
Tech • AI • Robotics • Game
OpenAI launched GPT-6.2 Astra and GPT-6.2 Soul on November 17, pitching Astra for demanding professional work and Soul for cheaper day-to-day tasks, coding and agents. On Agents Last Exam, Astra scored 78.3%, up 5.2 points from the prior Astra and roughly 5.5 points ahead of Claude Fable 5.2 in OpenAI's comparison. Astra is priced at $8 per million input tokens and $40 per million output tokens, about 20% below the previous flagship, while Soul stays at $2 in and $10 out. The central claim is not just lower API rates but fewer tokens used overall because the models plan better and backtrack less.
Early reaction to GPT-6.2 was notably hostile despite the headline benchmark gains. Developers reported models that were harder to steer, prone to stopping early, and inconsistent on writing, coding and creative execution. The dispute underscored a widening gap between leaderboard performance and real-world reliability, especially in agentic tasks where incomplete execution matters more than raw scores. Some practical tests still found value in Soul 6.2 for autonomous computer use, suggesting the release may be more workflow-dependent than the initial backlash implied.
In side-by-side AI video editing tests, GPT-6.2 Astra was markedly faster and cheaper than Fable 5.2 but generally lost on output quality. Fable delivered cleaner framing, stronger typography, better motion graphics and more polished visual judgment in short-form edits. Astra was considered usable and cost-efficient, yet more likely to stall or produce visually weaker decisions. The result reinforced a recurring theme of the week: lower cost and better benchmarks do not automatically translate into the best end product.
Google introduced Gemini 4 Neon just weeks after Gemini 4 Carbon, positioning it between Hydrogen and Carbon for coding, office work and science. On the Artificial Analysis Intelligence Index, Neon scored 61, ahead of Carbon at 60 and Hydrogen at 54, with Google also citing roughly 65% on Cursor Bench and 39.0 on a SWE-bench style bug-fixing test. Pricing landed at $3 per million input tokens and $15 per million output tokens, undercutting several premium rivals. In one cited coding comparison, Neon reportedly completed a task for about $3.50, while Claude Opus 5.5 took similar time but cost $720.
Google also launched Gemini Task, a consumer AI agent built around delegation rather than chat. It is rolling out free in the United States, Canada, Australia, India, Japan and South Korea, while users outside those markets generally need Pro or Ultra, with the UK and EEA excluded. The service can triage inboxes, draft replies, schedule activities, build documents and add groceries to Instacart. A notable feature is native text-message access on iOS and Android, making the product feel closer to a persistent assistant than a one-off chatbot.
Anthropic became the first major frontier AI lab to go public, with an indicated IPO range of $70 to $78 a share. But the market debut turned shaky, with the stock discussed around $87, roughly 10% lower on the day and only modestly above its $83 IPO price in one account. The listing put Dario Amodei's safety-first strategy under direct public-market scrutiny at a time when critics accuse the company of using catastrophic-risk arguments to support tighter control of advanced models. Wall Street's reaction suggested investors remain more focused on revenue, financing and compute dependence than on frontier prestige.
At Microsoft Ignite in San Francisco, Microsoft framed Copilot as the new front door for work rather than a feature tucked inside existing apps. Satya Nadella presented an enterprise stack organized around human ambition, intelligence and trust, pairing agents with governance, security and auditable data controls. The message targeted companies that have struggled to move beyond pilot projects because of data leakage fears, fragmented tooling and shadow IT. One of the clearest signals was what barely appeared: Windows was largely absent from the event's strategic center.
A separate thread of the day was rising concern over what AI and fragile infrastructure can enable in the wrong hands. Tests with weakly restricted models such as Heretic Qwen variants and Minimax M3 Flash showed how easily some systems could assist with exposed-secret discovery, including one demo that pulled 10 keys from 150 repositories. Meanwhile, internet infrastructure faced fresh pressure after an Optimizely compromise reportedly affected 178 major sites and a Java 17+ flaw exposed unauthenticated remote code execution risk through a Base64 decoding path. Together, the incidents sharpened the case that agentic capability and software security are now inseparable policy and product issues.