GPT-5.6 Sol, Terra, Luna: July 9 public release – all that is known at the time of release
Main chat
A chat for vibe coders: news, guides, live cases, marketplace, and finding executors.
On the evening of July 8, Sam Altman wrote in X: "GPT-5.6 Sol launches on Thursday." Happy development.” This confirmed what prediction markets have been waiting for the past three days: OpenAI announced the public launch of all three variants of GPT-5.6 — Sol, Luna and Terra — on Thursday, July 9. >
The history of the release of GPT-5.6 was atypical even by the standards of 2026, when state regulation of AI releases became the norm. Let’s take everything in order: the models themselves, prices, benchmarks and how events developed from June 26 to July 8.
Three models, one family
GPT-5.6 is not one model, but a family of three, and this is the largest change in the structure of OpenAI products in years. Each of the three occupies its own niche and is designed for different types of tasks and budget.
GPT-5.6 Sol - flagship
Sol is the strongest OpenAI model to date, built for deep reasoning and long-term agency work. Key benchmarks that OpenAI published in the preview system card:
Coding: Sol installs a new state-of-the-art on Terminal-Bench 2.1, a command-line test with tool planning, iteration and coordination. >
Biology: On GeneBench v1, which evaluates long-term genomic and quantitative-biological analysis, Sol performs better than GPT-5.5, using fewer tokens. >
Cybersecurity: Sol is OpenAI's most capable cybersecurity model. ExploitBench competes with Mythos Preview using only ~1/3 output tokens. This is a significant breakthrough in efficiency. >
Medicine: Length-adjusted HealthBench Professional: 60.5 (+8.7 to GPT-5.5). The biggest increase since GPT-5. HealthBench Hard: 33.1 (+1.6). >
GPT-5.6 Terra – Balance of price and quality
Terra is a balanced model for everyday performance, competitive GPT-5.5, at half the cost. That’s Terra’s key practical promise: If you’re using GPT-5.5 right now, Terra is cheaper.
Terra and Luna both significantly outperform the GPT-5.5, indicating a significant step forward in the price/quality ratio for more affordable models. >
GPT-5.6 Luna – speed and minimum cost
Luna is the fastest and most affordable model in the family. Positioned for high-volume tasks, where speed and cost are more important than the maximum quality of reasoning: classification, routine refactoring, documentation generation.
Prices: What it costs
Official pricing:>
| Модель | Входящие / 1M токенов | Исходящие / 1M токенов |
|---|---|---|
| Sol | $5 | $30 |
| Sol Fast (Cerebras) | $12,5 | $75 |
| Terra | $2,50 | $15 |
| Luna | $1 | $6 |
Prompt caching for the GPT-5.6 series has become more predictable: support for explicit cache breakpoints and a guaranteed minimum cache life of 30 minutes has been added. This is a direct response to complaints from developers who noted the unpredictability of caching in previous models.
Two new modes: max reasoning and ultra
With GPT-5.6, OpenAI introduces two fundamentally new modes, which were not in previous models.
Max reasoning effort
The new maximum level of effort in reasoning gives Sol maximum time for deep thinking. This is the development of the effort levels system, which already works in Codex and Claude – now OpenAI has a clear ceiling for the most demanding tasks.
Ultra mode
Ultra mode goes beyond the capabilities of a single agent, using subagents to accelerate complex work. It is a direct competitor to Missions in Factory.ai and Agent Sessions in Claude Code, where one task is decomposed into parallel subtasks, each solved by a separate agent instance.
Sol on Cerebras: 750 tokens per second
A separate track that is not directly linked to ChatGPT access:
GPT-5.6 Sol on Cerebras chips launches at speeds of up to 750 tokens per second in July - accessing frontline intelligence at an unprecedented rate. Access is initially limited to select customers as capacity expands. >
750 tokens per second is about 10-15 times faster than the standard speed of top-end models on standard infrastructure. For agent scenarios with long chains of steps, where the speed of inferencing is a bottleneck, this is a critical difference: a task that takes 10 minutes as standard will take less than a minute on Cerebras.
By comparison, Zhipu AI’s GLM-5.1 High-Speed, released in May, was 400 tokens/sec and was considered a record. 750 is the new bar.
History with state approval: June 26 to July 9
GPT-5.6’s path to public release was a standalone story – a direct continuation of the pattern set by Anthropic’s Fable 5 and Mythos 5 story.
OpenAI previously showed the U.S. government the plans and capabilities of the models before the announcement date. At the government's request, the company started with a limited preview for a small group of trusted partners whose involvement was reported to the government before releasing the models more widely. >
The reason for government attention is the same as with Fable 5: Sol is OpenAI's most capable cyber model. It shifts the performance frontier for long-term security challenges, including vulnerability research and exploitation. It is these opportunities that automatically attract state control.
The Department of Commerce, through its AI Standards and Innovation Unit, conducted additional tests. The company sent technical experts to Washington to respond to questions and concerns immediately. After additional testing and meetings with government agencies, the Trump administration gave OpenAI permission for a wider release. >
OpenAI has made no secret of its approach to such a procedure: "We do not believe that such a state approval process should become a long-term standard," the company said, adding that it complies with the requirement because it is the best way to ensure widespread public access to already created models. >
Community Reaction and Prediction Markets
While government approval was underway, the community monitored the situation through prediction markets.
On Kalshi, traders were actively betting on dates. “Now that Fable is unlocked by the government, I expect OpenAI to want to release 5.6 publicly sooner rather than later.” Another documented that people in X “are blatantly lying about the confirmed release date” and that “everything inside is under NDA.” >
Polymarket had July 9 as the leading outcome by July 7, followed by July 14. Earlier market activity boosted July 7 strongly, but the odds shifted when the official announcement that day did not follow. >
Noteworthy detail: in response to the expected launch of GPT-5.6 Anthropic on the same day announced the expansion of promotional access to Claude Fable 5. During the promotional period, users will be able to use Claude Fable 5 within up to 50% of the weekly subscription limit at no additional cost. Two companies release competing news on the same day - this has become virtually the new norm for the frontier.
What this means for users
** For developers on the API** - three clear options with clear positioning. Luna for high-volume routine tasks ($1/$6), Terra for most production cases ($2.5/$15), Sol for frontline tasks where maximum quality ($5/$30) is important. Explicit cache breakpoints make cost optimization easier.
**For ChatGPT users, GPT-5.6 is not yet available in the ChatGPT chat interface, only through the API and Codex for partners. GPT-5.6 is not available in ChatGPT during the preview period. From July 9, the situation should change.
** For vibcoding**, ultra mode with subagents and 750 token/sec Sol (Cerebras) is a fundamentally different iteration rate on long-term tasks. What used to take hours of agency work is getting much faster. The pair with Terminal-Bench 2.1 SOTA is especially relevant for claude-code scenarios – OpenAI is now clearly aiming for the same niche.
To choose between Anthropic and OpenAI, GPT-5.6 Sol competes with Mythos Preview for cybersecurity at ~1/3 output tokens. This is a direct blow to Anthropic's main trump card. Coincidentally, Anthropic expanded access to Fable 5 today, obviously not.
What is not known until July 9
A few questions that the official materials have not yet answered.
Exact ChatGPT access schedule: GPT-5.6 is not available in ChatGPT during the preview period. OpenAI plans to make the models widely available to ChatGPT, Codex and API users soon. The date inside ChatGPT for ordinary users has not been announced.
Availability for Russian users: Given the history of export controls around Sol’s cyber capabilities and the general pattern of restrictions, it is not obvious whether Sol will be available through APIs in Russia without additional restrictions.
Final benchmarks: OpenAI explicitly states that the extended set of evaluation results will be published when the model is widely available. Current data is a preview.
Outcome
GPT-5.6 is OpenAI’s biggest release this year. The three models cover the entire range from $1/million of input tokens (Luna) to $5/million (Sol). Ultra mode with subagents and 750 tokens/sec on Cerebras bring the speed of agent work to a new level. State approval passed - July 9 wide launch.
The totality of events of the last two weeks - Fable 5 is back, Sonnet 5 is out, GPT-5.6 Sol is out tomorrow - suggests that the companies have come out of the pause created by government regulation in June, and again switched to active model release. The next point of pressure is clear: who will be the first to widely roll out ultra mode / Missions / multi-agent scenarios for ordinary users, not just enterprise partners.
*Actual to July 8, 2026. Sources: OpenAI official announcement (openai.com), Engadget, Neowin, emergent.sh, Releasebot, Coursiv, Kalshi/Polymarket, OpenAI Help Center Release Notes, GPT-5.6 Preview System Card. The article will be updated following a public launch on July 9. *
