What OpenAI actually previewed
On June 26, 2026, OpenAI previewed GPT-5.6, and the headline is that it is not one model but three: Sol, the flagship; Terra, the mid-tier; and Luna, the budget tier. This is a real, announced release with published pricing and benchmarks — not a rumor. The rollout has been unusual. During the preview the models are available only through the API and Codex, and only to a small group of trusted partners and organizations with an OpenAI account representative; there is no public waitlist, and GPT-5.6 is not in ChatGPT yet. Access was initially gated by a US government review that requested vetting of who could use the models, which OpenAI complied with. That restriction was lifted on July 8, 2026, clearing the way for a broader rollout beginning around July 9, with general availability planned in the coming weeks. For a point release, that is a lot of ceremony — a signal that the capability jump is large enough to have drawn government attention. You can track when the GPT-5.6 models appear on live pricing feeds via the OpenAI provider page and the full model index, which refresh daily.
GPT-5.6 pricing: Sol, Terra, and Luna
OpenAI priced the family in three clean tiers, all per 1M tokens. Sol, the flagship, is $5.00 input and $30.00 output — identical to GPT-5.5's headline rate, so the flagship price did not move. Terra, the mid-tier, is $2.50 input and $15.00 output. Luna, the budget tier, is $1.00 input and $6.00 output. Every tier carries the same 6× output multiplier, which is the single most important number on the rate card: output tokens cost six times more than input tokens at every level. That structure means any workload generating substantial text — code, summaries, structured JSON, long answers — will see its bill dominated by the output line, not the input line. The naming replaces the old mini and nano labels, but the shape is familiar: one premium flagship, one balanced middle, one cheap high-volume option. Before assuming any tier is right for you, run your real input-to-output ratio through the API cost estimator — with a 6× multiplier, output volume drives the cost, and the tier that looks cheapest per token is not always cheapest per task.
The Sol / Terra / Luna split: which tier does what
Sol is the frontier model, aimed at the hardest reasoning and agentic-coding work, and it is the tier that ships with Ultra mode. Terra is the workhorse middle tier at half Sol's price, positioned for the bulk of production traffic where you want strong capability without flagship cost — the same role Sonnet-class and Flash-class models play elsewhere. Luna is the volume tier at $1/$6, built for high-throughput, cost-sensitive jobs like classification, extraction, and simple chat. The practical planning move is the same one that works across every provider: do not default your whole application to the flagship. Route most traffic to Terra or Luna and reserve Sol for the requests that genuinely need frontier reasoning. That single decision typically moves your bill more than any per-token price difference between providers. Model the split for your own traffic on the API cost estimator before you commit a default.
Ultra mode: Sol's high-effort setting
The most interesting new lever is Ultra mode, a compute-intensive, high-effort setting for Sol that trades cost and latency for capability. It is not the default — you opt into it — and it is where Sol posts its top benchmark numbers. On OpenAI's Terminal-Bench 2.1 agentic-coding evaluation, Sol Ultra scores 91.9% versus base Sol's 88.8%, a meaningful gap that shows Ultra mode is doing real work on the hardest tasks. Some early coverage describes Ultra mode spawning subagents to parallelize a problem, though OpenAI's own framing emphasizes it as a compute-intensive high-effort mode rather than detailing the mechanism. The parallel worth drawing is to Claude Opus 4.8's Effort Control, which similarly dials reasoning depth up at the cost of more output tokens — see our Effort Control deep-dive for how that hidden dial reshapes the bill. Ultra mode is the same idea with a name: the highest-capability configuration is also the most expensive to run, so reserve it for tasks that clear the bar to justify it.
Benchmarks: where GPT-5.6 lands
OpenAI led with agentic coding. On Terminal-Bench 2.1, Sol Ultra scores 91.9% and base Sol scores 88.8%, edging both Claude's Mythos-class model and GPT-5.5, which sit at 88.0%. Gemini 3.1 Pro's score on that benchmark was not yet published at preview time. The more striking claim is about efficiency: on ExploitBench², OpenAI reports Sol is competitive with the Mythos preview while using only about one-third of the output tokens. If that holds on real workloads, it partly offsets the 6× output multiplier — you pay six times per output token, but you may generate far fewer of them per task. OpenAI also reports a reported context window around 1.5M tokens, up from GPT-5.5's ~1.05M, and noted safety-review findings: the models could surface software vulnerabilities but were unable to carry out autonomous cyberattacks, alongside meaningful capability gains in biology. Treat vendor benchmarks as a starting point, not a verdict — the number that matters is how the model does on your own prompts.
The output-multiplier catch — and the token-efficiency twist
Here is the nuance most launch coverage will miss. GPT-5.6 Sol keeps the same 6× output multiplier as GPT-5.5 and prices output at $30 per 1M — higher per output token than Claude Opus 4.8's $25 at a 5× multiplier. On sticker economics alone, that makes Sol look like the more expensive way to buy frontier quality, exactly as GPT-5.5 was. But OpenAI's ExploitBench² efficiency claim, if it generalizes, changes the arithmetic: a model that reaches the same answer using one-third the output tokens can be cheaper per completed task even at a higher per-token rate, because the bill is token count times price, not price alone. That is the whole thesis behind measuring blended cost per task rather than staring at the rate card. The only way to know whether Sol's efficiency offsets its multiplier for your workload is to measure it — put your real input-to-output ratio and expected token volumes into the side-by-side price comparison and the cost estimator once you have API access, and compare cost per task, not cost per token.
How to get access, and what to do now
During the preview you cannot simply sign up. Access is limited to trusted partners and organizations with an OpenAI account representative, through the API and Codex only, with general availability in the coming weeks now that the government restriction has lifted. If you are not in the preview, the productive move is to prepare rather than wait. Route the bulk of your traffic today to the cheapest model that clears your quality bar — for most workloads a mid-tier model like Claude Sonnet 5 or Gemini 3.5 Flash — and reserve a flagship like Opus 4.8 for the hardest requests, so that when GPT-5.6 reaches general availability you can drop it into a narrow slice of traffic and measure it against your incumbents instead of re-plumbing everything. Keep a benchmark set of your own real prompts, and convert any pricing into a concrete monthly number with the token-to-USD tool before migrating. For the direct head-to-head against Claude's top tier, see GPT-5.6 Sol vs Claude Fable 5 and Opus 4.8 on max effort. One note on the bigger picture: GPT-5.6 is a point release; a generational GPT-6 is a separate, later step, so plan your architecture around swappability rather than waiting for a whole-number jump.