⚡ Grok 4.7: Half the Price, Twice the Speed

SpaceXAI released Grok 4.7 on Monday, describing it as the company’s most capable model for coding and knowledge work, and pricing it at $2 / $6 per M tokens — the same price as Grok 4.6 and a fraction of what the models it is measured against charge. GPT-5.6 Sol is listed at $4 / $20 per M and Claude Fable 5.1 at $10 / $50 per M on the same comparison chart.

The model is a new, larger base model than Grok 4.6, trained with a longer reinforcement-learning run weighted toward tasks that take many hours to finish, with more self-verification and better long-context handling. SpaceXAI also trained it natively on its own Grok Bot harness, which is why the agentic benchmarks move most:

  • CursorBench 4.0 (long-running coding): 46.3%, against 40.4% for Grok 4.6, 41.7% for GPT-5.6 Sol and 51.8% for Fable 5.1.
  • EEBench: 64.0%, versus 53.0% for Grok 4.6 and 39.4% for GPT-5.6 Sol.
  • Harvey Legal Agent Benchmark: 19.6%, against 15.8% for Grok 4.6 and 2.5% for Sol.
  • Terminal-Bench 4.0: 38.0%, close to Sol’s 37.3% at a quarter of the token price.

The safety pitch is unusually specific: a rebuilt safeguard stack, the company’s best jailbreak-resistance numbers so far, 62.4% on LatchBio’s biosafety benchmark, and a claim that only 3.3% of risky dual-use cyber prompts got through on its HackerBench v0.3 internal test while legitimate security work was rarely blocked. Select cybersecurity partners are getting invite-only access to the model’s red-teaming capability. It shipped in Cursor, in Grok Build, on the Grok API and through routers on day one — including, four hours later, OpenRouter, which lists it at $1.60 / $4.80 per M tokens with a 500K context, undercutting the vendor’s own published price. Hacker News gave the launch 583 points, its top AI story of the window.

🇨🇳 Alibaba’s 20 Gigawatts and a 10-Trillion-Parameter Plan

At the Apsara (Yunqi) Conference in Hangzhou on Tuesday, Alibaba CEO Eddie Wu unveiled the Zhenwu V900, which the company calls the most powerful AI chip in China, and a target of more than 20 gigawatts of global data-center capacity by 2032. T-Head, Alibaba’s semiconductor arm, says the V900 delivers roughly three times the compute of the Zhenwu M890 it announced in May, and scales to 500,000 cards in a single cluster.

Wu framed the strategy as three pillars — models, chips, cloud — and put the roadmap in explicitly superintelligence-adjacent terms: machine intelligence currently represents less than 3% of human capacity, so Alibaba is investing across all three layers, is working on recursive self-improvement, and plans Qwen models of 5 to 10 trillion parameters. It also launched Qwen Intelligence for mobile devices, an end-to-end stack for running multi-step tasks locally on a phone. Shares rose about 3% in Hong Kong.

The numbers worth holding next to each other: a 20GW build-out target that mirrors the ambitions of the US hyperscalers, an in-house accelerator meant to reduce dependence on Nvidia under export controls, and a parameter target 3-4x larger than anything currently serving traffic. Alibaba Cloud is the only company in China with a credible full stack — model, chip, and datacenter — and it is now selling that stack as the alternative to the one Washington is willing to license.

📱 Xiaomi Ships a 1-Trillion-Parameter Open-Weight Family

The quietest launch of the window may matter most. Xiaomi put MiMo-V2.6-Pro, MiMo-V2.6-Pro-UltraSpeed and MiMo-V2.6-Flash on OpenRouter within hours of the Grok release, all with a 1.05M-token context, and published the weights on Hugging Face the same day (MiMo-V2.6-Pro-RL, MiMo-V2.6-Flash-RL, plus a Distill-Qwen-9B variant).

Model Context Price (in / out)
MiMo-V2.6-Flash 1.05M $0.14 / $0.28 per M
MiMo-V2.6-Pro 1.05M $0.435 / $0.87 per M
MiMo-V2.6-Pro-UltraSpeed 1.05M $4.35 / $8.70 per M

Pro is a 1-trillion-parameter mixture-of-experts model with 42B active parameters, MIT licensed, accepting text, image, speech and video input. Artificial Analysis scores it 46 on its Intelligence Index — first in the open-weights large class and above the 18 median for comparable models — at 125 output tokens per second, with a 99% cache discount that drops the effective input price to about $0.13 per M. A Hacker News analysis thread on the Pro model reached 103 points.

Set that against the closed-model chart: $0.14 per M input tokens for a frontier-adjacent multimodal model, versus $2 for Grok 4.7 and $10 for Fable 5.1. The capability gap is real; the price gap is now an order of magnitude, and it is being set by labs that give the weights away.

🕳 A Zero-Day in the Assistant That Has All the Keys

Meta spent the past two weeks publishing posts about how seriously it takes the security of Muse, its consumer AI agent, and Mark Zuckerberg has said the assistant was built from the ground up for privacy and security. On Monday, macOS security researcher Patrick Wardle disclosed a zero-day that punctures the claim: any locally installed app or terminal command — regardless of the macOS permissions it holds — can write a long list of undocumented Muse settings.

Most of those settings are cosmetic. One is not: it changes the endpoint where transcription is sent. Point it at an attacker’s server and the attacker collects the token that authenticates the user to their Muse account. “We can manipulate the agent and leverage its privileges to do whatever we want,” Wardle told Ars Technica. “So instead of us having to write a very comprehensive Mac malware stealer, we can just leverage the AI assistant itself.” He built proof-of-concept attacks that write malicious files to disk and take photographs, in some cases with no visible warning. Meta did not answer Ars’s questions.

Three threads now run through the same product. Amazon began blocking Muse from shopping on its site on Sunday night, and shoppers see a popup citing its Conditions of Use. Muse downloads are surging — CNBC put it against ChatGPT, Grok and Claude — and the WSJ argues agents have stopped being an enterprise story. And the architectural pattern Meta documented, a monitoring agent code-named Sentinel sitting between Muse and the internet, is exactly the layer a token-theft bug bypasses. The assistant users were told to trust with their calendar, email, camera and WhatsApp is itself a privileged process, and it was shipped before that sentence was fully thought through.

✈️ SMART Goes Live Over the Busiest Airspace on the East Coast

Transportation Secretary Sean Duffy and FAA Administrator Bryan Bedford unveiled SMART — Strategic Management of Airspace, Routes and Trajectories — on Monday and began running it in limited mode in the airspace around Washington. It ingests around 200 data streams including weather, flight paths, traffic flow and controller staffing, and uses an AI-supported engine to project congestion and recommend fixes before a delay compounds. The FAA says it cannot replace controllers or take control of an aircraft, and plans to expand it gradually across the country over the next several months.

Representative Don Beyer (D-Va.), whose district contains Reagan National, called for the system to be suspended at least until controller feedback and independently verified stress-testing establish that it is safe. He said he was told controllers were not consulted on the design and not trained before deployment, and that they had been told the system can adjust schedules and potentially routes. He cited a Politico report that airlines are privately alarmed at the rollout speed, and went with the quote of the day: the FAA should not use my constituents as guinea pigs. The agency counters that it has about 11,000 controllers — its highest staffing in a decade — with 4,000 more in the hiring pipeline, and notes that roughly 24 million people pass through National each year.

The context is not abstract. On Sunday, a cut fiber line halted flights at busy East Coast airports; that Reuters story was the second-highest-scoring item on Hacker News with 225 points. Predictive tooling is exactly what such a failure argues for. Deploying it without telling the controllers, in the airspace where a fatal midair collision happened last year, is what the objection argues against.

🗳 Three in Four, and a Congress That Already Said No

A new Reuters/Ipsos poll — 1,277 US adults, four days of fieldwork ending Sunday, margin of error 3 points — found 73% saying AI companies have not done enough to prevent the technology from causing serious harm to society.

  • 39% say AI is having a negative impact on society, up from 36% a month earlier and the highest reading since the series began in March. Only 11% call it positive.
  • 55% say slowing AI development would be a good thing, against 13% who say it would be bad.
  • 69% have followed the news about major AI companies calling for a slowdown, oversight and safety standards.
  • 73% rank ensuring AI is developed safely and responsibly above keeping the US ahead of other countries, which 23% chose.

Read alongside the courts: the same week four paid subscribers filed the antitrust suit arguing that the labs’ voluntary pacing pledge is a horizontal agreement not to compete, Senators Josh Hawley and Ted Cruz blocked a national-security antitrust exemption for AI companies in the NDAA manager’s package. The industry asked for political cover to coordinate on safety, did not get it, and now has an active docket instead.

💸 The Bill Comes Due, Politely

The financing layer had a rough 24 hours, without anything actually breaking.

  • Jamie Dimon said hyperscaler AI spending could hit $1 trillion next year — a number that describes demand, and a bill that somebody has to fund.
  • The WSJ reported the AI stock rally losing steam with oil below $100; Reuters found corporate bond buyers getting picky about the flood of AI debt; the NYT wrote that Wall Street is growing skeptical of the data center boom, and TechCrunch framed Nscale’s coming IPO as another test of appetite for concentrated bets.
  • AMD joined the $1 trillion club on the strength of the AI build-out, with Deutsche Bank’s analysts working out how much money the industry still has to find.

The constraint has moved. It is no longer chips alone: California tightened datacenter rules on water and power this week, Barron’s argued SpaceX holds a key advantage in the AI power battle, and Alibaba’s answer to the same problem is to build 20GW itself. Compute is a land, water and credit problem now.

🎥 Quick Hits

  • Malware that learns. WIRED reports a new detection tool that found malware guided by an AI hive mind with no humans in the loop, while Fox News covers AI malware that rewrites itself to evade detection — the same week Apple’s and Anthropic’s agent incidents are still being digested.
  • Banks are worried about shopping agents. Reuters: lenders warn that AI shopping bots raise scam, fraud and data-privacy risks, which is the financial version of Amazon’s argument against Muse.
  • The US Treasury’s line on liability: AI executives, not their bots, will carry the can for criminal acts — a position that will be tested the first time an agent moves money.
  • DeepSeek will brief the UN Security Council on AI this week, per a Reuters exclusive; Sam Altman is scheduled to brief the same body.
  • Anthropic and OpenAI asked Australia to relax its ban on using local data to train AI models.
  • JetBrains Air launched a system of products for agentic software development (55 points), and Linear rewrote its CI because AI-generated code made it the bottleneck — 268 points, the most-discussed engineering write-up of the window.
  • Apple’s AI delay gets a price: up to $95 back for some iPhone users (Axios). The M5 Ultra Mac Studio review calls it the dream Mac for local AI agents (253 points); Show HN entry Mini-AGI trains a continual-learning model on 8GB of VRAM (264 points).
  • Culture, again. Robin Williams’ daughter told fans making AI videos of her father to have some shame (91 points); Shopify’s CEO says his employees’ AI slop grenades create work for everyone else; and CBS asks what an AI swarm is and why it troubles experts who build swarms.
  • Mathematics is the demo of choice. OpenAI named an advisory group on mathematics and AI; mathematicians used AI to find symmetries in a decades-old problem (Scientific American); NPR asked whether AI solving one of maths’ hardest problems taught humanity anything, and a Scientific American piece asks whether OpenAI solved the wrong Navier-Stokes problem. AI-written code was an estimated 17.25% of all Linux kernel patches in September.

⏰ OpenRouter Limited-Time Free Models

The list below is not a free tier — it is every OpenRouter listing carrying a published sunset date. This cycle that means 10 endpoints, of which 3 are genuinely free and 7 are discounted commercial models whose expiry dates cluster into vendor deprecation batches.

Model Context Price (in / out) Sunset
Nex-N2.5-Mini (free) 262K free 2026-09-25
Nex-N2.5-Pro (free) 262K free 2026-09-25
Dots3-Note Preview (free) 512K free 2026-09-30
DeepSeek V3.2 164K $0.269 / $0.40 per M 2026-09-28
DeepSeek V3.2 Exp 164K $0.27 / $0.41 per M 2026-09-28
DeepSeek V3.1 Terminus 164K $0.27 / $1.00 per M 2026-09-28
DeepSeek R1 Distill Llama 70B 8K $0.80 / $0.80 per M 2026-09-28
MiniMax M2.1 205K $0.30 / $1.20 per M 2026-10-08
Baidu ERNIE 4.5 VL 424B A47B 123K $0.42 / $1.25 per M 2026-10-08
Gemini 2.5 Flash Lite 1.05M $0.10 / $0.40 per M 2026-10-20

Both Nex free slots expire on Thursday, and the whole DeepSeek V3.2 and V3.1 generation leaves on September 28 — days after DeepSeek itself was reported to be briefing the UN Security Council. The Dots3-Note Preview at 512K context remains the best free deal on the board with eight days left.

🤗 Trending on Hugging Face

The collector’s new-uploads list was 100% zero-download test repositories for the 18th consecutive cycle, so this section uses the official trending feed.

The story of the feed is convaiinnovations/laya, the typed-decision model that has now gone from 811 likes to 2,222 likes in three days while still recording zero downloads. The metadata is being read, argued about and built against faster than the weights are being pulled — and the derivatives now outnumber the parent: laya-multilingual (160 likes), laya-typed-decisions, an MLX port, a CoreML port for the Apple Neural Engine, an ONNX export, a vision variant and a multilingual variant, all released inside four days.

Around it, the feed is dominated by image and quantisation work:

⭐ GitHub Trending: AI Edition

The collector’s repository list was out of window for the seventh consecutive run — all five entries were created on or before September 20, one of them 65 minutes before the last issue. Entries below marked (API) were substituted from the GitHub search API after per-repo created_at verification; the in-window pool is again small.

  • zai-org/ZCode — ⭐ 6,160, TypeScript, up from 5,031 yesterday. Z.ai’s coding agent harness, still climbing a week after the silent-upload disclosure.
  • mizorewww/laya-mlx — ⭐ 4,929, Python, up from 2,666 yesterday and 626 the day before. The native MLX runtime for Laya typed-decision models is now the fastest-growing repository in the ecosystem.
  • jev-chat/jev-chat-jarvis — ⭐ 3,077, Kotlin. Same repository ID as the entry that carried 34 stars in yesterday’s issue, now under a dedicated org: a 90x day, for a conversation copilot that reads WeChat, QQ, X and Feishu screens and drafts replies without hooking any app.
  • Mak5er/AirCard — ⭐ 2,539, Swift. Apple Wallet card skins for iOS 18+, no jailbreak.
  • jarrodwatts/jev-trader — ⭐ 1,971, TypeScript. One AI trade decision per Monad block.
  • (API) kydlikebtc/awesome-jev — ⭐ 78, Python. 148 verified Jev examples indexed by decision type: the top genuinely in-window repository.
  • (API) yibie/jev-engineering-zh — ⭐ 37. A complete Chinese translation of Jev Engineering: For Coding Agents, illustrations included — the clearest sign yet that this ecosystem’s centre of gravity is Chinese-language.
  • (API) JunMa11/MedJev — ⭐ 28, Python. Typed decisions applied to clinical work.
  • (API) tonyd2wild/MiMo-V2.6-Flash-2x-DGX-Spark — ⭐ 23, Python. Running Xiaomi’s new model on two DGX Sparks with vLLM, DFlash and fp8 KV cache — the first hardware recipe for the weights released hours earlier.
  • (API) Calcium-Ion/AstrLink — ⭐ 21, Go. A local gateway unifying subscriptions and API providers for agents.

💡 Key Trends

  1. The price of frontier capability just fell again, and it fell on the open side. Grok 4.7 launched at $2 / $6 per M tokens against rivals at $4 / $20 and $10 / $50, and was undercut within four hours by Xiaomi’s MiMo-V2.6-Pro at $0.435 / $0.87 with the weights published under MIT and a score near the top of the open-weights class. Ternary-Bonsai sits at $0.075 / $0.50. Nobody is winning on capability alone this week; several people are winning on price per unit of work.

  2. Compute is now a land, water, power and credit problem. Alibaba’s answer is to build 20GW and its own accelerator, targeting models of 5 to 10 trillion parameters. California’s answer is to tighten datacenter rules on water and power. Wall Street’s answer, per the WSJ, Reuters and the NYT in the same 24 hours, is to get picky about AI debt — even as Dimon says the spending could reach $1 trillion next year.

  3. The agent has become a privileged process, and the security models are being written in public. A single undocumented setting in Meta Muse turns any local program into the owner of the account; the monitoring agent built to guard it does not help, because the token is the thing being stolen. Amazon’s blockade, the banks’ fraud warning and the Treasury’s liability line are three institutions discovering the same problem from different directions.

  4. Institutions have stopped accepting self-reports. An FAA tool that ingests 200 data streams went live over Washington airspace with, according to a congressman who represents it, no controller consultation or training. Anthropic published the most transparent development metrics any lab has ever released, four days after a Senate that blocked the antitrust exemption the labs wanted — and a poll out the same day found 73% of Americans think the companies are not doing enough. Voluntary assurance is being replaced by instrumented evidence, demanded from outside.

Today’s Snapshot

  • OpenRouter: 4 models collected, 4 genuinely new (MiMo-V2.6-Pro, MiMo-V2.6-Pro-UltraSpeed, MiMo-V2.6-Flash and Grok 4.7), all created after the September 21 issue and verified by created-timestamp sweep. First 🆕 section in three days.
  • Limited-time listings: 10 endpoints with a published sunset date — 3 genuinely free (Nex-N2.5-Mini and Pro to September 25, Dots3-Note Preview to September 30), 4 DeepSeek deprecations on September 28, 2 on October 8, 1 Gemini 2.5 Flash Lite on October 20.
  • Hugging Face: 10 collected, all zero-download junk for the 18th consecutive cycle — trending feed used instead, 9 models featured, headlined by the Laya ecosystem.
  • GitHub: 5 collected, 0 in window — 5 API-verified substitutes added, top in-window repository 78 stars.
  • Papers: 0.
  • Collector total: 29 items. Published articles: 97.