OpenAI has unveiled its next generation of AI, GPT 5.6, which consists of three models: Sol as the flagship, Terra for everyday use, and Luna as an affordable, fast option. The highlight: According to OpenAI, the models deliver more performance with noticeably lower token consumption. The catch: For now, the launch is limited to a restricted preview for select partners because the U.S. government has a say in the matter.
TL;DR – The Most Important Points in Brief
-
- GPT-5.6 comprises three models: Sol (flagship), Terra (all-rounder), and Luna (affordable, fast).
- Advances are particularly evident in agent-based programming, biology, and cybersecurity.
- For cybersecurity tasks, Sol matches the performance of Anthropic’s Mythos Preview—using only about one-third of the output tokens.
- Launching as a limited preview via API and Codex, following consultation with the U.S. government.
- Prices per 1 million tokens: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 (input/output).
What’s behind GPT-5.6?
Instead of releasing a single new model, OpenAI is dividing the generation into three tiers. Going forward, the number in the name will denote the generation, while the names Sol, Terra, and Luna represent permanent performance tiers that can evolve independently of one another. With this approach, OpenAI aims to give users a clearer choice between intelligence, speed, and cost.
Sol is the most powerful model in the series and is designed for demanding reasoning, research, and coding. Terra is aimed at everyday use and, according to OpenAI, is expected to match the performance of its predecessor, GPT-5.5, while being half the price. Luna offers the essential performance at the lowest price in the family. Also new are two operating modes: a max-reasoning mode, which gives Sol more time to think, and an ultra mode, which accelerates complex tasks via subagents.
GPT-5.6 at a Glance: Models and Prices
The three models differ significantly in performance and API costs. Here’s a comparison of the key data:
| Model | Role | Input (per 1 million tokens) | Output (per 1 million tokens) |
|---|---|---|---|
| Sol | Flagship for reasoning, research, and coding | $5 | $30 |
| Terra | Well-balanced model for everyday use | $2.50 | $15 |
| Luna | Fast and affordable | $1 | $6 |
In addition, GPT-5.6 introduces more predictable prompt caching, including explicit cache breakpoints and a minimum cache lifetime of 30 minutes. Cache writes are charged at 1.25 times the regular input price, while cache reads retain the 90 percent discount on the input rate. In July, OpenAI also plans to offer Sol on Cerebras hardware at up to 750 tokens per second—initially only for select customers.
More Performance, Fewer Tokens: The Benchmarks
OpenAI substantiates this efficiency gain with several benchmarks. In coding, Sol sets a new record on Terminal-Bench 2.1 —a test that evaluates command-line workflows requiring planning, iteration, and the interaction of tools.
In the field of biology, the GeneBench v1 benchmark is used, which evaluates time-consuming genomics and quantitative analyses. Here, Sol achieves better results than GPT-5.5 while using fewer tokens. The leap in performance is most evident in cybersecurity: On ExploitBench, Sol matches Anthropic’s Mythos Preview but does so with about one-third of the output tokens. On the ExploitGym benchmark—co-developed by UC Berkeley researchers—Sol, Terra, and Luna show noticeable improvement the more reasoning they’re allowed to perform.
Importantly, according to OpenAI, GPT-5.6 is better at finding and fixing vulnerabilities than at carrying out complete attacks from start to finish. In tests with Chromium and Firefox, the model did find bugs and individual components for exploits, but it did not construct a fully functional exploit under the test conditions. The model therefore does not meet the “Cyber-Critical” threshold of OpenAI’s in-house Preparedness Framework.
Why the launch is limited
The most exciting part this time is less the technology than the distribution channel. OpenAI is initially releasing GPT-5.6 only as a limited preview via API and Codex—for a small group of trusted partners and organizations. Prior to the launch, the company presented the models and their capabilities to the U.S. government; at the government’s request, the rollout is beginning with this limited group, whose participants have been reported to the government.
OpenAI makes no secret of the fact that it is not happy with this process. It says that such a government approval process should not become the norm, because it keeps the best tools out of the hands of developers, companies, and cybersecurity defenders. They are taking this short-term step only to achieve broader availability more quickly, while working with the government on a repeatable framework for future model releases. General availability for ChatGPT, Codex, and the API is planned “in the coming weeks.” This is reminiscent of the situation with OpenAI GPT-5.5, which was also rolled out in phases—though without this form of regulatory oversight.
Greater Protection Against Misuse
Alongside these enhanced capabilities, OpenAI has expanded its safeguards. The company refers to this as its most robust safety stack to date and relies on multiple layers: protective mechanisms trained directly into the model, real-time classifiers during response generation, account-level checks, and tiered access.
Specifically, GPT-5.6 is designed to deny prohibited cyber assistance even if users attempt to disguise their intent or trick the model via a jailbreak. If the real-time classifiers detect a potential violation in risky queries, response generation can be paused while a larger reasoning model examines the context. If the response is deemed inadmissible, the user never sees it in the first place. To identify vulnerabilities in its own security measures, OpenAI reports having invested over 700,000 A100-equivalent GPU-hours in automated red-teaming—supplemented by external human testers.
OpenAI openly acknowledges that these safeguards can also block or delay legitimate work, especially during the preview phase—for example, in cases where offensive and defensive tasks initially appear similar. That is precisely what the preview is designed to test.
Context: Faster Pace, New Rules of the Game
OpenAI continues to ramp up its release pace. Only a few months separate ChatGPT-5.4, GPT-5.5, and now GPT-5.6. In terms of content, the direction is clear: greater autonomy, better agent capabilities, and significantly more efficient token usage—a tangible cost advantage, especially in professional applications.
What’s new, above all, is the political dimension. This is the first time an AI lab has released its flagship model only after consulting with the government and exclusively to a select group. How quickly it becomes widely available will likely depend on how swiftly the announced regulatory framework is put in place. In any case, the competition with models from Anthropic and budget-friendly challengers like DeepSeek remains exciting.
Conclusion
With GPT 5.6, OpenAI delivers a technically impressive generation: three clearly distinct models, measurable progress in coding, biology, and cybersecurity, and efficiency gains that directly translate into lower costs. However, the limited launch under regulatory oversight clouds the picture—those who want to use Sol, Terra, or Luna will have to be patient for now. Once general availability is reached, at the latest, it will become clear whether the efficiency promises hold up in everyday use.
Sources: OpenAI
