GPT-5.6 is here: OpenAI's Sol, Terra, and Luna, and when to use each
OpenAI's new models are named after the sky and priced by brightness. The invoice gap is 5x; the quality gap is smaller than you think.
OpenAI now sells the sun, the earth, and the moon. On July 9, after a two-week limited preview, GPT-5.6 went generally available across ChatGPT, Codex, and the API as a family of three models: Sol, the flagship; Terra, the balanced middle; and Luna, the fast, cheap one. The names are new. The idea behind them is the most useful thing about this launch.
If you only remember one line, make it this one: the name tells you the price, and the price gap is bigger than the quality gap. Sol costs 5x what Luna costs. On the benchmark that matters most for real work, it scores five points higher. That ratio, not any single headline number, is what should decide which of these models you actually use. (For readers who don’t live in API pricing pages: models bill by the token, roughly three-quarters of a word, so a million tokens is about 750,000 words in or out.) This post walks each tier, the numbers behind it, and a routing rule you can apply tomorrow.
The number is the generation. The name is the brightness
Until now, a GPT version number named one model: GPT-5.4, then GPT-5.5 in April, each a single flagship with discounts hanging off it. With 5.6, OpenAI split the label in two. The number now identifies a generation, while Sol, Terra, and Luna identify capability tiers that OpenAI says can advance on their own schedules. A future Luna could ship without waiting for a new flagship.
The names themselves are Latin: Sol the sun, Terra the earth, Luna the moon. Brightness maps to capability and to price, which makes the lineup unusually easy to hold in your head. The moon is $1 per million input tokens, the earth $2.50, the sun $5, and each output rate is 6x its input rate. Everything else is shared family DNA: a 1,050,000-token context window, 128,000 max output tokens, and a February 16, 2026 knowledge cutoff on all three tiers, with text and image input and text output.
With this move, OpenAI joins a convention the rest of the frontier settled on years ago. Anthropic has shipped small, medium, and large as Haiku, Sonnet, and Opus since 2024; Google splits Gemini into Flash and Pro. Named tiers exist because buyers think in tiers: the question in practice is rarely “which company” but “how much model does this job need.” OpenAI’s version arrives late and, characteristically, with the most memorable branding of the three.
The launch also carried unusual backstory. The family first appeared on June 26 as a limited preview under a US government directive, an arrangement OpenAI publicly objected to. That episode says more about where AI policy is heading than about these models, and we covered it in depth when it happened. This post is about the models.
Sol: a new flagship at the old flagship’s price
GPT-5.6 Sol is the tier OpenAI wants benchmarked, and the one the bare gpt-5.6 API alias routes to. It is pitched at frontier coding, long-horizon agentic work, and research, and it posts the numbers you would expect from a flagship six months of progress later: on Terminal-Bench 2.1, an agentic coding benchmark run in a real terminal, Sol scores 88.8%, and 91.9% in its compute-heavy Ultra mode, a new state of the art. The previous generation, GPT-5.5, sits at 88.0%.
The quieter story is the invoice. Sol costs $5 per million input tokens and $30 per million output tokens, which is exactly what GPT-5.5 launched at in April. OpenAI absorbed a full generation of capability gains and held the flagship price flat. For anyone budgeting AI work, that is the launch in one sentence: the frontier got better and did not get more expensive.
What does “frontier coding and agents” mean if you don’t write code? Sol is the tier for work that runs long and compounds its own mistakes: a task that takes fifty steps, where a small error at step 3 quietly ruins step 40. Deep research reports, multi-hour automations, refactoring a codebase, anything you hand off and walk away from. For a question you’d ask and read the answer to in one sitting, flagship depth is mostly wasted.
In ChatGPT, Plus, Pro, Business, and Enterprise plans get Sol, with a heavier “Sol Pro” reserved for Pro and Enterprise. Free and Go users get Terra. Reasoning effort spans six settings, from none up to a new max tier, so one model ID covers everything from quick answers to hour-long agent runs. Sol also ships with what OpenAI calls its most robust safety stack to date, with tightened handling of sensitive cybersecurity requests; the deployment story behind that is its own saga.
Terra and Luna: the production default and the volume tier
GPT-5.6 Terrais the tier for the work that pays the bills. OpenAI’s positioning is blunt: GPT-5.5-level quality at half the price, $2.50 in and $15 out. Independent testing broadly agrees. On the Artificial Analysis Intelligence Index, Terra scores 55 to Sol’s 59, and its measured cost to complete the full benchmark suite works out to $0.55 per task against Sol’s $1.04. If your workload was happy on GPT-5.5, Terra is the same class of brain with the bill cut in half.
GPT-5.6 Lunais the volume play: $1 in, $6 out, the fastest of the three, and $0.21 per task on the same suite. The surprise is how little quality that discount costs. On the Coding Agent Index, Luna scores 75 to Sol’s 80. Five points matters at the frontier; for classification, extraction, summarization, and high-volume chat, it usually doesn’t. Artificial Analysis notes that Luna and Sol sit on the cost-quality frontier ahead of Terra, which is a technical way of saying: the moon is the bargain of the family.
Make that concrete. Say you run a support assistant that handles 10,000 conversations a day, each around 2,000 tokens in and 500 tokens out. On Sol, that’s about $250 a day, or $7,500 a month. On Terra, $125 a day and $3,750 a month. On Luna, $50 a day and $1,500 a month. Same conversations, same API, a 5x spread on the invoice. The only question that matters is whether your customers can tell the difference, and for bounded, well-instructed tasks like this one, they usually can’t.
Route by the job, not by the launch video
Three tiers only pay off if you actually route between them. The failure mode is putting everything on the flagship out of habit and paying 5x for work a cheaper tier does indistinguishably. The other failure mode is the reverse: anchoring on the $1 price and feeding Luna the gnarly multi-step tasks that Sol exists for, then concluding the family is overhyped.
A practical default: start new work on Terra. When a task fails in ways that look like reasoning depth (lost threads in long agent runs, subtle bugs in generated code) escalate that task to Sol. When a task succeeds boringly for a week, try it on Luna and watch the quality bar. Judge each move on cost per completed task, not cost per token, since a smarter tier that finishes in one attempt is often cheaper than a discount tier that needs three.
Against Claude, it’s a split decision
The competitive picture at launch is genuinely mixed, which is worth saying plainly because launch weeks rarely are. On the Artificial Analysis Intelligence Index, Anthropic’s Claude Fable 5 holds the top spot at 60 with Sol one point behind at 59, at roughly a third of the cost per task. On the same firm’s Coding Agent Index, Sol takes the lead: 80 against Fable 5’s 77.2. OpenAI’s own reported results have Sol ahead by 13.1 points on Agents’ Last Exam, a 55-field professional benchmark, while Anthropic’s Mythos 5 wins SWE-Bench Pro by a wide margin, 80.3% to 64.6%.
Early hands-on impressions add a useful asterisk. Simon Willison, testing the family at launch, reported that Sol hadn’t struck him as better than Claude on the complex coding tasks he uses daily, benchmark wins notwithstanding. None of this settles anything; it re-opens it. The right frame is the one we laid out in our task-by-task comparison of the big three: there is no best model, only a best-for-this, and GPT-5.6 just made several of those calls closer.
Caching you can budget, tools the model can script
Two API changes shipped alongside the models, and for teams running AI in production they may matter more than the benchmark deltas. The first is prompt caching you can plan around. Caching lets the API remember the unchanging front of your prompt (the system instructions, the reference documents) so you stop paying full price to resend it on every request. GPT-5.6 adds explicit cache breakpoints and a guaranteed 30-minute minimum cache life. Cache writes now bill at 1.25x the normal input rate, and cache reads keep their 90% discount, so cached input on Sol costs $0.50 per million tokens instead of $5. Agents that reuse a large system prompt all day get most of their input bill back, predictably rather than opportunistically.
The second is programmatic tool calling in the Responses API. Instead of ping-ponging one tool call per turn through the conversation, the model writes JavaScript that orchestrates its tools inside an isolated, network-less V8 sandbox, and early customers report token reductions of 38% to 63.5% on tool-heavy workloads. One caveat hides in the pricing page: requests over 272K input tokens bill at 2x input and 1.5x output, so the million-token window is real but the top of it is premium-priced.
Start on Terra, move with evidence
The celestial naming will get the jokes, but it earns its keep: cheap, balanced, flagship is now moon, earth, sun, and you will never need the pricing page to remember which is which. Under the mnemonic sits a real strategy. OpenAI held the flagship price flat, cut the production tier to half of it, and shipped a volume tier that gives up five benchmark points for an 80% discount. And because the tiers can now version independently, the thing to watch next isn’t the next big number. It’s a quiet Luna update that moves the floor of what $1 buys.
The way to have an opinion about any of this is to run your own work through it. All three tiers are in the CSuite catalog: Sol, Terra, and Luna, each one click from a prompt, alongside the rest of the catalog if you want to line them up against Claude and Gemini on your own tasks. Start on the earth. Reach for the sun when the work proves it needs one.
