Executive Summary
The last 24 hours have been dominated by the rollout of OpenAI's GPT-6 Astra, which users are leveraging for highly complex 3D modeling, rapid application development, and seamless multi-agent orchestration. Developer tooling continues to evolve at a breakneck pace, with new agentic code editors like Zed Delta and advanced open-source frameworks like Hermes pushing token efficiency and sub-agent coordination. Meanwhile, compute constraints are shifting from GPUs to CPUs, as agents increasingly execute generated code in live environments, prompting a surge in local hardware investments and infrastructure re-evaluation.
Key Events
- GPT-6 Astra demonstrates advanced physical manufacturing capabilities, generating cm-accurate Blender models and 3D printable .stl files from simple photos and text prompts in minutes. → link
- Zed launches "Delta", a new agentic code editor, signaling a major shift in IDE paradigms towards AI-driven development. → link
- Ollama introduces off-peak hour token rates for DeepSeek-V4-Flash and Pro, offering 50% discounts outside peak business hours. → link
- Qwen 3.8 Flash Next support is added to DwarfStar, highlighting incredible prefill capabilities flat up to 262K context. → link
- Developers report massive compute costs, with the 90th percentile of OpenAI researchers spending over $7000 a day on coding agents, shifting compute bottlenecks to CPUs. → link
Analysis
Patterns: There is a clear escalation in autonomous coding capabilities, with users noting that models like Astra increasingly prefer writing raw Python to perform tasks rather than relying on built-in tools or APIs. This is creating UI/permission tensions for harness developers who must decide whether to force models into their toolchains or let them run raw code. Additionally, the cost of inference at scale is becoming highly visible, both in cloud spend ($7000/day for top researchers) and local compute bottlenecks (CPUs over GPUs as agents actively use computers).
What to watch next: How harness developers adapt their UI/UX to agents that want to write raw scripts, and whether smaller "taught" models (like the upcoming GPT-6 Sol) will be viable as sub-agents for the larger "trained" models, or if direct interactive use will be restricted to top-tier models only.
Tweet Feed
AI Models & Capabilities (Astra, GPT-6, Qwen, DeepSeek)
@LinusEkenstam · 2026-09-06T13:36
Now this is wild. After providing Astra with some rough ideas on our studio space, and a handful of photos. It created a full 3D model in Blender thats cm accurate. The crazy part, it modelled the entire thing in 11 minutes 🤯 → tweet link
@LinusEkenstam · 2026-09-06T17:19
I think the next insane thing with Astra and 3D will be in manufacturing. Vibe-manufacturing with 3D printers en masse. In all sorts of materials. [...] Gave it to Astra, alongside with the rest of the info. Astra actually found some manual online with a small 2D drawing, combined with my photos and the measurements it was able to model the part in less than 2 minutes. It made a .stl file, sent it to my printer in Sweden, and 1.5h later the part was printed and ready for my mom to pick up. → tweet link
@MengTo · 2026-09-06T08:05
Astra can create beautiful three.js UIs entirely in code. You don't need an assets library or MCP to create detailed 3D models. These are all code and are easy to customize for your agent. Give Astra a video, URL, or code reference, then ask it to keep comparing the result against the reference and fixing mistakes. Liquid Glass, shaders, motion design, procedural three.js models, and full environments. All from a few prompts. → tweet link
@steipete · 2026-09-06T01:01
Can’t remember last time we had such a large jump in capabilities. → tweet link
@jxnlco · 2026-09-06T05:33
RT @TetraspaceWest: GPT-3 davinci price: $60/$60 per 1M tokens in/out GPT-6 astra price: $10/$50 per 1M tokens in/out → tweet link
@ivanfioravanti · 2026-09-06T09:30
Qwen 3.8 Flash Next support in DwarfStar is ready! @kernelpool + @ddalcu + @Spangler3000 + me combined. Prefill with this architecture is incredible! Flat up to 262K! PR here: https://t.co/e2qGfX91de Trying to push decode with MTP even more, and working on Q2 Model for 64GB machines as foreseen by @antirez 🚀 → tweet link
@ollama · 2026-09-05T23:04
Introducing off-peak hour token rates. DeepSeek-V4-Flash and Pro are now half price outside of 12:00 to 18:00 UTC on weekdays (5am-11am pacific), and all day on weekends! Off-peak pricing will be available soon for more models. → tweet link
@kunchenguid · 2026-09-05T19:48
have you noticed something new with claude 5 and gpt 6? this is the first generation where they released the biggest model fable and astra ahead of the smaller ones [...] the most reasonable explanation i can think of is that only fable and astra are actually “trained” opus 5, sonnet 5, and the upcoming gpt 6 sol, terra, luna are all “taught” → tweet link
Developer Tools & Agent Frameworks
@Teknium · 2026-09-06T17:12
🔥🔥🔥 We've made huge improvements in token efficiency in Hermes over the last 2 weeks. Give your codex sub a try in Hermes Agent 🫡🫡 → tweet link
@jezell · 2026-09-06T17:37
Zed just launched delta. Their take on an agentic code editor: https://t.co/gP4KGc39BS → tweet link
@thdxr · 2026-09-06T15:34
stepping back there's tension right now with "agent just writes code" and - being able to see what it's doing - permissions [...] but if astra really wants to write raw python, should we even be fighting that and making it use our own thing? tricky questions → tweet link
@thdxr · 2026-09-06T15:16
so astra writes python to do everything which means we're going to have to rework this ui → tweet link
@levelsio · 2026-09-06T15:34
✨ Added an MCP and API endpoint now. You can also add trips via POST on /api/trips or via /mcp, which is nice because it's a lot of work to add all your old trips to your profile! Agent ready 😊 → tweet link
@thdxr · 2026-09-05T22:51
made a lua based widget system on top of opentui can prompt your agent to build widgets and create a little dashboard for yourself → tweet link
@thdxr · 2026-09-05T19:59
it's worth considering if you can make your apis authless as a hook for example, if a domain registrar just made domain search apis public and LLMs became aware of it, they'd naturally use them which would be a nice source of customers → tweet link
Hardware, Compute, & Infrastructure
@jezell · 2026-09-06T08:37
Yep. Welcome everyone to the CPU shortage. Like I've been saying, CPUs are the problem now, not GPUs. Doesn't matter how fast you can generate tokens because the tokens are using computers now. → tweet link
@jezell · 2026-09-05T20:21
Told you this was coming. Buy some local hardware while you still can. → tweet link
@jezell · 2026-09-06T16:42
RT @scaling01: Meanwhile the 90th percentile of OpenAI researchers is spending over $7000 a day on coding agents → tweet link
@gospaceport · 2026-09-05T19:04
Said it before, golden model. I'm back on it after my trist with Flash Next also 😻 If I had the GPUs to run FN at higher precision I think it would edge out but 24 subagents on 27B is a hell of a drug! → tweet link
@sudoingX · 2026-09-06T02:26
what’s the smallest machine you’ve successfully run an LLM on? i’ll go first. 6GB VRAM. beat that anon. 👀 → tweet link
AI Impact & Industry Trends
@kunchenguid · 2026-09-06T18:08
we’re entering an interesting age where AI makes repeating something that already exists drastically easier than creating something truly novel i worry the side effects are - 1. innovators who don’t have distribution get punished [...] 2. products with no network effect will all race to the bottom → tweet link
@TrungTPhan · 2026-09-06T17:00
RT @bearlyai: The Economist says AI has added 1 million new jobs in America since mid-2023 (offsetting 200k AI related job losses): “Rough… → tweet link
@nummanali · 2026-09-06T08:21
If an agent can successfully edit binaries Then, aren’t we at the point where code is simply to be compiled? In a six months to a year from now, it looks to be that code from LLMs will not be for humans to read Doesn’t it make sense to focus on the black box testing? → tweet link
@LinusEkenstam · 2026-09-06T09:50
The AI spam reply issue on here is 10x worse than before... sort it please @elonmusk → tweet link
@TheAhmadOsman · 2026-09-06T04:16
I have never worked harder in my life. AI is gonna give us all some really good opportunities, better be prepared and work your ass off → tweet link