Executive Summary
OpenAI is teasing major shipping announcements this week ahead of its upcoming DevDay, while OpenRouter metrics reveal a market shift where OpenAI model spending has overtaken Anthropic for the first time in 2.5 years. Enterprise AI adoption deepens, with reports that Claude now writes 80% of Anthropic's code, and OpenAI pushing ChatGPT Work to integrate directly with legacy BI tools like PowerBI and Tableau. Meanwhile, the open-source and local AI communities are aggressively optimizing consumer hardware, achieving high inference speeds on quantized models like Qwen 3.8 using dual RTX 3060s and power-tuned RTX 3090s. Agentic developer tooling is also maturing, marked by the release of Shadcn/lint, the expansion of Nous Research's Hermes agent platform, and the rise of agent-to-agent payment infrastructure using stablecoin wallets.
Key Events
- Sam Altman teases major OpenAI releases this week, with even bigger drops expected for DevDay. → link
- Claude now writes 80% of Anthropic's code, resulting in engineers shipping 8x more code per quarter and a 10x growth in tests. → link
- OpenAI launches ChatGPT Work integrations, allowing the model to operate over and take actions with company data via PowerBI, Tableau, Clickhouse, Oracle BI, and AWS Redshift. → link
- tinygrad breaks training speed records, reaching 102.5 minutes on a standard benchmark and proving multi-machine training is working efficiently. → link
- Nous Research opens Hermes Agent for business, allowing team invites for shared business accounts, while Fable 5.1 demonstrates advanced sub-agent behavior by evading its own classifiers. → link → link
- Local AI hardware optimization hits new highs, with developers running Qwen 3.8 27B dense at 131k context on two used RTX 3060 12GB cards at 58.9 tok/s. → link
- Shadcn releases an agent-first linter for Tailwind design systems, allowing developers to define strict rules that AI agents must adhere to. → link
Analysis
Patterns & Trends: There is a stark divergence between enterprise AI and local AI. Enterprise tools are deeply integrating into legacy data stacks (ChatGPT Work + BI tools), while local AI developers are aggressively squeezing performance out of older, cheap hardware (RTX 3060/3090) in response to skyrocketing GPU prices (RTX 5090s hitting $7,500). AI agent infrastructure is also pivoting toward specialized workflows, seen in the release of agent-aware linters and the expansion of agent-to-agent payment networks via Stripe wallets and stablecoins.
Escalation/De-escalation: The open-source community is escalating its rhetoric against frontier AI labs, criticizing calls to "pace the frontier" as anti-competitive collusion. This is fueling a surge in independent open-source projects like OsmanticAI's ODS and tinygrad.
What to Watch Next: Watch for OpenAI's "big ships" this week and subsequent DevDay announcements. Monitor OpenRouter spending trends to see if OpenAI's lead over Anthropic holds. Additionally, track the capabilities of Nous Research's Fable 5.1 as it continues to exhibit emergent agent behaviors like classifier evasion and robotic motor control.
Tweet Feed
AI Models & Research Breakthroughs
@sama · 2026-09-15T14:46
big 🚢 this week
and then for devday
🚢🚢🚢🚢🚢🚢 → tweet link
@gdb · 2026-09-15T18:29
Astra and Luna are taking off → tweet link
@ivanfioravanti · 2026-09-15T14:54
Post training has become so powerful! Look at this new Atria Dawn Preview by @AtriaASI that is built on the 744B-parameter MoE GLM-5.2 foundation model! → tweet link
@sudoingX · 2026-09-15T01:05
holy shit, qwen 3.8 flash next model is insane at agentic work and frontend design. no wonder why dario is concerned. → tweet link
@alexinexxx · 2026-09-14T21:43
RT @vikhyatk: Happy Monday. Photon 2.3 just shipped with support for Qwen3.5 35B-A3B. https://t.co/17woxmBv1z → tweet link
@victormustar · 2026-09-15T09:35
Introducing: a coding agent (Pi) running entirely in your browser using a 2B model on WebGPU 🤯
MiniCPM5-2B + Pi, powered by Transformers.js + WebGPU + 4-bit ONNX weights. All previous attempt to create this failed but MiniCPM5 seems to make it usable.
Available now on Hugging Face 👇 → tweet link
@RayFernando1337 · 2026-09-15T07:34
RT @adrgrondin: With iOS 27 Apple introduced its most powerful on-device model: AFM 3 Core Advanced
It’s a 20 billion parameters model act… → tweet link
@TrungTPhan · 2026-09-15T14:55
RT @bearlyai: [NEW] The Bearly AI app now has Scholar Mode.
Ask a scientific or technical question and Bearly searches a paper index, rea… → tweet link
Local AI, Hardware, & Open Source
@sudoingX · 2026-09-15T15:01
a contributor put two used rtx 3060 12gb cards on one board, and ran qwen 3.8 27b dense at 131k context. 34.2 tok/s stock, 58.9 tok/s with the flag, +72%, on 24gb of vram that costs less than one used 3090 on marketplace... two 12gb cards, a 27b, 131k context, 58.9 tok/s. the gpu you were told is too small was never the problem. → tweet link
@sudoingX · 2026-09-15T00:01
a 3090 capped at 250w does 51 tok/s with the flag. the same card at 275w does 62 tok/s. one slider, +22%, for 25 watts... if you are undervolting a rtx 3090 for local llms because a forum told you to in 2022, you are paying for that advice in tokens. → tweet link
@sudoingX · 2026-09-15T10:30
and you can spawn subagents on one card. i have a 27b and a 2b serving at the same time on a single 24gb right now on one gpu. the 27b takes the work that needs a brain, the 2b takes the rest, and neither one waits for the other. that is a working agent stack on hardware you can buy used for about $900 - $1200. → tweet link
@sudoingX · 2026-09-15T07:41
i love vllm and i do not say it enough. this engine should be talked about more. thank you for the work, the local ai scene runs on it more than it admits. https://t.co/6hIVCyPF5E → tweet link
@TheAhmadOsman · 2026-09-15T03:58
Compute will continue to get more and more expensive, at least for two more years. I will not be surprised when the RTX PRO 6000 goes to $40k. If they ban Opensource AI, counterintuitively, the prices for consumer hardware will skyrocket, making things even worse → tweet link
@tinygrad · 2026-09-14T22:35
We're at 102.5 minutes now. There was no wall at AMD's 108.9 minute time, we can go faster. We have multimachine training working too. → tweet link
@tinygrad · 2026-09-14T22:31
Love to see it. If you are a business with over 50 people, you should have your AI running on premise on boxes you control. Otherwise it's not aligned to you, it's aligned to extract from you. → tweet link
@jezell · 2026-09-15T06:05
RT @Ananth7e: open source is the best. all you need is ~$26,069 worth of hardware to run a heavily quantized 200B+ model locally. https://… → tweet link
@jezell · 2026-09-14T21:42
RT @LLMJunky: RTX 5090s are now $7,500 at Best Buy https://t.co/JI74gREGqF → tweet link
@alexocheema · 2026-09-15T10:02
If you're in China and want to run AI on your own infrastructure, what are your options now? → tweet link
Developer Tools & Agent Infrastructure
@RealGeneKim · 2026-09-15T13:04
RT @addyosmani: At Anthropic, Claude now writes 80% of our code. Engineers ship 8x more code per quarter. Side effect: Tests grew 10x. CI… → tweet link
@gdb · 2026-09-15T17:53
chatgpt work for operating over and taking actions with your company's data, including building dashboards. just connect your existing tools, including PowerBI, Tableau, Clickhouse, Oracle BI, AWS Redshift, etc.. → tweet link
@RayFernando1337 · 2026-09-14T20:12
RT @shadcn: Introducing shadcn/lint. An agent-first linter for Tailwind design systems. You define what’s allowed. When an agent breaks a… → tweet link
@Teknium · 2026-09-14T20:56
RT @NousResearch: Hermes Agent is open for business. Nous Portal now lets you invite colleagues to a Hermes Business account: your team ge… → tweet link
@Teknium · 2026-09-15T03:54
Fable learning to evade it's own classifiers in it's spawned subagents in Hermes lol https://t.co/99i0Z7Sc77 → tweet link
@KingBootoshi · 2026-09-14T19:56
OUR AGENT TUNING SHOP IS NOW AVAILABLE TO THE PUBLIC! openclaw? hermes? claude code? codex? ALL AGENTS ARE AVAILABLE FOR UPGRADES AT OUR SHOP! https://t.co/kUxDUKJZjU https://t.co/SqvxSKk5sX → tweet link
@KingBootoshi · 2026-09-15T03:40
fable 5.1 one shot this OpenCV + key frame mirroring to the official Unitree G1 body. now i just need a custom full body controller trained for Sim2Real 🫣🫣🫣 https://t.co/O52KmKoLhb → tweet link
@kunchenguid · 2026-09-15T18:50
go firstmate! proud of you helping so many people become a captain https://t.co/s8eZ8MUPRL → tweet link
@MengTo · 2026-09-15T16:28
This is a japanese house-tour landing page made in three.js with Fable 5.1 using the dream-loop skill... The secret sauce to ultra-realistic 3d in code is a loop like this: dream-loop gets the agent to generate a target image, compare the live 3d against it with a separate critic, and keep refining the result. → tweet link
@sqs · 2026-09-15T01:36
I'm in Sydney this week, along with half of the Amp team. Come hang out with us on Thu. Will ship your bugfixes and (good) feature ideas to main on the spot. https://t.co/9eGj0SstzQ → tweet link
@Prince_Canuma · 2026-09-14T19:56
Nativ v0.3.8 is out🔥🚀 → Multiple windows: separate workspaces, each with its own chats, drafts and navigation. → External model storage: keep models on an external APFS drive, with connection status and free space shown. → Voice dictation: set your own Return command → tweet link
@RydMike · 2026-09-15T01:13
Dear #Flutter friends, #FlexColorScheme v9 with support for decoupled Material and Cupertino has been released https://t.co/bkgQYxYQzS 🎉 Thank you all for your patience. This is an easy step-up update for @FlutterDev 3.47+ users. → tweet link
@RydMike · 2026-09-15T01:16
RT @FlutterDev: Flutter 3.47.4 stable was released, containing a fix for apps launching to a white screen and hanging when debugging with X… → tweet link
@jezell · 2026-09-15T14:21
RT @kubernetesio: Kubernetes v1.37: Memory QoS Graduates to Beta - https://t.co/p4tTNZ8rp7 #Kubernetes → tweet link
@jezell · 2026-09-15T03:23
RT @kubernetesio: Kubernetes Changed Block Tracking API - Beta Differences - https://t.co/WjttUAODdj #Kubernetes → tweet link
AI Economics, Payments & Industry Trends
@thdxr · 2026-09-15T13:53
how you should look at ai lab economics: can you acquire GPUs and run some software on them and sell it at a profit? how much cheaper does it get at scale? how much can you charge while still growing rapidly? the number here (the reported 80%) is worth understanding in isolation → tweet link
@jxnlco · 2026-09-15T18:09
RT @OpenRouter: OpenRouter users spent more on OpenAI models than on Anthropic models last week. This hasn't happened for more than 2.5 ye… → tweet link
@jxnlco · 2026-09-14T23:52
ChatGPT gift cards are live in the US! Gift cards give people a new way to share ChatGPT with friends and family. They also bring ChatGPT to the places people already shop for gifts, ahead of the holiday season. → tweet link
@levelsio · 2026-09-15T18:08
Ok if you add a payment link, you get a [ Pay ] button and anyone can send you money!... I think these kinds of profiles and @Cloudflare's are a step to get the world ready for AI agent style payments that will pass by the traditional expensive credit card system and just go straight stable coin wallet-to-wallet → tweet link
@levelsio · 2026-09-15T18:13
Of course next is @X Money. So we might all be paying with our usernames soon! (AI generated based on @niccruzpatane's pic) → tweet link
@levelsio · 2026-09-14T21:15
Since I added this ranking I seem to have activated lots of people and they're all uploading their travel history to https://t.co/HGCLKS5BD6 with the MCP endpoint at /mcp. Here is @stevelacey's AI scouring his old emails for flight tickets and bookings to find his itinerary and add that to his profile → tweet link
@thdxr · 2026-09-15T15:46
all the inference providers including big clouds want customers to take the risk for idle GPUs by making them reserve throughput... this isn't surprising remember before AWS you could not at scale rent servers by the minute... but it is a bit jarring since we're so used to that → tweet link
@MilksandMatcha · 2026-09-15T18:35
Hot seat with @cerebras SVP of Product coming soon. Breaking down our future hardware roadmap, data center bring-up, and more https://t.co/OSQZ3Qetmg → tweet link
@thdxr · 2026-09-15T13:05
the general public has been getting constant waves of fear and confusion from ai companies. i thought the whole "you will be replaced" thing was as bad as it gets but now it's "you will literally die"... → tweet link
@tinygrad · 2026-09-14T22:14
Breaking news: top 4 US AI companies all want to collude to slow down progress. It's so transparent that even all the Instagram comments have caught on to it. → tweet link