What Is GPT-5.6 Sol (the 3 Models)
How the three GPT-5.6 models are positioned
GPT-5.6 Sol is the most capable model from OpenAI, released as a limited preview on June 26, 2026. Terra and Luna were announced alongside it, giving three models that differ in performance and price for different uses. Let's look at where Sol sits, how it differs from Terra and Luna, and the pricing, in order.
What GPT-5.6 Sol Is (the flagship position)
GPT-5.6 Sol is OpenAI's strongest model to date. What the company put forward in this preview is a set of evaluations showing improved agentic capabilities (an AI deciding its own steps and carrying the work through) in three areas: coding, biology, and cybersecurity. In cyber specifically, it says Sol shifts the performance-efficiency frontier for long-horizon security tasks.
GPT-5.6 Sol is our strongest model yet. To give a preview of model performance, we share a set of evaluations highlighting improved agentic capabilities in coding, biology, and cybersecurity / GPT-5.6 Sol is our most capable model yet for cybersecurity. It shifts the performance-efficiency frontier for long-horizon security tasks including vulnerability research and exploitation. — from the description of Sol's position and the evaluation areas published
How Sol, Terra, and Luna Differ (price and speed)
The three models split their roles along the balance of "performance" versus "cost and speed." Terra keeps performance on par with the previous generation, GPT-5.5, while holding the price to about 2x cheaper; it suits workloads that run large volumes. Luna, the fastest and cheapest in the family, fits light, routine work handled at low cost. In short, the basic split is Sol for hard problems, Terra for high volume, and Luna for light tasks.
Terra, a balanced model for everyday work; and Luna, a fast and affordable model. Terra has competitive performance to GPT-5.5 while being 2x cheaper and Luna brings strong capability at our lowest cost. — from the description of how Terra and Luna are positioned
GPT-5.6 Pricing Compared (per 1M tokens)
Pricing is usage-based on the number of tokens, with separate rates for input and output. Terra and Luna were cut on July 30, 2026, so the launch prices no longer apply. The table below shows the current official per-1M-token rates (Standard, short context).
| Model | Input (/1M tokens) | Output (/1M tokens) | At launch (June 2026) | Best for |
|---|---|---|---|---|
| Sol | $5.00 | $30.00 | unchanged | Hard coding, security, long-horizon agents |
| Terra | $2.00 | $12.00 | $2.50 / $15 | GPT-5.5-class performance at high volume |
| Luna | $0.20 | $1.20 | $1 / $6 | Fast, cheap, light routine tasks |
Luna took the bigger cut—80% less in OpenAI's wording, or one-fifth of the launch price. Terra came down 20%, and top-end Sol was left unchanged. That widens the gap between Luna and Sol from 5x to 25x, so which model you assign light routine work to now moves the bill far more than it used to.
The official pricing page also lists rates for cached input, cache writes, and long context. The table above shows the most basic tier.
For a wider view of the major models, you may also find our ChatGPT (GPT-5) guide and What is Claude (Anthropic's generative AI) guide helpful.
Starting today, GPT-5.6 Luna, our fastest and most affordable model, will cost 80% less, while GPT-5.6 Terra, our balanced model for everyday work, will cost 20% less. / Starting July 30, API pricing is $2 per million input tokens and $12 per million output tokens for Terra, and $0.20 per million input tokens and $1.20 per million output tokens for Luna. Sol pricing remains unchanged. — from the description of the size of the cut and the new rates
gpt-5.6-sol $5.00 $0.50 $6.25 $30.00 / gpt-5.6-terra $2.00 $0.20 $2.50 $12.00 / gpt-5.6-luna $0.20 $0.02 $0.25 $1.20 — current rates for the three models (columns in order: Short context input / Cached input / Cache writes / Output, per 1M tokens)
GPT-5.6 is priced per 1M tokens across three model sizes: Sol is $5 input / $30 output; Terra is $2.50 input / $15 output; and Luna is $1 input / $6 output. — as the prices at launch (June 2026)
GPT-5.6's New Features and Performance
Two new ways of thinking
What stands out in GPT-5.6 is a pair of new modes that control how the model thinks. With the same model, you can choose to have it think more deeply, or split the work to move faster. Let's look at what each does, and the official benchmark results.
What max reasoning effort Is (thinking more deeply)
max reasoning effort raises the model's "depth of thinking" to the maximum, letting it spend more time reasoning on hard problems. By deepening a single chain of reasoning, it lets Sol work through tough problems carefully. It helps in complex analysis and design where accuracy matters more than speed.
With GPT-5.6, we're introducing a new max reasoning effort to give Sol the most time to reason deeply. — from the description of max reasoning effort
What ultra mode Is (subagents in parallel)
ultra mode takes the approach of "splitting" the work rather than "deepening" it. Instead of one model working alone, it hands complex work to several subagents (helper AI workers) and drives them together at high speed. The announcement does not name any specific areas where the gain is largest. From the standpoint of running AI on several tasks in parallel day to day, having the work split up often finishes faster than asking a single model to think for a long time, so the intent behind ultra mode's subagent approach is easy to follow.
Additionally, we're introducing a new ultra mode that goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work. — from the description of ultra mode (the announcement's only mention of it)
Benchmark Performance (Terminal-Bench, GeneBench)
Official benchmark results across three areas
OpenAI shared benchmark results across three areas: coding, biology, and cybersecurity. Sol set a new state of the art on Terminal-Bench 2.1, which measures command-line work requiring planning, iteration, and tool coordination. It is a result that backs up its strength in hard coding work.
There is progress beyond coding too. On GeneBench v1, which evaluates long-horizon genomics and quantitative-biology analyses, it achieved a higher score than GPT-5.5 while using fewer tokens. On the cybersecurity benchmark ExploitBench, it produced a comparable result while holding output tokens to about one-third. This efficiency—matching or beating with fewer tokens—is another feature of this release.
GPT-5.6 Sol sets a new state of the art on Terminal-Bench 2.1, which tests command-line workflows requiring planning, iteration, and tool coordination. / On GeneBench v1, which evaluates long-horizon genomics and quantitative-biology analyses, it achieves stronger results than GPT-5.5 while using fewer tokens. / …GPT-5.6 Sol is competitive with Mythos Preview using only ~1/3 of the output tokens. — from the results for Terminal-Bench 2.1, GeneBench v1, and ExploitBench (the third follows "On ExploitBench²,")
Can You Use GPT-5.6 Sol Now (availability)
GPT-5.6's current availability
For all its capability, GPT-5.6 is not yet open to everyone. Let's look at the current scope, the U.S. government's request behind it, and the outlook for general availability.
Current Availability (API, Codex, limited preview)
Right now GPT-5.6 is a limited preview that only a select group of trusted partners and organizations can use, through the API and Codex. During the preview, access is confined to the API and Codex, so it is not yet available from everyday ChatGPT. The reach is limited to a small group of partners and organizations.
During the preview, GPT-5.6 models will initially be available through the API and Codex to a select group of trusted partners and organizations. — from the description of preview access and who it covers
Limited Rollout at the U.S. Government's Request
Behind this limited rollout is a request from the U.S. government. OpenAI says it is starting with a limited preview for a small group of trusted partners whose participation has been shared with the government—a phased approach to releasing a highly capable model carefully. That said, OpenAI is clear that this kind of government access process should not become the long-term default.
For more on why the rollout is this cautious, you may also find our OpenAI Daybreak guide, focused on cyber defense, helpful.
At their request, we are starting with a limited preview for a small group of trusted partners whose participation has been shared with the government, before releasing more broadly. / We don't believe this kind of government access process should become the long-term default. — from the description of the limited rollout at the government's request and how OpenAI frames it
GA Outlook and Safety
OpenAI says it plans to make GPT-5.6 Sol, Terra, and Luna generally available within the coming weeks. After further testing, it intends to move to broad availability in ChatGPT, Codex, and the API. It also says the models ship with its most robust safety stack to date, including protections for higher-risk activity, sensitive cyber requests, and repeated misuse.
On the cyber side, OpenAI says Sol does not reach the top "Cyber Critical" threshold in its Preparedness Framework. In evaluations involving Chromium and Firefox, it identified real bugs and the building blocks used to assemble an exploit (attack code that abuses a vulnerability to make software behave in unintended ways), but under the conditions tested it did not autonomously produce a fully working end-to-end exploit. Pairing stronger capability with stronger safeguards is what shows in this phased rollout.
We believe in broad access, and we plan to make GPT-5.6 Sol, Terra, and Luna generally available in the coming weeks. / GPT-5.6 Sol launches with our most robust safety stack to date. We strengthened protections for higher-risk activity, sensitive cyber requests, and repeated misuse, and spent multiple weeks finding weaknesses, pressure-testing our system, and hardening it against real-world attacks. / GPT-5.6 Sol does not cross the Cyber Critical threshold under our Preparedness Framework. In evaluations involving Chromium and Firefox, it identified bugs and exploitation primitives—the building blocks of an exploit—but did not autonomously produce a functional full-chain exploit under the conditions tested. — from the GA plan, the safety stack, and the Stronger cyber capabilities section
GPT-5.6 Sol Summary
GPT-5.6 lines up three models—the top-end Sol, the cheaper Terra, and the fastest, cheapest Luna—to cover everything from hard problems to high volume and light tasks by use. With the new ultra mode and max reasoning effort, you can choose to think deeply or split the work to move faster, and it showed a high level on benchmarks such as Terminal-Bench 2.1. What you can reach today is limited to the API and Codex preview, but general availability is expected within the coming weeks, making this an update worth waiting for.
Official announcements are often long English documents, and there are times you want to grasp the key points quickly. Turning a web page into Markdown keeps the structure of headings and tables intact, which tends to improve the accuracy when you have an AI read and summarize it.
