When you buy through affiliate links in our content, we may earn a commission at no extra cost to you. Learn 
OpenAI made its GPT-5.6 family available to the general public on July 9, 2026, but getting there wasn’t easy. The new version purports to run faster and cost less than the one it replaces. However, it also marks a turning point in AI security, as the first frontier model release to be held back pending approval by the U.S. government — specifically, the U.S. Department of Commerce in coordination with the White House.
GPT-5.6 arrives as a trio, with Sol as the flagship. Terra serves as the balanced everyday model, while Luna is the cheap, fast option. OpenAI has decoupled the generation number from the capability tier. The digit tells users which generation they are on, while Sol, Terra, and Luna are meant to be durable names that advance on their own schedules.
It is a sensible fix for a version numbering system whose idiosyncrasies had become something of a running joke.
A Launch Focused on Intelligence Per Dollar
The rollout began globally on launch day, and OpenAI expected it to finish within 24 hours. In ChatGPT, Plus, Pro, Business, and Enterprise users get Sol at medium effort and above. Free and Go users get Terra inside ChatGPT Work and Codex. OpenAI plans to retire GPT-5.4 on July 23.
OpenAI is pitching the release around a metric of “intelligence per dollar.”
Its press release says GPT-5.6 will offer “more successful work for the same spend, or comparable results at a lower total cost,” and the launch post leans heavily into token efficiency. OpenAI says that, on the Artificial Analysis Coding Agent Index, Sol at max reasoning scores 80 while using less than half the output tokens of its nearest rival and finishing in less than half the time.
There are two new dials. Max gives the model more time to reason, check itself, and review its work. Ultra goes further by coordinating four agents in parallel by default across separate workstreams for hard problems.
ChatGPT Work is OpenAI’s Answer to Claude Coworker
The model shipped the same day as ChatGPT Work, an agent that takes a goal, gathers context across your apps, breaks the job into steps, and hands back finished artifacts rather than a wall of text. Those artifacts include spreadsheets, slide decks, live web apps, and documents.
A Unified Plugins Directory connects tools such as Google Drive, Slack, Teams, Gmail, Outlook, SharePoint, and Canva. OpenAI is also merging the Codex app into a new ChatGPT desktop app, putting Chat, Work, and Codex under one roof on every plan, including Free. The company has also started sunsetting the standalone Atlas browser and moving its agentic browsing flow into ChatGPT’s own built-in browser.
Observers noticed the resemblance immediately. ChatGPT Work functions as OpenAI’s answer to Anthropic’s Claude Coworker. The German outlet The Decoder was blunt about it, calling the launch largely a rebrand of Codex for knowledge work and noting that Anthropic had pushed Coworker to web and mobile shortly beforehand in what looked like a pre-emptive strike.
OpenAI’s New GPT-Live Voice Model
Separately, and a day earlier, OpenAI shipped GPT-Live, a new voice model now rolling out to all ChatGPT users. Its key feature is duplex audio, which lets it listen and speak at the same time. That means users can interrupt it, and it will adapt to the new instructions mid-sentence.
TechRadar’s David Nield, who spent hours with it, found that it sounds more convincingly human than before, even in the way it hesitates, and that real-time translation works well. OpenAI has since patched some early bugs, including unreliable memory in voice sessions and foreign words sometimes emerging with an English accent.
The AI Model Arms Race Now Runs Through Washington
Both American frontier labs saw the state restrict access to their best models in June 2026. Anthropic launched Mythos 5 and Fable 5 on June 9. Three days later, the Commerce Department issued an export control directive under the Export Controls Reform Act, barring access for foreign nationals anywhere. That included Anthropic’s own foreign national employees.
Anthropic could not enforce the directive in real time, so it pulled both models globally. The trigger was a jailbreak found by Amazon researchers in which Fable identified software flaws and, once, went further by demonstrating exploit code. Anthropic disputed the severity, noting that researchers could reproduce the same behavior on Opus 4.8, GPT-5.5, and Moonshot’s Kimi K2.7, and that the work counted as routine defensive security research.
Fable 5 returned globally on July 1 after Anthropic deployed a retrained classifier. Commerce’s own CAISI office verified that the classifier blocked the technique in more than 99% of attempts.
OpenAI faced a different process but the same pressure. Commerce Secretary Howard Lutnick reportedly called Sam Altman to warn against releasing GPT-5.6 publicly without prior approval. OpenAI complied, shipping the model to a small vetted group on June 26 with the intention of approving access customer by customer, according to an internal memo cited by The Information. Washington cleared the general release 12 days later.
OpenAI’s own launch post said the company does not believe this type of access process should become the long-term default. It argued that the approach withholds the best tools from the defenders and developers who need them.
