China tests robots in traffic surveillance — The Nordic Times
OpenAI has begun a limited preview of GPT-5. 6, with the flagship model Sol and two smaller models intended to offer developers a range of options for intelligence, speed and cost. The rollout is being. . .
OpenAI has begun a limited preview of GPT-5. 6, with the flagship model Sol and two smaller models intended to offer developers a range of options for intelligence, speed and cost. The rollout is being conducted in coordination with the US government. The model family was presented by OpenAI on Friday and consists of GPT-5. 6 Sol, GPT-5. 6 Terra and GPT-5. 6 Luna. Sol is described as OpenAI's most powerful model to date. Terra is a more balanced option for everyday work, while Luna is a faster and cheaper model for large-volume use. OpenAI emphasizes that the release is only a limited preview. A smaller group of trusted partners and organizations will get access through the API and Codex, while broader access to ChatGPT, Codex and the API is promised "over the coming weeks". Introducing a limited preview of GPT-5. 6 Sol, our next generation frontier model, as well as GPT-5. 6 Terra, a balanced model for efficient, everyday work, and GPT-5. 6 Luna, a fast and affordable model for high-volume work. https://t. co/OoM83SyISN— OpenAI (@OpenAI)June 26, 2026The rollout confirms rumors of political friction that surfaced earlier this week. The company writes that the models and plans were shown to the US government before launch, and that the preview is being limited at the government's request. OpenAI emphasized that this process should not become the norm. Such a government access process, the company argues, keeps the best tools away from users, developers, companies, IT security experts and global partners who need them. Three models, three tiersThe new naming means that 5. 6 marks the model generation, while Sol, Terra and Luna are intended to be more durable capability tiers. Sol is the flagship, Terra the general-purpose work tool and Luna the fastest and most cost-effective variant. OpenAI says Terra reaches performance on par with GPT-5. 5 at half the price. Luna is said to offer strong capability at the lowest cost in the family. Pricing is listed per million tokens. Sol costs $5 per million input tokens and $30 per million output tokens; Terra is priced at $2. 50 and $15, respectively, and Luna at $1 and $6. For cached input, a 90 percent read discount still applies, while cache writes for GPT-5. 6 and later models are billed at 1. 25 times the model's standard input price. GPT-5. 6 also introduces a new reasoning level, "max", where Sol gets more time to reason deeply. OpenAI is also launching an "ultra" mode in which subagents are used to speed up complex workflows. Record performance in command-line tasksOpenAI highlights coding, biology and cybersecurity in particular. In TerminalBench 2. 1, which tests command-line work where AI systems must plan, iterate and coordinate tools, GPT-5. 6 Sol Ultra topped the benchmark with 91. 9 percent. The standard version of Sol is listed at 88. 8 percent. GPT‑5. 6 Sol sets a new state of the art on Terminal‑Bench 2. 1, which tests complex command-line workflows requiring planning, iteration, and tool coordination. pic. twitter. com/gWA1pzpOmK— OpenAI (@OpenAI)June 26, 2026The figures are based on OpenAI’s own benchmarks, but show how the company wants to position GPT-5. 6: not merely as a chatbot, but as a system for more extensive agentic tasks where the model controls tools, interprets results and keeps going until the task is solved. OpenAI also reported clear improvements in biological research capabilities. On GeneBench v1, which measures longer-running workflows in genomics and quantitative biology, Sol is said to perform better than GPT-5. 5 with fewer tokens. Cybersecurity at the centerThe most politically sensitive aspect of the release concerns cybersecurity. OpenAI describes GPT-5. 6 Sol as its most advanced model to date for security work, where it can help with vulnerability analysis, development of security fixes, code review and defensive testing. At the same time, the company says Sol is better at finding and fixing vulnerabilities than at reliably carrying out complete attacks from start to finish. In the system documentation, GPT-5. 6 isclassifiedas high capability in cybersecurity and biology/chemistry, but not critical in any of the tracked risk categories. OpenAI writes that Sol, in tests against Chromium and Firefox, could identify bugs and parts of exploit chains, but could not independently create a working end-to-end exploit chain under the tested conditions. This distinction matters. The same capabilities that can help defenders find weaknesses in critical infrastructure can also be weaponized to scale up attacks. OpenAI is therefore trying to present the release as both a leap in capability and a security test. Several layers of safeguardsOpenAI describes GPT-5. 6 as its most closely monitored release so far. According to the company, the safety architecture consists of model training, real-time classifications, account review, differentiated access, monitoring, enforcement, mitigations and continued testing. At higher risk levels, the model's generation can be paused while a larger reasoning model reviews the conversation and context. If the response violates policy, it is blocked before reaching the user. OpenAI also says flagged activity can lead to account review across multiple conversations. The aim is to distinguish normal security research from more systematic attempts to use the model for prohibited activity. The company writes that legitimate users may see some delays or false blocks during the preview period, particularly in areas where defensive and offensive activity look technically similar. Automated attacks against the safeguardsA central part of the safety work is automated red teaming. OpenAI says more than 700,000 GPU hours, equivalent to A100 capacity, were spent automatically searching for universal jailbreaks – attacks that work across many types of questions and contexts. According to the company, such tests are intended to complement human experts and third-party testers. The goal is to find general weaknesses in the safeguards, close them and then add similar patterns to ongoing evaluations. At the same time, OpenAI acknowledges that no test can cover every product configuration, multistep attack or real-world workflow. The company is therefore maintaining a rapid process for reproducing, prioritizing and fixing new jailbreaks during the preview period. Government oversight remains centralThe launch comes only days after reports that the Trump administration wanted OpenAI to release GPT-5. 6 gradually for security reasons. OpenAI’s announcement now confirms the substance of those reports: the models were previewed for the government and access begins with a group of partners whose participation has been disclosed to the government. The company describes the measure as temporary and linked to work on the president's executive order on cybersecurity. In practice, however, the arrangement means Washington gets insight into which actors first gain access to the strongest commercial AI model. This is not a formal licensing system. OpenAI says the goal is broad access. Even so, the release shows how frontier models, the most advanced AI systems, have quickly moved from product launches into security policy. OpenAI also plans to make GPT-5. 6 Sol available on Cerebras in July, with up to 750 tokens per second. That access, too, will initially be limited to selected customers as capacity is expanded. The question is therefore not only how smart GPT-5. 6 is, but who gets to use it first – and on what terms.