
Tech • AI • Robotics
OpenAI launched GPT-6 Astra on 3 September with a staged rollout across ChatGPT, Codex and the API. The flagship release is positioned as a step beyond the earlier GPT-5.6 line, with a sharper focus on long-horizon reasoning, tool use, browsing, software work and professional knowledge tasks. OpenAI is offering Astra and Astra Pro, pricing them at $10 per million input tokens and $50 per million output tokens. Early access began with selected organizations, with wider availability expected shortly after.
Nvidia moved to acquire Hugging Face for $12.93 billion, a deal that would reshape control of the open-source model ecosystem. The transaction stands out not just for size, but for what it says about the strategic value of developer distribution, model hosting and community influence. It also signals that infrastructure leaders no longer want to sit only underneath the AI stack. If completed, the purchase would tighten Nvidia's grip across chips, software and model discovery.
Anthropic unveiled Claude Fable 5.1 and Claude Mythos 5.1, claiming the strongest headline score of the week on the Artificial Analysis Intelligence Index. Fable 5.1 was reported at 66, ahead of Opus 5 at 63 and the earlier Fable at 62. The company also said improved caching cuts normal workload costs by 25% and long-horizon agentic jobs by 45%. That mix of higher benchmark performance and lower operating cost is aimed squarely at enterprise buyers.
The most consequential part of Astra may be its stronger computer use rather than any single benchmark. OpenAI showed the model operating software and browsers through natural instructions, handling screenshots, navigating interfaces and continuing multi-step work with less supervision. Demonstrations ranged from developer QA and routine office workflows to creative editing inside apps such as Krita. The broader message is that frontier models are being sold less as chatbots and more as operational co-workers.
OpenAI reported a near-saturating 99.9% score for Astra on ARC AGI 3 using its own provider adapter harness, versus 7.8% for GPT 5.6 Soul. But ARC Prize cited a much lower figure of roughly 63% under a different setup, exposing how much evaluation now depends on methodology. OpenAI argued the harness settings better reflect real-world performance rather than tuning specifically for ARC. Even with that dispute, reports that Astra builds compact symbolic representations in novel environments added to claims of a meaningful capability jump.
Astra's most talked-about demos came in Blender, where it built detailed 3D scenes from minimal prompts. Examples included recreating San Francisco's Palace of Fine Arts, generating real-estate interiors and preparing assets that could move into Unreal Engine 5 workflows. Those showcases shifted attention from abstract scores toward practical compression of many-hour creative pipelines. They also reinforced a broader industry change: users increasingly judge models by what they can ship inside existing software.
Google introduced Gemini 3.8 Flash along with a cybersecurity-focused variant, while Meta released Muse Spark 1.3. The timing, clustered with launches from OpenAI and Anthropic, underscored how quickly major labs are iterating after a quieter summer stretch. The emphasis was split between frontier capability and efficiency-oriented deployment. At the same time, new limits on Gemini Notebook highlighted the cost pressure behind advanced AI features.
France's tax administration confirmed a breach affecting data from 678,437 people after two intrusions in late June and late July. The stolen files reportedly covered 392,867 individuals, 285,570 professionals and 386 people with assets above €1 million. Cyber specialists said the intrusion appears to have stemmed from a compromised legitimate account rather than an exotic attack chain. The case sharpened concerns over cyber hygiene and identity protection as governments centralize ever more sensitive records.