GPT-5.6 Is Here: OpenAI's Sol, Terra, and Luna Explained (Pricing, Benchmarks, and Which to Use)

OpenAI released GPT-5.6 on July 9, 2026, and this time it is not one model but three: Sol, the new flagship; Terra, a balanced everyday tier; and Luna, the cheapest and fastest. All three are live now in the API and in Codex, priced from $1 per million input tokens, and OpenAI says the top tier sets a new state of the art on agentic coding benchmarks. Here is what actually changed, what each tier costs, and which one you should reach for.

The launch lands in a crowded fortnight. Anthropic shipped Claude Sonnet 5 on June 30 and, by its own reporting, has pulled ahead of OpenAI on revenue run-rate. GPT-5.6 is OpenAI's answer, and the interesting part is the shape of it: a model family split by price rather than one do-everything model, aimed at developers who now watch cost per task as closely as raw capability.

What is GPT-5.6? Sol, Terra, and Luna explained

GPT-5.6 is a three-tier family, first previewed on June 26 and made generally available on July 9, 2026. Sol is the flagship, built for frontier reasoning and long-horizon agentic work such as multi-step coding and tool use. Terra is the balanced middle: OpenAI positions it as competitive with the older GPT-5.5 while costing roughly half as much, which makes it the sensible default for everyday production work. Luna is the fast, cheap tier for high-volume tasks where latency and price matter more than squeezing out the last few points of accuracy. The naming is new for OpenAI, which has usually shipped a single numbered model plus "mini" and "nano" variants. Splitting the release into three named siblings is a bet that most teams do not need the flagship for most calls, and would rather pick a tier per workload.

GPT-5.6 pricing: what Sol, Terra, and Luna cost

Pricing is the clearest signal of who each tier is for. These are the published per-million-token rates at general availability:

Tier	Input / 1M	Output / 1M	Positioning	Best for
Sol	$5	$30	Flagship reasoning	Hardest agentic coding, long-horizon tasks
Terra	$2.50	$15	Balanced (≈GPT-5.5 at ~2x cheaper)	Everyday production default
Luna	$1	$6	Fast and cheap	High-volume, latency-sensitive work

Two things stand out. Luna at $1 input is aggressive: it undercuts most frontier-adjacent models and pushes OpenAI further into the "good enough, cheap, fast" territory where the flash-tier models have quietly been winning production. And the whole family changes how prompt caching is billed. GPT-5.6 adds explicit cache breakpoints and a 30-minute minimum cache life, but cache writes now cost 1.25x the uncached input rate while reads keep the 90% discount. If your workload re-reads a big cached prompt many times, that is a win; if you write caches you rarely reuse, do the math before assuming it is free.

Do the benchmarks mean GPT-5.6 beats Claude?

Not cleanly, and this is where the marketing and the reality diverge. OpenAI's headline is that Sol sets a new state of the art on Terminal-Bench 2.1, the harness that tests real command-line workflows, at roughly 88.8% against Claude Opus 4.8's 78.9%. That is a genuine result on an agentic-coding benchmark that is hard to game.

But pick a different test and the picture flips. Independent developer Simon Willison notes that Fable 5 scored 80% on SWE-Bench Pro versus Sol's 64.6%, even as OpenAI argues about 30% of that benchmark's tasks are broken. On OpenAI's own Agents' Last Exam, Sol leads at 53.6 to Fable 5's 40.5. Willison's more useful observation is the hands-on one: the model "hasn't struck me as better than Fable at the kind of complex coding tasks" he actually runs. Benchmark leadership is now split across suites, so the honest read is that GPT-5.6 Sol trades wins with Anthropic's top models rather than clearing them. If you want the current cross-model picture, our best AI model for coding in 2026 ranking tracks this by benchmark and real cost.

What's new for developers: programmatic tool calling

The most concrete change for people building agents is programmatic tool calling in the Responses API. Instead of the usual round-trip loop, where the model asks for one tool call, waits for the result, then decides the next step, GPT-5.6 can write JavaScript that runs in an isolated V8 runtime with no network access and composes several tool calls in code. For workflows that chain many small operations, that cuts latency and token spend, because the orchestration happens inside one turn rather than across a dozen API hops. The sandbox has no network by design, so the model composes and transforms tool outputs rather than reaching out on its own, which keeps the security surface smaller than a fully open code interpreter.

Which GPT-5.6 tier should you actually use?

Match the tier to the job, not to the leaderboard. For most production traffic, start with Terra: it is the balanced default, roughly half the price of the flagship, and close enough to GPT-5.5 that you will not notice the difference on ordinary tasks. Drop to Luna for anything high-volume or latency-sensitive, classification, extraction, first-draft generation, cheap agent steps, where $1 input keeps a large workload affordable. Reserve Sol for the genuinely hard problems: long agentic coding runs, deep multi-tool reasoning, the calls where a better answer is worth six times the input price. A practical pattern is to route by difficulty, Luna or Terra for the bulk of calls and Sol only when a task escalates, rather than paying flagship rates for everything. If you are weighing this against Anthropic's line-up, our write-up of Claude Sonnet 5 covers the closest competitor to Terra on price and capability.

FAQ

Is GPT-5.6 out? Yes. GPT-5.6 was previewed on June 26, 2026 and became generally available on July 9, 2026 across the API and Codex, with all three tiers (Sol, Terra, and Luna) accessible to developers.

What is GPT-5.6 Sol? Sol is the flagship tier of the GPT-5.6 family, built for frontier reasoning and long-horizon agentic work. OpenAI reports it sets a new state of the art on the Terminal-Bench 2.1 coding benchmark. It is the most expensive tier at $5 input and $30 output per million tokens.

Is GPT-5.6 better than Claude? It depends on the task. Sol leads on Terminal-Bench 2.1 and OpenAI's Agents' Last Exam, but Anthropic's Fable 5 scores higher on SWE-Bench Pro, and some developers find no clear real-world coding advantage. Treat the two families as trading wins rather than one dominating.

Is GPT-6 coming out? OpenAI has not announced GPT-6. The GPT-5.6 release, a point update split into three tiers, suggests the near-term focus is on cost, speed, and agentic tooling rather than a single next-numbered flagship.

Sources

• OpenAI — Previewing GPT-5.6 Sol: official announcement, tier positioning, pricing, and Terminal-Bench 2.1 claims.
• OpenAI Help — A preview of GPT-5.6 Sol, Terra and Luna: preview date and availability details.
• MarkTechPost — OpenAI releases GPT-5.6, a three-tier family with programmatic tool calling: GA date, pricing table, caching changes, and V8-sandbox tool calling.
• Vellum — GPT-5.6 benchmarks explained: Terminal-Bench 2.1 scores across tiers and competitors.
• Simon Willison — The new GPT-5.6 family: Luna, Terra, Sol: independent benchmark caveats (SWE-Bench Pro, Agents' Last Exam) and hands-on impressions.
• CNET — OpenAI's GPT-5.6 is dropping on Thursday: launch timing and competitive context with Anthropic.