OpenAI announces GPT-6 Astra for computer tasks

#1 𝕏

OpenAI announced GPT-6 Astra, stating it can quickly perform anything a user can do on a computer.

Also covered by: @OpenAI News, @How I AI Podcast, @claire vo 🖤, @Dan Shipper, @Randy Counsman, @Aravind Srinivas, @Peter Yang, @There's An AI For That, @There's An AI For That, @claire vo 🖤, @Sam Altman, @Nicholas Thompson

#2 𝕏

Hugging Face shared a link to NVIDIA’s blog titled “NVIDIA to Acquire Hugging Face.” The source provides no acquisition details, dates, terms, or named people.

Also covered by: @Mira Murati

#3 𝕏

Google AI announced WeatherNext 3, a global weather AI model from Google DeepMind and Google Research that uses live geostationary satellite data and real-world observations to update forecasts every hour. Its high-spatial-resolution predictions are up to 5x sharper than WeatherNext 2, supporting localized forecasts for fast-evolving storms, temperature shifts, and wind power output.

Also covered by: @Google AI, @Google DeepMind, @Google Research

#4 𝕏

Sundar Pichai says new Gemini voice capabilities are rolling out now for Google AI subscribers, enabling conversational Gmail searches, task and thought organization in Keep, and creation of new Docs. His post also demonstrates Docs Live.

#5 𝕏

Mustafa Suleyman shared MAI-Transcribe-2, claiming it is 10x faster than GPT-Transcribe, 5x faster than Gemini 3.5, ranked #1 on Artificial Analysis’s accuracy-latency Pareto frontier, and priced at the lowest level on the market.

#6 𝕏

LlamaIndex 🦙 released Turbo mode for Extract in beta, delivering structured document extraction roughly 4× faster than its Cost Effective Tier at comparable accuracy, with median latency of 3.7 seconds per page. By processing pages in parallel, latency stays nearly flat as document size grows.

#7 𝕏

Qwen announced E-Commerce Bench, a benchmark for long-horizon autonomous business operations in which agents start with ¥100,000 and run online stores for 365 days using real e-commerce data. Its seven-axis evaluation found that almost no model learned to buy more cheaply or improve over the year, and no single model dominated across all dimensions.

#8 𝕏

Guillermo Rauch shared the `vercel ai-gateway coding-agents setup` command, which points all coding agents to AI Gateway. He says it provides 100% uptime, observability, budgets, and easy switching.

#9 𝕏

Philipp Schmid shared that Managed Agents offers an easy way to try Gemini 3.8 Flash in an agentic environment, providing a dedicated remote sandbox in a single API call. It supports coding tools, full network access, Google Search, function calling, MCP, persistent filesystems, background tasks, and cron triggers through the Google AI Studio free tier.

#10 𝕏

Boris Cherny requested feedback on an early concept for making Anthropic’s Claude Code more extensible, describing it as “a little crazy, and very exciting.” More details are available in issue 91870 of the anthropics/claude-code GitHub repository.

Also covered by: @Thariq is on vacation

#11 𝕏

Aravind Srinivas announced that Portable Computer, described as a fully local runtime of Perplexity Computer, is now compatible with NVIDIA RTX GPUs on Linux, with Windows compatibility planned next.

#12 𝕏

Harrison Chase said agent workspaces should be durable, inspectable, and swappable, highlighting MongoDB support for LangChain’s Deep Agents virtual file system. Its BackendProtocol lets agent code use read, write, glob, and grep operations while teams choose their production storage layer.

#13 𝕏

Teresa Torres recently created an in-depth guide explaining what AI evaluations are and why product teams should use them. Evals measure AI product or workflow performance, helping teams maintain quality, catch issues before they reach users, and create a feedback loop similar to interviewing and assumption testing.

#14 𝕏

DeepLearning.AI recapped Self-GC, built by Xiaohongshu researchers to use a planner LLM to decide which context tokens to keep, fold, or prune. In tests, Self-GC retained necessary details 84.85 percent of the time, compared with 54.55 percent for standard methods.

#15 𝕏

Santiago shared an unnamed video-game benchmark where agents speedrun a game to test planning, action, learning from prior attempts, and long-horizon optimization through an autoresearch-like loop. A public leaderboard compares model performance.

#16 𝕏

Cognition announced that GPT-6 Astra is coming to Devin, claiming it performs within 0.4 points of Fable 5 on FrontierCode 1.1 at a 64% lower cost. Cognition also says Astra sets a new state of the art on its internal testing benchmark, producing more comprehensive tests, clearer reports, and better video evidence.

Also covered by: @Cognition

#17 𝕏

Santiago said there is no universally best model and that the best applications he has seen combine multiple models to leverage their strengths. He highlighted comparing models by task, response quality, cost, and reliability, as well as assessing different model combinations.

#18 𝕏

Rowan Cheung shared a six-step Wispr Flow + Claude Cowork journaling workflow that captures unstructured thoughts in Apple Notes, syncs them to his Mac, and uses a scheduled nightly Claude task to structure them in Notion into ideas, action items, and suggested GCal deep-work blocks. He says it flags follow-ups, organizes half-formed ideas, books focus time, and creates a searchable log of ideas from all year.

#19 📝 OpenAI News

Daybreak for Frontline Defenders - OpenAI is committing $1 billion in subsidized Daybreak access, training, technical support, and partnerships—targeted to be consumed over the next six months—to help frontline cyber defenders protect essential services like water, electricity, local government, community banks, nonprofits, and open-source maintainers. The effort includes Daybreak for America, Daybreak Blue and Red models, a public-sector and water pilot with MS-ISAC, prior emergency support of up to $1 million in no-cost API credits for attacked water systems, thousands of defenders across 2,000 approved organizations already using Daybreak, and a Daybreak Defense Network with more than 35 partner products and services.

#20 𝕏

Google Research shared an evaluation of approaches to cross-population genetic risk prediction, finding that transfer learning from European cohorts improves prediction in small populations but degrades accuracy as target cohort sample sizes grow.