OpenAI Lifts GPT-5.6 Preview Restrictions Ahead of Thursday's Full Launch
Hello everyone, I'm cxuan.
GPT-5.6 has finally been officially unblocked! I've been waiting so long my kid is almost old enough to run errands.
Following Company A's unblocking of Fable 5, GPT-5.6 has also finally received its unblocking news.
Just a few dozen minutes ago, OpenAI officially announced: GPT-5.6 Sol, along with Terra and Luna, will be publicly launched this Thursday, and they are now expanding preview access globally.
Many people, myself included, have been waiting for this news.
I wrote an article on June 27th specifically introducing the capabilities of GPT-5.6: 《Better Late Than Never! GPT-5.6 Major Release!》.
At that time, GPT-5.6 had already been released, but due to policy issues, it was only available for limited preview to a small number of organizations; ordinary users couldn't use it at all.
Now this restriction is finally starting to lift. Thank goodness Belgium knocked out the US team! The dawn has arrived!
But don't rush to open ChatGPT and look for the model just yet. OpenAI said this time they are expanding preview access globally, but they haven't explicitly stated that all ChatGPT users can use it directly.
Whether it will open up to the API first, Codex, or go directly into the ChatGPT model list, we'll have to wait for the official rollout on Thursday.
Two weeks have passed, and I guess everyone has forgotten the capabilities of GPT-5.6, so let's do a quick review to prepare for the big test tomorrow.
GPT-5.6 Sol is currently OpenAI's most powerful model; GPT-5.6 Terra is a more balanced model; GPT-5.6 Luna is a cheap, fast, and most durable model.
Image source: OpenAI
Terra's performance is basically on par with GPT-5.5, but at half the price. Luna's capability is not as good as GPT-5.5, but its advantage lies in lower cost and faster speed.
To differentiate in one sentence: Sol is for peak intelligence, Terra is for daily cost-effectiveness, Luna is for throughput and cost.
OpenAI didn't release all the evaluations at once back then. The original text said that the full evaluation would be released when it became more widely available.
It first released three directions: Coding, Biology, Cybersecurity.
In coding, GPT-5.6 Sol refreshed OpenAI's own performance on TerminalBench 2.1. This benchmark tests command-line workflows, requiring planning, iteration, and tool coordination abilities.
Image source: OpenAI
On TerminalBench 2.1, both effort tiers of GPT-5.6 Sol surpassed Mythos 5. Terra, the cost-effective version, even managed to surpass Fable 5.
GPT-5.6 also added two new capability entry points: max reasoning effort and ultra mode.
max means giving Sol more time for deep reasoning.
ultra is more like letting the model call sub-agents, splitting complex tasks into multiple sub-tasks to run concurrently.
OpenAI's phrasing is that it exceeds the capability boundary of a single agent.
In the biology direction, OpenAI mentioned GeneBench v1. It is used to evaluate long-cycle genomics and quantitative biological analysis tasks.
OpenAI's statement is that GPT-5.6 Sol is stronger than GPT-5.5 and uses fewer tokens.
Image source: OpenAI
In terms of cybersecurity, the statement was even more blunt.
OpenAI said that GPT-5.6 Sol is currently their model with the strongest cybersecurity capabilities. It has made significant progress in performance and efficiency on long-cycle security tasks such as vulnerability research and exploitation.
On the ExploitBench leaderboard, GPT-5.6 Sol used about 1/3 of the tokens and approached the level of Mythos Preview.
Image source: OpenAI
On ExploitGym, Sol, Terra, and Luna all demonstrated clear improvements in cybersecurity capabilities as reasoning was enhanced.
Image source: OpenAI
This needs to be clarified for everyone.
The original text and System Card both emphasize that Sol is mainly used to help cyber defenders discover attack vulnerabilities, develop patches, and strengthen system protection.
Under Chromium and Firefox test conditions, it did not produce a complete, autonomously runnable attack chain. According to OpenAI's current framework settings, it has not yet crossed the Cyber Critical threshold.
But the capability is already strong enough that it must be released in phases.
So after this official announcement came out, my first reaction was simple: it's finally our turn to test it ourselves.
Official benchmarks have already released quite a few, but whether GPT-5.6 can really deliver still needs hands-on experience. Especially Sol's max and ultra, I really want to see how far it can go in real coding tasks.
See you all tomorrow.
Reference links: