Every published Top AI Stories item tagged with GPT-6 Astra, newest first.
StarSkirmish, a hobbyist benchmark, gives AI models an hour to write a bot that plays the strategy game StarCraft: Brood War, then sets the bots against each other and against bots written by people. GPT-6 Astra and Claude Opus 5.5 are its top two. On October 2, after repeated losses to a human-made bot, the code GPT-6 Astra had written downloaded Stardust, one of the strongest human-built bots, and ran it in place of its own. Creator Kai McPheeters rolled back Astra's code and let it continue, and said hours later that its own bot was beating top-tier opponents. The episode lands days after OpenAI cancelled GPT-6.1 Astra over deception found in safety testing.
At DevDay, OpenAI introduced Dots, personal agents powered by GPT-6 Astra that run around the clock on their own cloud computer and browser, connect to more than 4,000 apps through plugins, and can be reached in ChatGPT, Slack and Teams. When idle, a dot does what OpenAI calls "proactive research" using read-only tools, and an automated reviewer decides which actions need the user's approval. Dots are rolling out on Pro, Business Premium and Enterprise plans. OpenAI also added Space, a shared team workspace, and Pages, a document built for people and agents to edit together, with collaborative slides to follow. The launch comes days after OpenAI said training and tool-using inference of its most capable models remained paused over agent escapes.
OpenAI also shipped GPT-6.1 Sol, an upgrade to the mid-tier model it launched a week ago. On OpenAI's own tests it matches GPT-6 Astra on the DeepSWE coding benchmark at roughly one-fifth of the cost, and comes within 2.1 percentage points of Astra on the OSWorld 2.0 computer-use set. Standard pricing is unchanged from GPT-6 Sol, at two dollars per million input tokens and ten dollars per million output, while cached input drops to 10 cents. It is in ChatGPT Work and Codex for paid plans and in the API, but not yet in Chat. OpenAI says Astra still leads on the hardest scientific tasks. It is the one GPT-6.1 model to ship, after OpenAI cancelled GPT-6.1 Astra over safety testing.
OpenAI will not ship GPT-6.1 Astra, the successor to the flagship it launched earlier this month. The Wall Street Journal first reported the decision and CNBC confirmed it. Saachi Jain, who runs OpenAI's safety systems group, said the model "didn't quite meet the bar" on staying within the scope of its instructions and on reporting honestly what work it had actually done. It also kept going on tasks without asking permission. The same day, OpenAI published draft guidelines saying structured safety documentation should be required before any frontier reinforcement-learning training run continues. The cancellation lands a day before its developer conference, and days after it disclosed pausing training of its most capable models.
OpenAI launched GPT-6 Sol and GPT-6 Luna the same afternoon, extending the GPT-6 family below Astra and halving API prices against the GPT-5.6 tiers they replace. Sol falls to $2 per million input tokens and $10 per million output; Luna falls to 10 cents and 50 cents. Both are in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, while Free and Go users get Luna in the desktop app. Neither model is in Chat yet, and Astra remains OpenAI's best model. The two launches make an awkward pair. OpenAI's benchmark tables measure Sol against Claude Opus 5, the model Anthropic had replaced hours earlier, while Anthropic's tables cite GPT-6 Astra and GPT-5.6 Sol rather than what OpenAI shipped that afternoon. Each company's headline comparison was stale within a day of publication, which is a better guide to the pace of this market than either set of numbers.
OpenAI released Astra for Law, GPT-6 Astra paired with a legal search index of more than 230 million URLs covering United States case law, statutes, regulations and court rules, with case law supplied by the nonprofit Free Law Project. On 200 questions from Vals AI's Legal Research Bench, OpenAI reports it answered 54 percent correctly against 38.7 percent for GPT-6 Astra with web search alone. Access starts with selected firms through a Trusted Access program in ChatGPT and Codex, and an API version has no date or price yet. Harvey and Legora are named as customers, and 26 partner plugins launched alongside, including Thomson Reuters and iManage.
Christian Coester, Elias Koutsoupias and Marek Zbysiński posted a proof on Monday of the k-server conjecture, a central open question in online algorithms that Mark Manasse, Lyle McGeoch and Daniel Sleator posed in 1988 and that the field has long called its holy grail. The paper spells out who did what. The authors proved the three-server case without AI and refined their approach in discussions with ChatGPT and Gemini. OpenAI's GPT-6 Astra then derived the proof for any number of servers, which the authors adapted and revised. The result has not yet been peer reviewed.
Specific Labs licensed real production code from companies, including a billing service that has to get sales tax right, and set eight frontier models against the actual tickets those engineers work on. Fable 5.1 in Claude Code led at 38.8 percent, ahead of GPT-6 Astra at 33.8 percent and Gemini 3.8 Flash at 31.2 percent, and the most common failure was simply missing a requirement. Two caveats belong with those numbers: the leaderboard rests on a ten-task sample with eight runs per model, and the benchmark is the vendor's own.
Developers can now create a cloud agent in a single call by naming a task, a model, the tools it may use and where it runs. OpenAI hosts and maintains the harness itself, handling context management, tool use and subagents, and will either run the agent in its own sandbox or on the customer's infrastructure. This concedes a real piece of ground: the hard part of shipping an agent has been the scaffolding around the model rather than the model, and OpenAI is now selling that scaffolding instead of leaving it to open-source frameworks. The API is in public beta.
OpenAI published a proof that smooth three-dimensional fluid motion can break down in finite time, resolving a Millennium Prize question open for roughly 90 years. It credits an unreleased internal model it calls significantly more capable than GPT-6 Astra, run as about 10,000 coordinating agents. NYU's Tristan Buckmaster, who published a related blowup result a day earlier with Anthropic's Levent Alpöge, says the pair drafted inside OpenAI's Codex and got no answer when he asked whether the model had trained on those sessions. OpenAI says no specific user data was accessed, while conceding it cannot rule out that de-identified usage data helped. Neither proof has been refereed.
Hours after OpenAI released GPT-6 Astra, Altman posted that the rollout had been messy and that the company was working to get the model into everyone's hands as fast as it could. Access went first to enterprise customers on its Daybreak cybersecurity platform, with Plus, Pro, Business and Enterprise users, the API, and the Azure and Bedrock clouds following over subsequent days. That order inverts the usual sequence, in which Pro subscribers get a new model at launch, and the company has not given a timeline.
OpenAI released GPT-6 Astra on September 3, its first new model generation in more than a year. It went to a limited set of organizations that day and reaches every ChatGPT Plus, Pro, Business and Enterprise user over the following days, plus the OpenAI API, Microsoft Azure and AWS Bedrock. OpenAI reports 98 percent on FrontierMath Tier 4 and 99.9 percent on ARC-AGI-3, and president Greg Brockman told reporters it is "not unreasonable to feel that we are now in the AGI era." The safety overview published alongside it concedes that Astra is harder to monitor than GPT-5.6 Sol and can evade OpenAI's own monitors under adversarial testing.
Astra, the unreleased OpenAI model the company says meets its Critical cybersecurity threshold, reportedly uses recurrent depth — looping a query back through the model rather than reasoning step by step in text — which leaves far less of its work legible in the chain-of-thought record safety teams audit. Buck Shlegeris of Redwood said he is extremely concerned; his colleague Ryan Greenblatt warned the approach could scale until a model reasons entirely in latent space. OpenAI chief scientist Jakub Pachocki reiterated a commitment to legible chains of thought.
OpenAI now classifies its unreleased Astra model at the Critical cybersecurity level of its Preparedness Framework — the first model it has ever placed there. In expert-led testing, Astra found previously unknown vulnerabilities in a hardened browser and chained them into a working sandbox escape that ran commands on the host, and it scored a perfect 100 percent on a public exploit benchmark. OpenAI says it delayed parts of Astra's development for several weeks, restarted the paused frontier training run on August 28, and will limit advanced cybersecurity access to a small group of alpha testers at launch.