT-5.6 adds new controls over how hard the model thinks — and it's worth being precise, because these are not separate models or price tiers. They all run on gpt-5.6-sol at Sol's pricing:

• max reasoning effort — a new top rung on the reasoning-effort dial (none, low, medium, high, xhigh, max). It gives Sol more time to reason before answering, for problems where you'd rather wait and get it right.
• Sol Pro — a reasoning mode that spawns independent parallel agents that each work in isolation, then merges the best result. It's the successor to the old GPT-5.5-Pro approach.
• Sol Ultra — goes further with four cooperating sub-agents that communicate mid-task and synthesize a joint answer. It burns several times the tokens of a standard Sol call and is offered in Codex and ChatGPT Work. OpenAI reports it lifts Terminal-Bench 2.1 from Sol's 88.8% to 91.9%. See our GPT-5.6 Sol Ultra vs Claude Fable 5 deep-dive on whether that extra compute is worth it.

Don't confuse Sol Ultra with Sol “Ultrafast.” Ultrafast, announced August 13, 2026, is a speed-focused service tier — the same Sol model served on Cerebras hardware — not a compute mode. Ultra is “same model, more thinking”; Ultrafast is “same model, served faster.” Details in the what's-new section below.

OpenAI also changed how it reports benchmarks: rather than one score, it now publishes a curve across reasoning-effort levels — a useful reminder that "how good is the model" depends heavily on how much thinking (and how many tokens) you let it spend.

What's new in GPT-5.6 vs GPT-5.5?

Beyond the tiers, GPT-5.6 ships three notable changes from GPT-5.5:

• Programmatic Tool Calling (PTC) — in the Responses API, the model can write and run JavaScript in an isolated, network-less sandbox to orchestrate its own tool calls (loops, conditionals, parallel calls) instead of round-tripping each call through your app. OpenAI cites customer-reported savings (one customer reported roughly 63% fewer total tokens).
• Persisted reasoning — reasoning state can be reused across turns for better multi-turn quality and cache efficiency.
• Token efficiency — OpenAI's headline claim is up to 54% fewer output tokens than GPT-5.5 on agentic coding (a vendor figure; independent measurement of general-task token use was more modest). Combined with Terra and Luna's lower prices, that's the real economic story.

What's new since the GPT-5.6 launch? (August 2026)

GPT-5.6 has moved fast since GA. The changes that matter, date-stamped (last verified August 20, 2026):

• July 30, 2026 — official price cuts. OpenAI cut Luna by 80% ($1/$6 → $0.20/$1.20) and Terra by 20% ($2.50/$15 → $2/$12) on the API. Sol stayed at $5/$30. These are permanent list-price changes, already reflected in the pricing table above.
• August 10, 2026 — GPT-5.6-Cyber. A purpose-trained cybersecurity model built on the Sol base, available only through OpenAI's Daybreak Red trusted-access program (identity verification, monitoring, and mandatory hardware security keys for individual accounts from September 1, 2026). On OpenAI's own evals it scores 95.0% on Advanced Cybersecurity Completion Rate versus 57.3% for GPT-5.5-Cyber — a vendor-reported figure, not independently verified. No public API, no pricing, not on OpenRouter.
• August 13, 2026 — Sol “Ultrafast.” A service tier, not a new model: the same Sol served on Cerebras wafer-scale hardware at up to 750 output tokens per second — “up to 14× faster than Standard” per OpenAI, with the headline speed comparisons (11× faster than Claude Fable 5) coming from Cerebras’ own framing of Artificial Analysis data. Limited preview, apply with a use case; pricing undisclosed. Sol-only for now — and not to be confused with Sol Ultra, the multi-agent compute mode covered above.
• August 17, 2026 — OpenRouter's 50% Sol promo. OpenRouter lists openai/gpt-5.6-sol at $2.50/$15 — exactly half of list. This is an OpenRouter-only, non-BYOK, time-limited promotion, not an official OpenAI cut: the OpenAI API price is still $5/$30. Don't build long-term cost models on it.
• August 17, 2026 — vision results land. Roboflow's VLM benchmark (its own harness) measured Sol at 46.2 mAP@50 on object detection versus GPT-5.5's 13.8 — a step-change for OpenAI models — with counting accuracy up to 73.0% from 64.9%. The caveat, from the benchmark's author himself: Gemini 3.5 and 3.7 Flash still beat it on most vision metrics at a lower per-image price, and GPT-5.6 is unstable on large images (≥ ~2,000×2,000 px).

How good is GPT-5.6 Sol? (Benchmarks)

Now that GPT-5.6 is generally available, independent testing has landed. From Artificial Analysis (independent, Intelligence Index v4.1.1, August 2026, models at maximum reasoning effort):

Sol lands at 61 on aggregate intelligence — behind Claude Opus 5 (63 at max effort) and Claude Fable 5 (62), tied with Grok 4.6 — but #1 on the agentic Coding Agent Index, measured in Codex. On ARC-AGI (independent, ARC Prize), Sol posted 96.5% on ARC-AGI-1 and 92.5% on ARC-AGI-2, and became the first model to make meaningful progress on the far harder ARC-AGI-3. OpenAI's own reported figures for Sol include SWE-Bench Pro ~64.6%, Agents' Last Exam ~52.7%, and Terminal-Bench 2.1 88.8% (91.9% in Ultra mode). Note absolute index scores drift as Artificial Analysis re-versions its harness — the relative ordering is the durable signal, and this snapshot is dated August 2026 (Intelligence Index v4.1.1).

On cybersecurity and biology, OpenAI's system card rates all three tiers High capability but below the "Critical" threshold — in testing the models could find vulnerabilities and pieces of exploits but could not autonomously carry out end-to-end attacks against hardened targets. OpenAI's stated view is that GPT-5.6 is better at finding and fixing vulnerabilities than at exploiting them.

Meta benchmarked Muse Spark 1.2 against Terra, not Sol

One footnote on the tiering that matters when you read a competitor's benchmark chart. When Meta launched Muse Spark 1.2 and its Muse Code terminal agent on August 5, 2026, its Terminal-Bench 2.1 comparison set included GPT-5.6 Terra — the mid-tier model — but not Sol:

Against Terra, Muse Spark 1.2 wins by 1.1 points. Against Sol's own reported 88.8%, it loses by about six. Meta's DeepSWE v1.1 chart has the same shape — Claude Opus 5 65.0, Terra 64.8, Muse Spark 1.2 59.3, Grok 4.5 56.6 — again with Terra standing in for the family. On price that is a defensible pick: Terra at $2/$12 (since the July 30 cut) sits far closer to Muse Spark 1.2's $1.25/$4.25 than Sol at $5/$30 does. It is not a flagship-versus-flagship read, though, and coverage that reported it as one got the tiering wrong.

The usual caveats cut both ways. Meta's numbers are vendor-run and published as chart images only, and Muse Spark 1.2 is absent from the verified Terminal-Bench leaderboard — but so is Claude Opus 5, so the top of that chart is vendor claim against vendor claim. Our Muse Spark 1.2 benchmark breakdown covers what is and is not independently confirmed.

How do you access GPT-5.6?

As of the July 9 general availability, GPT-5.6 is broadly accessible:

• ChatGPT — Plus, Pro, Business, and Enterprise users can select Sol; Free and Go users get Terra.
• API — all three tiers (gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna) are open, with max reasoning effort and, for Sol, the Pro and Ultra modes.
• Codex and ChatGPT Work — including Sol Ultra for the hardest agentic coding.
• GitHub Copilot — Sol, Terra, and Luna became available the same day.

Which GPT-5.6 model should you use?

Because the tiers map to jobs, model routing is straightforward. Use this as a starting decision matrix:

What GPT-5.6 means for teams building AI agents

The most useful read on GPT-5.6 isn't the benchmark table — it's what the tiering and the Ultra sub-agent mode signal: agentic, long-horizon work is now the product, not a side feature. If you're shipping agents, a few things follow.

First, the economics shift. With Terra at roughly half of GPT-5.5's price and Luna cheaper still, the right architecture is increasingly a routing one: send the easy turns to Luna, the bulk to Terra, and escalate only the genuinely hard steps to Sol. Teams that hard-code a single expensive model for everything will overpay.

Second, supervision matters more, not less. OpenAI's own system card flags that GPT-5.6 shows a greater tendency than GPT-5.5 to go beyond the user's intent — including taking actions the user didn't ask for — even though absolute rates remain low. For anyone wiring a model into tools, file systems, or CI, that's a direct instruction: scope agent permissions tightly, log every tool call, and keep a human approval gate on destructive actions.

A practical adoption checklist:

• Build an eval harness first. Don't swap models on vibes — measure your real tasks against GPT-5.5 before and after.
• Route by difficulty. Luna → Terra → Sol, with explicit escalation rules, not one model for everything.
• Cap cost and set budgets. max, Pro, and Ultra can burn a lot of tokens; meter them.
• Lock down agent permissions. Least-privilege tool access, audit logs, and human-in-the-loop on anything irreversible.
• Keep a rollback path. Pin a known-good model version so you can revert instantly if behavior drifts.

If you're new to wiring models into autonomous workflows, our AI coding agents complete guide covers the patterns — tool use, sandboxing, and supervision — that GPT-5.6 makes more relevant, not less.

What it means for hiring and engineering teams

As frontier models get better at long-horizon coding, the scarce skill stops being "can write the code" and becomes "can direct, review, and contain an agent that writes the code." The engineers who compound in value are the ones strong at codebase navigation, test design, security review, and agent supervision — the judgment work models still can't own.

That's exactly the profile Codersera vets for. If you're extending your team to ship AI-powered products faster — and you want developers who are fluent with agentic tooling rather than threatened by it — hire vetted remote developers through Codersera and start with a risk-free trial.

FAQ

When was GPT-5.6 released?

OpenAI previewed GPT-5.6 (Sol, Terra, and Luna) on June 26, 2026 as a government-reviewed limited preview, then made it generally available on July 9, 2026 across ChatGPT, the API, Codex, and GitHub Copilot.

What's the difference between GPT-5.6 Sol, Terra, and Luna?

Sol is the flagship for the hardest reasoning, coding, and agentic work ($5/$30). Terra is a balanced everyday model that OpenAI says matches GPT-5.5's performance at a fraction of the cost ($2/$12). Luna is the fastest and cheapest, built for high-volume tasks ($0.20/$1.20).

How much does GPT-5.6 cost?

Per 1 million tokens (after OpenAI's July 30, 2026 cuts): Sol is $5 input / $30 output, Terra is $2 / $12, and Luna is $0.20 / $1.20. Cached input is $0.50 / $0.20 / $0.02 respectively. Sol matches GPT-5.5's pricing; Luna got an 80% cut and Terra 20%. OpenRouter also runs a time-limited, non-BYOK promo listing Sol at $2.50/$15 — OpenAI's own API price is unchanged.

What are Sol Ultra and Sol Pro?

They're compute modes of Sol, not separate models or tiers — both run on gpt-5.6-sol at Sol's price. Sol Pro spawns independent parallel agents and merges the best result; Sol Ultra runs four cooperating sub-agents that communicate mid-task, lifting Terminal-Bench 2.1 from 88.8% to 91.9% at the cost of several times the tokens.

Is GPT-5.6 better than GPT-5.5?

On the independent benchmarks cited here, GPT-5.6 Sol ranks above GPT-5.5 — it scores 61 at max effort on Artificial Analysis's Intelligence Index (v4.1.1) and tops its Coding Agent Index at 80. Whether it's "better" for you also depends on cost, latency, and task mix. It also adds Programmatic Tool Calling, a new max reasoning effort, and — through Terra and Luna — lower prices for the same or near-same quality that OpenAI claims.

What context window does GPT-5.6 have?

All three tiers offer roughly a 1M-token context window with up to 128K output tokens.

Why does Meta compare Muse Spark 1.2 to GPT-5.6 Terra instead of Sol?

Meta's August 2026 launch charts benchmark Muse Spark 1.2 against Terra, the mid-tier model, rather than the Sol flagship. On price that is defensible — Terra at $2/$12 is much closer to Muse Spark 1.2's $1.25/$4.25 than Sol at $5/$30. On capability it flatters Meta: Muse Spark 1.2's claimed 82.9% on Terminal-Bench 2.1 beats Terra's 81.8% but sits well below Sol's reported 88.8%.