GPT-5.6 Sol, Terra & Luna: Tiers, Pricing, Benchmarks

GPT-5.6 is not a single model. It's a family of three — Sol, Terra and Luna — that OpenAI previewed on 26 June 2026 and made generally available on 9 July 2026 across ChatGPT, the API, Codex and GitHub Copilot. Instead of one frontier model plus a couple of "mini" spin-offs, OpenAI ships three models tuned to three jobs, plus finer controls over how hard the model thinks.

This page is the tier reference: what each model is for, what it actually costs today, what independent testing says, and how to choose. Vendor claims and independently-measured numbers are kept clearly apart. All pricing and specifications below were re-verified against OpenAI's developer documentation, and every Intelligence Index figure against Artificial Analysis, on 5 October 2026.

What are GPT-5.6 Sol, Terra and Luna?

They are three separate models released under one version number, split along the classic capability-versus-cost curve:

• Sol — the flagship of the 5.6 generation, built for frontier reasoning and long-horizon agentic work: complex coding across large codebases, multi-step agents, scientific reasoning and defensive security research. OpenAI's docs still describe it as its "flagship model for complex professional work", and the bare gpt-5.6 alias routes to it.
• Terra — the balanced production workhorse. OpenAI positions it as matching GPT-5.5's performance at a fraction of the cost, aimed at high-volume business work: customer support, internal tools, document analysis, RAG.
• Luna — the cheapest and most latency-friendly tier, for summarization, drafting, classification and routine automation where price and speed beat reasoning depth.

The API names are gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna. All three are reasoning models and all three carry identical context and output limits — the difference is capability and price, not window size.

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

Here is the family at a glance, with the verified hard specifications alongside the positioning. Astra is included in the last row purely for scale.

Two details worth knowing. Sol's usable input is capped at 922,000 tokens of the 1.05M window, with the remainder reserved for reasoning and output. And all three GPT-5.6 tiers support the full reasoning-effort ladder — none, low, medium, high, xhigh, max — defaulting to medium. Astra, notably, drops none: it always reasons.

How much does GPT-5.6 cost?

Current published API pricing per 1 million tokens, verified against OpenAI's pricing documentation on 5 October 2026. Batch requests run at exactly half the standard rate across every tier.

One warning if you price Sol from a third-party aggregator. OpenRouter's catalogue currently lists openai/gpt-5.6-sol at $2 / $10, which is wrong — that is GPT-6 Sol's rate. OpenAI's own pricing page and Artificial Analysis independently agree on $4 / $20 for GPT-5.6 Sol. OpenRouter's Terra and Luna rows do match OpenAI. If you have built a cost model off an aggregator feed, check the Sol row.

Sol's promotional pricing is still live — through at least 21 November 2026

This is the single most important cost fact on this page, and it is easy to get wrong because most write-ups still quote the launch price. On 21 August 2026 OpenAI cut Sol to $4 input / $20 output — 20% off input and 33% off output against the $5/$30 list price — and its documentation states the promotional pricing is "available at least through November 21, 2026."

That is a promotion with a floor, not a permanent list change. It makes Sol 2.5× cheaper than Astra on both input and output rather than 2×. If you are building a 12-month cost model, budget against the $5/$30 list price for anything past November, not the promo.

It is worth noting how much that discount has been overtaken, though. Even at the promotional $4/$20, Sol is twice the token price of GPT-6 Sol, which OpenAI lists at $2/$10 — and GPT-6 Sol scores higher on the independent index. The promo made Sol better value against Astra; it did not make it better value against Sol's own successor.

Terra and Luna reached their current prices earlier, in OpenAI's 30 July 2026 cut — Luna down 80% (from $1/$6) and Terra down 20% (from $2.50/$15). Those were permanent list changes, not promotions.

Long-context and reasoning-token billing

Two billing rules catch teams out:

• Requests over 272,000 input tokens are surcharged on every tier: 2× the input and cached-input rate, 1.5× the output rate. So a long-context Sol call bills at $8 input / $0.80 cached / $30 output. The 1M-token window is real, but the last two-thirds of it costs double.
• Reasoning tokens are billed as output tokens. OpenAI's reasoning guide is explicit: they "are not visible via the API, they still occupy space in the model's context window and are billed as output tokens." At xhigh or max effort, invisible reasoning can dominate your bill. Pro mode behaves the same way — it "aggregates the model work performed to produce the final answer and bills those tokens at the selected model's standard token rates."

The practical consequence: comparing tiers on the headline output price alone is misleading, because higher-effort settings on Sol generate far more billable reasoning tokens than a medium-effort Terra call. Measure cost per completed task, not cost per token. For a worked comparison against a cheaper competitor, see our DeepSeek V4 vs GPT-5.6 cost breakdown.

Is GPT-5.6 still OpenAI's newest model?

No — not since 3 September 2026, when OpenAI released GPT-6 Astra, described as "our most capable model for the most demanding work." Since then the GPT-6 line has filled out: as of 5 October 2026 OpenAI's docs carry model pages and pricing rows for gpt-6-astra, gpt-6-sol, gpt-6.1-sol and gpt-6-luna. Notably there is no GPT-6 Terra — the mid tier did not survive the generation change.

What that does and does not change for GPT-5.6 users:

• Astra is no longer gated. It launched behind the Trusted Access Program, but it now has a full public model page, a documented reasoning-effort ladder and a published API rate. GPT-5.6 is no longer the only generally-available option.
• The cheap half of GPT-6 is the real story. gpt-6-sol is $2/$10 and gpt-6-luna is $0.10/$0.50 — half the token price of GPT-5.6 Sol and Luna respectively, and both score higher on the independent index. OpenAI describes gpt-6.1-sol as delivering "near-Astra performance at a lower cost", which the benchmark section below confirms.
• Astra is more constrained. It does not support none reasoning effort, does not accept custom temperature, top_p or log probabilities. Function calling is listed as a supported feature and both Chat Completions and Responses are supported endpoints, though the hosted tools (web search, computer use, MCP and the rest) are documented for the Responses API. If your integration leans on those hosted tools or on sampling parameters, GPT-5.6 is the drop-in and Astra is a migration.
• The price gap is wide. $10/$50 against Sol's promotional $4/$20 is 2.5× on both sides of the meter.

We keep Astra coverage deliberately brief here. The dedicated pages go deep: the complete GPT-6 Astra guide, Astra pricing and API costs, and the direct GPT-6 Astra vs GPT-5.6 Sol comparison.

Is GPT-5.6 being deprecated?

No. Re-checked on 5 October 2026: OpenAI's deprecations page still lists no shutdown or retirement date for gpt-5.6-sol, gpt-5.6-terra or gpt-5.6-luna. Neither the Astra launch nor the GPT-6 Sol launch came with a GPT-5.6 sunset notice.

The reverse is true: GPT-5.6 models appear all over that page as the recommended replacements for models that are retiring. Two waves matter.

• 23 October 2026 — a large batch of legacy models retires, including gpt-4, gpt-4-turbo, gpt-4o-2024-05-13, o1, o3-mini and o4-mini. The named successors are GPT-5.6 Sol, Terra and Luna depending on the model.
• 11 December 2026 — the original GPT-5 line goes: gpt-5-2025-08-07 → gpt-5.6-sol, gpt-5-mini-2025-08-07 → gpt-5.6-terra, gpt-5-nano-2025-08-07 → gpt-5.6-luna, and gpt-5-pro-2025-10-06 → gpt-5.6-sol with reasoning.mode: pro. o3 and o3-pro retire the same day onto Sol.

If you are on original GPT-5 or GPT-5-mini, 11 December 2026 is your real deadline, and OpenAI's named target is GPT-5.6. If you are migrating anyway, read the benchmark section below before you take that suggestion literally — GPT-6 Sol is cheaper per token than GPT-5.6 Terra and scores above GPT-5.6 Sol, so OpenAI's documented replacement is not the best-value destination. Anyone already on GPT-5.6 has no forced move at all.

What are the max, Pro and Ultra modes?

These are not separate models or price tiers. They all run on gpt-5.6-sol and bill at Sol's standard rates — there is no gpt-5.6-sol-pro line item on OpenAI's pricing page.

• max reasoning effort — the top rung of the effort dial (none, low, medium, high, xhigh, max). More thinking time before answering, for problems where you would rather wait and be right. Remember the reasoning tokens are billed as output.
• Sol Pro — spawns independent parallel agents that each work in isolation, then merges the best result. Successor to the old GPT-5.5-Pro approach. Billed by aggregating all the work at Sol's standard rates.
• Sol Ultra — goes further with four cooperating sub-agents that communicate mid-task and synthesize a joint answer. Burns several times the tokens of a standard Sol call; offered in Codex and ChatGPT Work. OpenAI reported it lifting Terminal-Bench 2.1 from Sol's 88.8% to 91.9%. Our Sol Ultra vs Claude Fable 5 piece digs into whether the extra compute pays.

Don't confuse Sol Ultra with Sol "Ultrafast." Ultrafast, announced 13 August 2026, is a service tier — the same Sol model served on Cerebras wafer-scale hardware — not a compute mode. Ultra is "same model, more thinking"; Ultrafast is "same model, served faster." It remains a limited preview with undisclosed pricing.

How good is GPT-5.6? (benchmarks)

Independent measurement

From Artificial Analysis, Intelligence Index v4.3.2, checked 5 October 2026. Two things to get straight before reading the numbers.

First, AA rebased the index from v4.1.1 to v4.3.2. The benchmark basket changed, so every model's score moved; v4.1.1 and v4.3.2 figures are not comparable and there is no conversion factor. If you have an older copy of this table, or you find GPT-5.6 scores in the high 50s or 60s elsewhere, those are v4.1.1 numbers.

Second, every row below names the reasoning-effort variant, because for this family the variant matters more than the model. AA publishes a separate score per effort level, and the spread inside GPT-5.6 Sol alone runs from 28.33 to 46.97. Rows are at max effort unless stated:

This table reverses the conclusion this page used to carry. When Astra launched, its score and Sol's both rounded to 61 on index v4.1.1, which made "stay on Sol, it is the same intelligence for less money" the right call. On v4.3.2 that is no longer true in either direction:

• Astra is now clearly ahead of GPT-5.6 Sol, 52.67 against 46.97 — a 5.7-point gap, not a rounding artefact. It costs 64% more per Index task ($3.26 vs $1.99), so it is still a premium, but it is a premium for measurably more capability.
• GPT-6 Sol beats GPT-5.6 Sol on every axis at once. Higher score (47.63 vs 46.97), 48% cheaper per Index task ($1.04 vs $1.99), 32% faster output (107.1 vs 81.2 tok/s), and half the token price ($2/$10 vs $4/$20). There is no dimension on which GPT-5.6 Sol wins.
• GPT-6.1 Sol is the cost-performance outlier. 51.83 — within 0.8 points of Astra — at $0.72 per Index task, which is cheaper per task than GPT-5.6 Sol, GPT-6 Sol and GPT-5.6 Terra. Its weakness is latency: 273 seconds median time to first chunk at max effort, against 108 seconds for GPT-5.6 Sol. For interactive use that is disqualifying; for batch and agentic work it is the best value in the table.
• GPT-6 Luna undercuts GPT-5.6 Luna too, 38.12 against 37.32, at $0.07 per Index task against $0.18 and half the token price.

Note also that Terra is not the fastest tier in the family, which this page previously claimed on v4.1.1 data. At max effort Luna is fastest at 126.1 tok/s, then Terra at 110.8, then Sol at 81.2.

One figure we have deliberately deleted rather than updated: the leaderboard ranks (#10 and #8 "of 202"). AA's field has grown to 224 models, so a rank captured against the old denominator is wrong even when the score behind it is right, and there is no honest way to carry it across a rebase.

How much does reasoning effort change each tier?

This is the most useful table on the page for anyone actually routing traffic, and it is the one almost nobody publishes. AA scores each GPT-5.6 tier separately at every rung of the effort ladder. Index score on the left of each pair, USD per Index task on the right:

Three things fall out of it that change how you should route:

• Effort outweighs tier over much of the range. Luna at max scores 37.32 for $0.18 per task. Sol at low scores 33.47 for $0.26. The cheap model thinking hard beats the expensive model thinking lightly, on both score and cost.
• Sol at high is the sweet spot, not Sol at max. 42.35 for $0.81 against 46.97 for $1.99: the last 4.6 points cost you 2.5× per task. Whether that is worth it is an eval question, not a default.
• The default is medium, and the defaults are mediocre. Nobody comparing "GPT-5.6 Sol" against a rival's maximum-effort number at Sol's default 39.24 is running a fair test.

Vendor-reported figures

These come from OpenAI's own GPT-5.6 launch materials and have not been independently reproduced. Useful for direction, not for settling arguments: SWE-Bench Pro around 64.6%, Agents' Last Exam around 52.7%, and Terminal-Bench 2.1 at 88.8% for standard Sol, rising to 91.9% in Ultra mode. On ARC-AGI — independently administered by the ARC Prize team — Sol posted 96.5% on ARC-AGI-1 and 92.5% on ARC-AGI-2.

OpenAI's system card rates all three tiers High capability but below the Critical threshold on cybersecurity and biology: in testing the models could find vulnerabilities and pieces of exploits but could not autonomously run 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.

A tiering trap when you read competitors' charts

When Meta launched Muse Spark 1.2 in August 2026, its Terminal-Bench 2.1 comparison chart included GPT-5.6 Terra — the mid-tier model — but not Sol. Muse Spark's claimed 82.9% beats Terra's 81.8% by 1.1 points, but sits about six points below Sol's reported 88.8%. On price that substitution is defensible; as a flagship-versus-flagship read it is not, and several outlets reported it as one. Our Muse Spark 1.2 benchmark breakdown covers what is and isn't confirmed. The same care applies to any chart naming "GPT-5.6" without saying which tier — the gap between Luna and Sol is far larger than the gap between most competing flagships.

The same trap exists one level deeper, inside Artificial Analysis's own labels, and it has caught us out. AA's unqualified model name does not map to the same effort tier across models. Bare GPT-5.6 Sol on AA is its low variant — 33.47, byte-identical to the row labelled GPT-5.6 Sol (Low). Bare GPT-5.6 Terra is its medium variant, 30.09. Bare GPT-6 Astra is its xhigh variant, 52.39. So quoting two unqualified AA figures against each other can manufacture an 18-point gap that is almost entirely a tier artefact. Always resolve the variant before you compare, and always name it in the text so a reader can check you.

For a like-for-like flagship comparison, see Claude Opus 5 vs GPT-5.6.

How do you access GPT-5.6?

• ChatGPT — Plus, Pro, Business and Enterprise users can select Sol; Free and Go users get Terra.
• API — all three tiers are generally available, with the full reasoning-effort ladder 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, all available since GA.

A fuller timeline of what shipped when is in our GPT-5.6 release date and what's new rundown.

Which GPT-5.6 tier should you pick?

Because the tiers map to jobs rather than to raw quality bands, routing is straightforward. Start here:

The step-up rule has changed, and so has the step-sideways rule. The old version of this page told you to stay on GPT-5.6 Sol because it matched Astra's measured intelligence more cheaply. On index v4.3.2 that is no longer the case, so:

• GPT-5.6 Sol → GPT-6 Sol is close to a free upgrade. Higher score, lower cost per task, faster output, half the token price, same 1.05M context. The only reasons to defer are a pinned model version you have not re-evaluated, or an integration that depends on behaviour you have characterised on 5.6.
• GPT-6 Sol → Astra is the real step-up decision, and the old three-part test still applies: GPT-6 Sol at max must measurably fail tasks you need to pass on your eval set; the work must be the long end-to-end reasoning or computer-use shape Astra was built for; and you must be able to absorb 5× the token price and a Responses-API migration.
• Latency is the hidden axis. GPT-6.1 Sol wins on score-per-dollar and loses badly on time to first chunk. Decide which of those your users feel before you pick on the table alone.

And a budget note that outlives this article: if you do stay on Sol on the strength of $4/$20, model your spend past 21 November 2026 at the $5/$30 list price. The promotion is guaranteed only to that date — which makes the gap to GPT-6 Sol's $2/$10 wider still.

What does the GPT-5.6 tiering mean for teams building agents?

With a 20× price spread between Luna and Sol inside one family — and the effort ladder adding another ~60× inside each tier — the right design is a routing one: send easy turns to the cheap tier, the bulk to the middle, escalate only genuinely hard steps, and reserve the flagship for the slice that fails everything below it. Teams that hard-code one expensive model at max effort for everything overpay by a large multiple, and that is now true within OpenAI's own lineup, not just across vendors. Note that the GPT-6 generation collapsed the middle rung: there is no GPT-6 Terra. The family is Astra, Sol and Luna, with GPT-6 Sol at $2/$10 sitting roughly where GPT-5.6 Terra used to on price while scoring well above it.

Supervision matters more, not less. OpenAI's 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 stay low. If you are wiring a model into tools, file systems or CI, scope agent permissions tightly, log every tool call, and keep a human gate on destructive actions.

A practical adoption checklist:

• Build an eval harness before you switch anything. It is the only thing that answers Sol-versus-Astra for your workload.
• Route by difficulty — Luna → Terra → Sol, with explicit escalation rules.
• Meter reasoning effort. xhigh and max bill invisible tokens as output; cap them.
• Batch anything asynchronous. Half price across every tier, for work that can wait.
• Watch the 272K threshold. Crossing it doubles input cost — trim context before you pay the surcharge.
• Pin a known-good model version so you can revert instantly if behaviour drifts.

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

The bottom line

GPT-5.6 remains a perfectly good family: nothing in it is deprecated, all three tiers are generally available, Sol is still discounted, and if you are running it in production today there is no fire to put out. But the honest reading of index v4.3.2 is that GPT-5.6 has been passed on both axes at once. GPT-6 Sol scores higher than GPT-5.6 Sol, runs 32% faster, costs 48% less per benchmark task and half as much per token. GPT-6 Luna does the same to GPT-5.6 Luna. There is no longer a price-performance argument for choosing GPT-5.6 on a new build.

The decision rule, updated. On a new project, start at GPT-6 Sol and drop to GPT-6 Luna wherever quality is not the binding constraint; use GPT-6.1 Sol for batch and agentic work where a four-minute first token is acceptable; evaluate Astra only when Sol at max demonstrably fails your own evals. On an existing GPT-5.6 deployment, nothing forces a move — but run your eval set against GPT-6 Sol before you renew a budget, because the comparison is not close. And whatever you pick, set the effort level deliberately: the ladder table above shows it moves scores more than the tier choice does over much of the range.

Re-check the pricing page after 21 November 2026 — that is the one dated thing on this page that will change on its own.

If you're building AI-powered products and want engineers who are fluent at directing, reviewing and containing agents rather than threatened by them, you can hire vetted remote developers through Codersera.

FAQ

What are GPT-5.6 Sol, Terra and Luna?

They are OpenAI's three GPT-5.6 model tiers, released together in July 2026. Sol is the flagship for hard coding, agents and deep reasoning. Terra is the balanced production model for support, internal tools and document work. Luna is the cheapest and fastest, built for summarization, classification and high-volume automation. All three share a 1.05M-token context window and a 128K max output.

Which GPT-5.6 tier should I use?

Within the family: Terra for production volume at $2/$12, Luna where quality is not the binding constraint, Sol for multi-step agents and hard refactors. Note that Luna, not Terra, is the fastest at max effort — 126 tokens/sec against Terra's 111. But on a new build, compare GPT-6 Sol first: at $2/$10 it is cheaper than Terra on output and scores 5.5 index points above it.

How much does GPT-5.6 cost?

Per 1M tokens as of 5 October 2026: Sol is $4 input / $20 output (promotional), Terra $2/$12, Luna $0.20/$1.20. Cached input is $0.40, $0.20 and $0.02 respectively. Batch requests are exactly half price on every tier. Prompts above 272K input tokens are surcharged at 2× input and 1.5× output. Reasoning tokens bill as output tokens. Ignore OpenRouter's $2/$10 figure for Sol — that is GPT-6 Sol's rate, not GPT-5.6 Sol's.

Is GPT-5.6 still OpenAI's newest model?

No. GPT-6 Astra launched on 3 September 2026 as OpenAI's most capable model at $10 input / $50 output per 1M tokens, and OpenAI has since published GPT-6 Sol, GPT-6 Luna and GPT-6.1 Sol. GPT-5.6 remains generally available and fully supported, but it is no longer the value choice: GPT-6 Sol costs $2/$10 and scores above GPT-5.6 Sol on Artificial Analysis's Intelligence Index v4.3.2.

Is GPT-5.6 deprecated?

No. OpenAI's deprecations page lists no shutdown or retirement date for gpt-5.6-sol, gpt-5.6-terra or gpt-5.6-luna as of 5 October 2026, and neither the Astra nor the GPT-6 Sol launch carried a sunset notice. The opposite applies: GPT-5.6 models are the named replacements for retiring ones. A batch of legacy models including gpt-4, gpt-4-turbo, o1 and o4-mini retires on 23 October 2026, and the original GPT-5 line follows on 11 December 2026 — GPT-5 to Sol, GPT-5-mini to Terra, GPT-5-nano to Luna.

Should I use GPT-5.6 Sol or GPT-6 Astra?

Neither, as a first choice — start with GPT-6 Sol. On Artificial Analysis's Intelligence Index v4.3.2 at max effort, Astra scores 52.67, GPT-6 Sol 47.63 and GPT-5.6 Sol 46.97, while cost per Index task runs $3.26, $1.04 and $1.99 respectively. GPT-6 Sol beats GPT-5.6 Sol on score, speed, cost per task and token price simultaneously. Move up to Astra only when GPT-6 Sol at max measurably fails your own evals on long end-to-end reasoning or computer-use work and you can absorb 5× the token price.

Is GPT-6 Sol better than GPT-5.6 Sol?

Yes, on every measure that is published. On Artificial Analysis's Intelligence Index v4.3.2 at max effort GPT-6 Sol scores 47.63 against GPT-5.6 Sol's 46.97, costs $1.04 per Index task against $1.99, produces 107 output tokens/sec against 81, and lists at $2 input / $10 output per 1M tokens against GPT-5.6 Sol's promotional $4/$20. Both have the same 1,050,000-token context window. There is no axis on which GPT-5.6 Sol currently wins.

What does reasoning effort do to GPT-5.6 scores?

A great deal — more than the choice of tier over much of the range. On Intelligence Index v4.3.2, GPT-5.6 Sol runs from 28.33 with reasoning off to 46.97 at max, with cost per Index task rising from $0.03 to $1.99. Luna spans 15.53 to 37.32. The API default is medium, which puts Sol at 39.24 — so any comparison quoting "GPT-5.6 Sol" without naming an effort level is ambiguous by roughly 18 index points.

Is GPT-5.6 Sol's discounted price permanent?

No. The $4/$20 rate is promotional, introduced on 21 August 2026 as a 20% input and 33% output cut against the $5/$30 list price. OpenAI's documentation, re-checked on 5 October 2026, still guarantees it only "at least through November 21, 2026." Terra and Luna's prices are different — those came from a permanent list-price cut on 30 July 2026 and carry no expiry. Either way, GPT-6 Sol's permanent $2/$10 undercuts the Sol promotion.

What context window does GPT-5.6 have?

All three tiers have a 1,050,000-token context window with up to 128,000 output tokens. For Sol the usable input portion is capped at 922,000 tokens, with the rest reserved for reasoning and output. Be aware that any request exceeding 272,000 input tokens is billed at 2× the input rate and 1.5× the output rate on every tier.