OpenAI releases GPT-6 Astra in Codex and API

#1 𝕏

OpenAI released GPT-6 Astra to all Pro, Enterprise, and Business Premium users in ChatGPT Work and Codex, and it is also live in the API. Rollout to Plus and Business users might take a few days.

Also covered by: @How I AI Podcast, @Randy Counsman, @AI Explained, @Sebastian Raschka, @OpenAI, @v0, @Sam Altman, @Fireship, @Nicholas Thompson

#2 𝕏

Google AI recapped this week’s shipments: Gemini 3.8 Flash and Flash Cyber, Lyria 3.5, WeatherNext 3 from Google DeepMind and Google Research, and Agentic Video Understanding. The updates span coding, agentic workflows, multi-step reasoning, cybersecurity, music generation, weather AI, and more accurate video analysis with reduced token usage and costs.

#3 𝕏

Muse Spark 1.3 with max reasoning is now available on Muse Code and the Meta Model API.

Also covered by: @Alexandr Wang

#4 𝕏

Mustafa Suleyman says MAI-Image-2.6-Flash is available to try now, generating images 2x faster than GPT-Image-2 with 72% more efficient GPU usage. He also claims it has the world’s best price-performance score, though no benchmark methodology or pricing was provided.

Also covered by: @Mustafa Suleyman

#5 𝕏

Anthropic announced that Claude completed the first formalized proof of Fermat’s Last Theorem last month—a project expected to take many years and the largest Lean proof ever written. The machine-verifiable proof spans over 13 million lines of code and proves over 29,000 additional required theorems.

#6 𝕏

LlamaIndex 🦙 shared its document extraction accuracy-versus-cost evaluation of 14 frontier systems across 370 enterprise documents, reporting that higher per-page costs did not yield better extraction. Agentic Plus achieved the highest overall accuracy at less than one-third the cost of the runner-up, while Agentic and Cost Effective routinely outperformed systems costing several times more per page.

#7 𝕏

Aravind Srinivas shared a deep dive into how Perplexity serves search results at scale, covering embeddings for ranking, GPU-based model inference, request batching, inference servers, and latency/throughput trade-offs.

#8 𝕏

NVIDIA AI shared five practical guidelines for using speculative decoding to speed up LLM inference without sacrificing accuracy, balancing throughput and latency by tailoring draft length and drafting method to the model, workload, and hardware.

#9 𝕏

Andrew Ng presented the AI Engineering Skills Map, describing it as covering important skills for using AI coding agents effectively.

#10 𝕏

Dharmesh Shah demonstrated YouSpot, an AI-native Solo CRM for one-person companies, using cloud-based Tracker Agents that run natural-language searches and automatically add results to a Second Brain for access via YouSpot Chat or MCP in ChatGPT or Claude. He said it costs $10/month and can be canceled anytime.

#11 𝕏

Santiago recapped how Cline was the second most popular coding agent at a friend’s unnamed company, behind Claude Code. After the Pentagon contractor dropped Claude Code and moved its use to Codex, it also sought to migrate Cline users but could not. He characterized Cline as having one of the AI community’s most committed fan bases.

#12 𝕏

Garry Tan recapped using AsideAI’s AI harness to set up OpenClaw with Slack, saying a two-hour process took less than three minutes with Aside’s full integrations, browser integration, full features, and smart access-control defaults.

#13 𝕏

Alexandr Wang commented on the updated artificial analysis index, saying Muse Spark 1.3 Max still performs quite well and that the efficient frontier consists entirely of Muse, Claude, and GPT.

#14 𝕏

Santiago commented that videos generated with Visko Orbis 1.0 ranked best in a blind test, particularly for keeping people, objects, and surroundings consistent as scenes unfold. He noted that longer AI-generated videos still accumulate errors until the output falls apart.

#15 𝕏

DeepLearning.AI recapped highlights from The Batch: Andrew Ng on the fundamental skills required to use coding agents, new enterprise data retention policies from OpenAI and Anthropic, Zai’s cost-efficient open-weights multimodal GLM-5.3-Flash, and Thomson Reuters’ 397B-parameter model for law, finance, and news.

#16 𝕏

Madhu Guru recommended choosing a familiar personal or work workflow and automating it end to end with AI, experimenting with a few products if needed. The exercise raises 4 questions—experience design, MCPs and tools, human involvement, and evaluation—and, he claimed, doing it once teaches more than reading about AI product building for a month.

#17 𝕏

Guillermo Rauch described an @eve agent as orchestrating the software development lifecycle, characterizing it as a new SDLC orchestrated with AI.

#18 𝕏

Guillermo Rauch said all the software factories he referred to as “our software factories” have human-in-the-loop operators, but the percentage of human interventions will decline over time.

#19 𝕏

Garry Tan announced that GStack now uses AsideAI as its preferred remote-session browser, citing its credential handling, integrations, harness, and memory system. He called it his new favorite AI browser and the “#1 absolute best” option he has found for agent web and credential access.

#20 𝕏

Lenny Rachitsky characterized Google Alerts as “Tired” and Overheard @bot as “Wired,” commenting on a Grok Bot post about marketplace Bot templates. The post presented Haggle Bot as an AI procurement specialist and claimed $100K+ in savings.