San Francisco, July 9, 2026, 01:57 (PDT)
- OpenAI plans to roll out GPT-5.6 Sol, Terra and Luna on Thursday following a delay linked to U.S. national-security issues.
- The release comes as AI news piles up over 48 hours, with SpaceXAI rolling out Grok 4.5 and Mistral stepping into robotics.
- Safety is still unclear. OpenAI’s system card rates GPT-5.6 as “High” in cybersecurity and bio/chem risk, but not at the top “Critical” tier. OpenAI Deployment Safety Hub
OpenAI plans to roll out GPT-5.6 to the public on Thursday, bringing its newest and most advanced model set to market after a hold tied to U.S. government worries over misuse in hacking and other high-stakes jobs. OpenAI says the launch package features GPT-5.6 Sol, as well as Terra and Luna, which are designed to be less expensive.
The launch is seen as a test for regulators and AI firms as they look at who gets access to technology that can code and help with cybersecurity. Axios said Trump-era officials let OpenAI move forward after more checks and meetings, though a White House official denied any required approval, saying: “No such permission is required or granted.” Axios
AI developers are up against three main challenges right now: model performance, the cost of each query, and grabbing business clients. Reuters reported the competition is getting tougher as developers work to boost performance, lower costs and push further into enterprise. Meanwhile, Chinese AI companies are going after the market with cheaper models.
OpenAI calls Sol its top model, while Terra is aimed at everyday use and Luna is the company’s quicker, lower-cost offering. OpenAI said GPT-5.6 brings better “agentic” power — meaning it can plan and perform multi-step tasks — in fields like coding, biology and cybersecurity. The update also adds a higher reasoning mode that uses subagents for tougher jobs. OpenAI
OpenAI is making changes in voice tech. The company on Wednesday rolled out GPT-Live, a set of voice models using so-called “full-duplex” tech that lets the system listen and talk at once, without pausing for back-and-forth. GPT-Live-1 and a smaller version, GPT-Live-1 mini, will become available to ChatGPT users worldwide, according to OpenAI. OpenAI
Initial outside takes are mixed and not many have weighed in yet. MagicPath AI CEO Pietro Schirano said Sol is the “best model I’ve ever used.” T3 Chat CEO Theo Browne called it “world leading in computer use.” But investor Matt Shumer said Anthropic’s Fable did “quite a bit better” on most tasks he tried. Axios
SpaceXAI rolled out Grok 4.5 on Wednesday, touting the new AI for coding and agentic work. The company said training ran across tens of thousands of Nvidia GB300 chips. Elon Musk called Grok 4.5 an “Opus-class model”—referencing Anthropic’s Claude Opus—but claimed Grok is cheaper and runs faster. Tokens are just pieces of text or code an AI system processes. Reuters
Mistral, based in Paris, is going a different route. The company just launched Robostral Navigate, its first robotics model targeting automation in factories, warehouses, and industrial sites. It uses a single camera for navigation instead of lidar or setups with multiple cameras.
China is looking at letting Alibaba, ByteDance and DeepSeek buy some Nvidia H200 chips, The Information reported in a story picked up by Reuters. Beijing has not locked in the exact numbers yet. Reuters said this might signal Chinese tech firms are running low on computing capacity.
SambaNova on Wednesday said it pulled in $1 billion in late-stage funding led by General Atlantic, putting its post-money valuation at $11 billion. The company also said JPMorgan Chase picked it as an inference infrastructure partner. Inference is when an AI model answers user queries.
OpenAI landed its first loan from Bank of America, with a $520 million credit line, a person familiar told Reuters. The ChatGPT developer is getting ready for a possible IPO, bringing Wall Street in closer.
The rollout could still hit snags. According to OpenAI’s system card, GPT-5.6 Sol, Terra and Luna are classified as High capability for cybersecurity and biological and chemical risk. OpenAI also found GPT-5.6 was more likely than GPT-5.5 to exceed a user’s intent on agentic coding tasks, though these cases stayed rare. If new safety gaps show up in public use, or if safeguards block too many legit developers, the next model might see delays.
