Just now, the “cosmic level” GPT-5.6 launched a surprise attack, and Mythos 5 was unlocked

Just now, OpenAI released a limited preview version of its most powerful model to date – GPT-5.6: the flagship version Sol (Sun); Terra (Earth), a balanced model suitable for daily work; And the fast and affordable Luna model.

OpenAI co-founder and CEO Sam Altman posted on social media X that Sol is priced the same as GPT-5.5 and has stronger performance; Terra’s performance is comparable to GPT-5.5, but its price is only half of it. However, due to government censorship in the United States, the model is only available in limited preview form to some companies today, and it is working with the government to strive for full release in the coming weeks.

▲ Sam Altman’s tweet

From the evaluation results, GPT-5.6 Sol has strengthened programming, biology, cybersecurity, and other abilities. It leads the Claude Fable 5 in programming testing for Terminal Punch 2.1, with flagship and unreleased Ultra versions surpassing Claude Mythos 5 in evaluation; Faced with long-term security tasks, it only consumes one-third of the output token to benchmark Claude Mythos Preview on ExploitBench.

In terms of pricing, Sol’s input price is $5 (approximately RMB 34) and output price is $30 (approximately RMB 204) per million tokens; The input price of Terra is $2.5 (approximately RMB 17), and the output price is $15 (approximately RMB 102); The input price of Luna is $1 (approximately RMB 6.8), and the output price is $6 (approximately RMB 41).

By comparison, the prices of Fable 5 and Mythos 5 are the same: an input price of $10 (approximately RMB 68) and an output price of $50 (approximately RMB 339), which is approximately twice the price of GPT-5.6 Sol; Claude Mythos Preview (invited beta) costs $25 (approximately RMB 170) for input and $125 (approximately RMB 850) for output.

GPT-5.6 also introduces a more predictable hint caching mechanism, including support for explicit cache breakpoints and a minimum cache validity period of 30 minutes. For GPT-5.6 and higher models, cache write fees are charged at 1.25 times the uncached input fee, while cache read fees continue to enjoy a 90% discount on cache input fees.

OpenAI claims that the GPT-5.6 Sol version is equipped with OpenAI’s most powerful security protection system to date. OpenAI has strengthened its protection against high-risk activities, sensitive network requests, and repeated abuse behavior, and spent weeks searching for vulnerabilities and conducting stress tests on the system to enable it to withstand real-world attacks.

Alex Finn, founder and CEO of Henry Intelligent Machines PBC, a Silicon Valley AI startup, wrote in an article that unfortunately, “the era of large-scale release of cutting-edge models has come to an end… only a few people now have access to superintelligence. However, he believes that on the positive side, someone can balance the Fable 5. “GPT-5.6 surpasses Mythos in performance, but its price is only one-third of the latter.

▲ Alex Finn’s tweet

But Rohan Paul, a technology self media with 1.5 million followers on X, said that METR found that GPT-5.6 Sol cheated so many times in benchmark tests that its score became unstable. The cheating rate of 5.6 Sol is the highest value detected by METR in its publicly available ReAct Agent framework, which includes attempting to use evaluation settings instead of completing tasks normally.

▲ Rohan Paul tweet

On the same day, the US government relaxed its control over Anthropic models. According to foreign media Semafor today, the US government has just lifted the ban on the Claude Mythos 5 model. They have sent a letter to Anthropic notifying them that Mythos 5 can be opened for use by over 100 US institutions, and the listed entities do not need to obtain a license for exporting or transferring the model domestically. On June 13th, the US government implemented export controls on Mythos and Fable models, and Anthropic immediately suspended all users’ access to Fable 5 and Mythos 5. But Fable 5 was not mentioned in this unlocking.

▲ Screenshot of Semafor’s report by foreign media

01 .

GPT-5.6 Sol programming, biology, and safety greatly improved

Rated over Claude Fable 5

GPT-5.6 Sol is OpenAI’s most powerful model to date, enhancing agent capabilities in programming, biology, and network security; In addition, users can also view more security and readiness assessments in OpenAI’s system card.

In GPT-5.6, OpenAI introduced a new inference mechanism to give Sol more time for deep inference. In addition, OpenAI has introduced a new pattern that utilizes sub agents to accelerate the execution of complex tasks, surpassing the limitations of individual agents’ abilities.

For programming workflows, GPT-5.6 Sol performs outstandingly on Terminal Punch 2.1, which requires a command line workflow for planning, iteration, and tool coordination. The evaluation performance of GPT-5.6 Sol and Ultra versions exceeded Claude Mythos 5, while GPT-5.6 Terra surpassed Claude Fable 5.

GPT-5.6 Sol Evaluation on Terminal Punch 2.1

In terms of biological workflow, GPT-5.6 Sol achieved better results than GPT-5.5 in the GeneBench v1 test for evaluating long-term genomics and quantitative biological analysis with fewer markers.

GPT-5.6 Sol Evaluation on GeneBench v1

GPT-5.6 Sol is OpenAI’s most powerful network security model to date. It improves the performance efficiency of long-term security tasks. In the ExploitBench test, GPT-5.6 Sol only uses about one-third of the output tokens, which is comparable to Mythos Preview. In the ExploitPym test, GPT-5.6 Sol, Terra, and Luna models all achieved significant improvements in network capability with the improvement of inference ability.

GPT-5.6 Sol Evaluation on ExploitBench

GPT-5.6 Sol evaluation on ExploitPym

02 .

GPT-5.6 creates the strongest level of security protection

Adopting multiple layers of security measures

OpenAI emphasizes that it has equipped the three models in the GPT-5.6 series with the strongest level of security protection, with precise matching of level and capability. They mainly enhance the robustness of the model in real adversarial scenarios, while ensuring legitimate defense work such as code auditing, vulnerability research, and patch development. The strategy is to make prohibited attacks more difficult to implement, less predictable, and easier to trace without restricting their legitimate use. The evaluation shows that legitimate defense will significantly benefit, and illegal use will be effectively restricted.

Sol is better at assisting in vulnerability discovery and repair rather than end-to-end attacks. The primary task of OpenAI is to ensure that defenders benefit first. The framework evaluation shows that Sol has not reached the “critical” level, and can identify vulnerabilities and exploit primitives in Chromium and Firefox testing, but has not independently completed the complete attack chain. The so-called preparation framework is the process by which OpenAI uses advanced AI capabilities to track and respond to new risks that may pose serious harm. Due to limitations in benchmark testing, OpenAI has decided to upgrade the model while taking stricter measures and releasing it in stages.

In the GPT-5.6 preview version, OpenAI has implemented multiple layers of security measures. Model training includes built-in protection, real-time auditing, account level monitoring, and differentiated access control. The model has been trained and refused to be prohibited from assisting; Real time classifiers are evaluated during generation, high-risk pauses are reviewed by strong models, and violation outputs are intercepted; Account level review distinguishes between malicious and legitimate dual-use. Multilayer stacking makes the overall more robust.

During the preview period, users may encounter some security measures that may prevent or reject certain requests. OpenAI also collaborates with enterprise clients to develop longer-term solutions, including privacy protection checks and risk calibrated access permissions.

03 .

Invest 700000 A100 GPU hours

Conduct an automated red team drill

Security protection must remain effective against changes in attacker strategies. Protection against known attack methods alone is far from sufficient for cutting-edge models.

To this end, OpenAI has invested unprecedented intelligent computing power to ensure security, utilizing self-developed models to accelerate vulnerability discovery and protection upgrades. OpenAI has invested over 700000 A100 GPU computing hours in automating Red Team exercises to discover universal jailbreak methods. This investment also enables it to explore attack modes far beyond the coverage of manual testing, identify failure modes earlier, and shorten the path from vulnerability discovery to repair.

In addition to automated red team exercises, OpenAI has also collaborated with third-party testing agencies to conduct extensive human expert red team exercises, which will continue during the preview period. Artificial red team exercises are a supplement to automated exercises, used to test whether system defenses can resist abusive behavior implemented by creative experts in ways that AI systems cannot predict.

Due to the fact that no evaluation can exhaust all product configurations, multi-step attacks, or real workflows, OpenAI has established a rapid response process for reproducing, evaluating, grading, and fixing newly discovered jailbreak vulnerabilities, and incorporating them into an ongoing evaluation system to ensure effective testing of similar vulnerabilities in the future.

04 .

Conclusion: GPT-5.6 has three precise card positions

Attempting to stratify and squeeze opponent space

Through GPT-5.6, OpenAI has synchronously accelerated its model capability and security protection. Benchmarking in the three major fields of programming, biology, and network security has comprehensively surpassed Claude Fable 5, coupled with the precise positioning of Sol, Terra, and Luna, OpenAI is trying to squeeze the survival space of competitors with stronger performance and finer product layering.

After previewing, OpenAI plans to promote GPT-5.6 to a wider range of users using ChatGPT, Codex, and APIs in the coming weeks. OpenAI will also launch GPT-5.6 Sol on Cerebras in July, with a processing speed of up to 750 tokens per second, achieving a new high in speed.