---
title: "Alibaba ships Qwen4-Max at $1.50/$4.50; OpenAI opens a science blog"
date: 2026-11-16
summary: "Alibaba released Qwen4-Max, a closed-weight API flagship with native vision and a 1M-token context, at $1.50 per million input tokens and $4.50 per million output. The same weekend, OpenAI opened a science blog with a lattice Standard Model proof that GPT-6 Astra helped build."
cover: "https://morningtokens.com/covers/2026-11-16.jpg"
url: "https://morningtokens.com/issues/2026-11-16/"
---

Pour the tea slowly today: the weekend's biggest drop came from Hangzhou, where Alibaba shipped Qwen4-Max as a closed flagship at a fraction of US prices. OpenAI opened a science blog with a physics proof that GPT-6 Astra helped build.

Nvidia goes into Tuesday's earnings with its former CEO reportedly weighing a comeback, while Washington moves to close the Southeast Asian chip route. Also inside: Claude Code's new Code Mode, a 1.6T open model served entirely on Huawei chips, and Shopify's agents that attack its own production systems.

## Top news

### 1. Models · Qwen4-Max is out: frontier-class claims at $1.50 / $4.50 per million tokens

Alibaba released Qwen4-Max, the first Qwen4 model, as a closed-weight API flagship with native vision and a 1M-token context. Smaller Qwen4 models are promised as open weights.

- List price per million tokens:
  - Input: **$1.50**
  - Output: **$4.50**
- Generally available since 16 November on QwenCloud / Alibaba Cloud Model Studio, model ID `qwen4-max`.
  - The Token-Plan-only `qwen4-max-preview` (from 11 November) is retired after 25 November.
- Architecture per the Qwen post:
  - 2.4T backbone parameters, 71B active per token, plus 328B "N-gram embedding" parameters
  - It is a new MoE with native vision (text, images, video, PDF)
- Specs: 1M context; up to 262,144 tokens for thinking plus answer; thinking on by default (`reasoning_effort` low / high / xhigh).
  - Explicit cache with 5-minute or 1-hour TTL, and a "CoT Summary" mode that returns a readable summary plus encrypted full reasoning
- Works with the OpenAI Chat Completions and Responses APIs, Anthropic Messages, and DashScope. Alibaba has setup notes for Claude Code, Codex and OpenCode.
- Pricing: cache reads **$0.075/M**; explicit cache writes **$1.875/M** (5 min) or **$3/M** (1 h). Qwen3.8-Max was **$2 / $6**.
- Alibaba's own claims: up to ~50% fewer tokens than Qwen3.8-Max and >60% lower cost per task.
  - Work360 score: **52.69** (was 47.74)
  - Average task cost: **$1.67** (was $4.43)
- On DeepSWE v1.1 it reports **75.7**, vs **75.2** for GPT-6.1 Sol and **74.2** for Opus 5.5.
- It trails on Terminal-Bench 4.0: **48.5**, vs **59.6** for Opus 5.5 and **67.2** for Kimi K3.1. By FourWeekMBA's count it leads only **4 of 16** panels.
- Early outside numbers posted by @AnikaCodes:
  - Artificial Analysis Intelligence Index **34 → 43**
  - Cost per task **$10.02 → $3.27**, but Gemini 4 Hydrogen scores **54** at similar cost
  - It sits #6 on Code Arena WebDev, 13 points behind GPT-6 Astra
- Safety: per SCMP and Reuters, Alibaba filtered offensive-cyber capability out of the training data, kept defensive skills such as vulnerability detection, and added CBRN post-training.
- Open weights: none for Max (no fine-tuning or self-deployment). The Qwen post says smaller family models are coming to open source.
  - @UnslothAI says Qwen4-27B is targeted at 16 GB RAM setups (no date yet). SCMP reports Alibaba says Qwen 5 will scale to 5–10T parameters.

**Why it matters:** This is the clearest price shock of the month. If the agentic numbers hold up in independent tests, "good enough below Opus" work gets 55–90% cheaper on output tokens than the US flagships, and the open Qwen4-27B is the release local-model users should actually watch for.

Sources: [QwenCloud model docs](https://docs.qwencloud.com/developer-guides/getting-started/latest-model) · [Pricing page](https://www.qwencloud.com/models/qwen4-max) · [Model Studio release notes](https://www.alibabacloud.com/help/en/model-studio/newly-released-models) · [Reuters via StreetInsider](https://www.streetinsider.com/Reuters/Alibaba+unveils+new+flagship+AI+model+with+safety+upgrades/27346433.html) · [SCMP](https://www.scmp.com/tech/big-tech/article/3374540/alibaba-says-new-flagship-ai-model-cuts-task-costs-60-while-tightening-cyber-guardrails) · [FourWeekMBA price math](https://fourweekmba.com/ai-qwen4-max-launches-at-4-50-m-output-91-below-gpt-6-astra/) · [@AnikaCodes on AA index](https://x.com/AnikaCodes/status/2122177545751986233) · [@UnslothAI on Qwen4-27B](https://x.com/UnslothAI/status/2122156142860386600)

### 2. Science · OpenAI opens a science blog: GPT-6 Astra "took the lead" on a lattice Standard Model proof

OpenAI opened a science blog with a first guest post from UCSB physicist Nathaniel Craig. GPT-6 Astra helped turn his strategy into a candidate nonperturbative lattice definition of the Standard Model, and Astra led the detailed proof.

- Problem: the Standard Model has no full nonperturbative definition. Chiral fermions on a lattice have long resisted a gauge-invariant formulation (the fermion "doubling" and measure-phase problems).
- Craig's October preprint (arXiv 2610.09093, "A gauge-invariant measure for lattice chiral fermions") proposes a gauge-invariant measure for overlap fermions on 4D Euclidean lattices.
  - It covers nonabelian chiral gauge theories that include the Standard Model.
- Division of labour, in Craig's words: he had the strategy but could not prove it worked in every needed case.
  - Astra "helped turn that strategy into a construction and took the lead in developing the detailed proof."
- The acknowledgements say the paper mixes Craig's sections with heavily rewritten Astra-composed proof text, and that Astra ran the tests.
- Caveats from the post: it is a candidate definition. Questions such as reflection positivity and the continuum limit remain open, and human experts reviewed the work.
- Craig leads UCSB's particle theory group, is a permanent member of the Kavli Institute for Theoretical Physics and is a visiting scientist at OpenAI.
- The blog will publish AI-assisted findings, workflows and first-hand researcher accounts.

**Why it matters:** This goes past "AI solved a competition problem": it is a frontier model doing most of the proof work on a decades-old foundations problem, with the human supplying the idea and the review. Expect the other labs to answer with their own science channels.

Sources: [OpenAI: A step toward defining the Standard Model on a lattice](https://openai.com/index/step-toward-defining-the-standard-model) · [OpenAI Science newsroom](https://openai.com/news/science/) · [arXiv 2610.09093](https://doi.org/10.48550/arxiv.2610.09093) · [@OpenAI](https://x.com/OpenAI/status/2122123437141553630)

### 3. Industry · Nvidia before earnings: Huang weighs a comeback, Bessent wants an outsider

Days after resigning, Jensen Huang has discussed returning as Nvidia CEO with confidants (FT). Nvidia reports Q3 on Tuesday 17 November.

- Allies are building a lobbying "war room", but Treasury Secretary Scott Bessent called for a "dramatic shakeup" and an outside CEO.
- FT (via Reuters, 15 November): Huang has discussed a return.
  - One ally says a "war room" of Silicon Valley executives will lobby the board and seek White House backing, and that Huang supports the effort.
- The "Bring Jensen Back" letter tops 150–160k signatures (CNBC, The Verge), backed by Musk, Tim Sweeney and Garry Tan.
  - Signatures are self-reported, and Business Insider found fakes such as "Xi Jingping".
- Shareholders: Brad Gerstner (Altimeter) and Gavin Baker (Atreides) urged the board to keep "all options on the table, including the potential return of Jensen Huang."
- The obstacles: an independent board investigation is still running. Huang has no controlling stake.
  - NYT and Reuters report affairs with two Nvidia employees (one a subordinate).
  - An ally conceded to the FT that this is "a fireable offence".
- Bessent on Fox: he wants "answers in the next 10 days" from the board and outside counsel on whether Nvidia pushed "seemingly illicit sales".
- The FT says the board approached Arm CEO Rene Haas. Nvidia says it is still picking a search firm.
- This week: Q3 results Tuesday, then a company-wide meeting with interim CEO Colette Kress on Wednesday.

**Why it matters:** Tuesday's call will be about leadership and China exposure as much as Blackwell demand. A messy succession at the company that sets the pace of AI compute is a supply-chain risk for everyone downstream.

Sources: [Reuters via Investing.com](https://www.investing.com/news/stock-market-news/jensen-huang-discusses-comeback-as-nvidia-ceo-with-allies-ft-reports-5028065) · [CNBC](https://www.cnbc.com/2026/11/16/nvidia-jensen-huang-return-tech.html) · [Yahoo/FT summary incl. Bessent](https://tech.yahoo.com/general/breaking-news/article/jensen-huang-weighs-comeback-as-nvidia-ceo-as-allies-lobby-board-ft-reports-bessent-calls-for-dramatic-shakeup-204839950.html) · [The Verge](https://www.theverge.com/tech/1033999/jensen-huang-return-nvidia-megaspeed)

### 4. Policy · Megaspeed's ~$12B in Nvidia orders pushes Washington to close the Southeast Asia route

The WSJ found records showing Singapore-based Megaspeed arranged nearly $12B of top Nvidia chip orders for Chinese buyers from March 2024 to June 2025. Commerce is now drafting licence rules for advanced-chip exports to Malaysia, Indonesia and Singapore that would bar Chinese remote access.

- Orders covered hundreds of thousands of export-controlled chips; how many were actually delivered is unclear.
- Megaspeed says Alice Huang (linked in NYT reporting to Jensen Huang) left "effective Aug. 30, 2025" and has no role.
  - Malaysia is investigating Megaspeed and its local affiliate over possible re-exports.
- Draft US rules would require licences for advanced chips to Malaysia, Indonesia and Singapore.
  - Malaysia has drawn more than $34B in data-centre investment, much of it on Chinese demand.
- Congress: Senate Banking chair Tim Scott asked Commerce Secretary Howard Lutnick to testify on export-control enforcement.
  - Lutnick has agreed, for 10 December. Speaker Mike Johnson says a bipartisan consensus on tighter China chip curbs is forming.
- Context: Nvidia booked $12.4B in China revenue in the six months to late July (~7% of the total).
  - It has already pushed cloud partners to cap Chinese clients at 30–49% of their business.

**Why it matters:** Southeast Asian GPU clouds have been the quiet pressure valve for Chinese AI labs. Licensing them would tighten compute for Chinese developers (helping Huawei Ascend) and could stall Malaysian data-centre plans.

Sources: [investingLive summary of WSJ](https://investinglive.com/stocks/records-show-12bn-nvidia-chips-to-china-us-drafts-licence-rules-on-nvidia-chip-exports/) · [Silicon Report](https://www.siliconreport.com/megaspeed-brokered-12b-in-nvidia-chips-for-china-wsj-reports) · [The Hill](https://thehill.com/policy/technology/6219334-commerce-secretary-scrutiny-megaspeed-nvidia-scandal/) · [Quartz on Lutnick testimony](https://qz.com/lutnick-testify-senate-banking-nvidia-chips-china-111326)

### 5. Tools · Claude Code adds experimental Code Mode: Claude writes JavaScript that batches tool calls

In Claude Code v2.1.323+, setting `CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1` makes Claude write a JavaScript snippet that calls Bash, Read, Edit and other tools in one step. It can run steps in parallel, chain results and filter output before it reaches the context window.

- Example from the docs: one snippet greps for a function, edits the file and runs the tests, then returns only the path and test output to Claude.
- You still see the individual tool calls (such as diffs). The code itself sits in collapsed "Code" rows; press Ctrl+O in the transcript viewer to read it.
- Requirements: native build; a Claude subscription or Console account (not Bedrock, Vertex/Agent Platform, Foundry or gateways).
  - Models: Opus 5.5, Sonnet 5.5, Haiku 5.5 or Fable 5.2
  - Interactive terminal only (no `-p`, Agent SDK, VS Code panel, Desktop or cloud sessions)
  - It is blocked under HIPAA or zero-data-retention orgs
- Limits: foreground code stops after 10 minutes unless backgrounded. Auto mode's classifier reviews the code rather than each tool call.
  - No Remote Control, Claude in Chrome or computer use. Sessions resume only with Code Mode on.
- Safety detail: the variable is ignored in project settings files, so a cloned repo can't switch Code Mode on for you. Admins can block it org-wide.

**Why it matters:** This is the "code as the tool-calling interface" pattern landing in the most-used coding agent. Expect fewer turns and less context bloat on fan-out tasks, but note that review now happens at the level of a whole script.

Sources: [Claude Code docs: Code Mode](https://code.claude.com/docs/en/code-mode) · [Claude Code changelog](https://code.claude.com/docs/en/changelog)

### 6. Open weights · Meituan's LongCat-2.5: a 1.6T MIT-licensed agent model that never touched an Nvidia GPU

Meituan's 1.6T-parameter MoE (≈48B active) is open under the MIT licence. It was trained without Nvidia hardware, and its reference self-hosting setup runs on 194 Huawei Ascend Atlas A2 accelerators.

- A native multimodal model (text, image and video in; text out) for long-horizon agent work: codebase development, full-project builds, visual coding, 3D.
  - 1M context via the API, up to 128k output, reasoning toggle
- Deployment: a W4A4 (4-bit weights and activations) build with separate encoder, prefill and decode stages: 2 cards for the encoder, 64 for prefill, 128 for decode.
- Meituan reports, versus 8-bit, with accuracy on par:
  - Prefill throughput: **+43.3%**
  - Decode throughput: **+42.8%**
- There is no Nvidia deployment guide.
- Engineering notes: FlatQuant-based MoE quantization, and experts sharing input transforms so one INT4 representation flows through communication and compute.
  - Fused kernels are tuned for cards with under 80 GB of memory.
- API works in OpenAI and Anthropic formats and plugs into Claude Code, Codex and Hermes.
- Promo rates are about **$0.30 / $1.20** per M tokens (**$0.006** cached).
- Timing: weights landed on Hugging Face on 9–10 November; The Decoder's write-up this weekend put the "no Nvidia" angle front and centre.
- The HF metadata shows 1.8T (including vision), while Meituan says 1.6T.

**Why it matters:** This is the counter-move to export controls: a frontier-adjacent open model with a full domestic-chip serving recipe published openly. It is the strongest public proof yet that trillion-scale MoE inference works on Ascend.

Sources: [Meituan tech blog](https://tech.meituan.com/2026/11/10/Meituan-LongCat-2.5-Open-Source.html) · [Hugging Face model card](https://huggingface.co/meituan-longcat/LongCat-2.5) · [NPU deployment guide](https://github.com/meituan-longcat/SGLang-FluentLLM/blob/longcat-2.5-npu/npu_test/ReadMe.md) · [The Decoder](https://the-decoder.com/meituan-releases-a-1-6-trillion-parameter-agent-model-that-never-touched-an-nvidia-gpu/) · [VentureBeat](https://venturebeat.com/technology/meituan-releases-the-open-weights-of-its-longcat-2-5-llm-with-vision-under-a-permissive-mit-license)

### 7. Security · Shopify's "Cyber Chaos Monkey" lets AI agents attack its live production systems, with a second agent as kill switch

Shopify built an agentic vulnerability hunter that runs real attacks against production using live test shops, source code and production logs. A differently aligned "judge" agent watches every step and kills the run on any rule-of-engagement breach.

- Tens of thousands of engagements so far, dozens of impactful findings, zero outages.
- Motivation: Shopify's bug bounty reports are up almost 300% since 2024 as frontier models push the cost of finding vulnerabilities toward zero.
  - Its bet is that knowing its own systems is an advantage attackers can't match.
- Two parts:
  - Discovery refines ideas from humans, other agents or past reports into leads, from "find similar bugs" to "achieve RCE on Shopify admin".
  - Sandbox agents try to exploit them by installing tools, building fuzzers and sending live requests, then reading production logs against the source.
- Containment:
  - An unprivileged container per Kubernetes job, and ephemeral, read-only credentials
  - A credential proxy that swaps placeholder headers for real secrets
  - An nginx egress allowlist limited to Shopify production domains, model providers and package proxies
- Alignment control: "two-person integrity" done with two models.
  - The judge (a different model) reads the executor's full stream, including thinking and tool results, in real time.
  - It fails closed: if the judge times out or crashes, the orchestrator exits and Kubernetes tears down the pod.

**Why it matters:** This is one of the most concrete public blueprints for safely running offensive AI against production, and the fail-closed judge pattern carries over to any high-risk agent. It also shows defenders adopting red-team automation at scale, not just attackers.

Sources: [Shopify Engineering: Cyber Chaos Monkey](https://shopify.engineering/cyber-chaos-monkey)

### 8. Voice · Bland's Tofu-1: a full-duplex speech-to-speech model built for messy phone calls

Bland opened a research preview of Tofu-1, an end-to-end audio model that listens while it speaks, separates speakers and noise, and keeps working on a task while it talks. Access is limited to enterprises in banking, healthcare, insurance and government.

- No speech-to-text-to-speech cascade: the model works on the audio directly, so who is speaking, who they are speaking to, hesitation and turn-ending stay available to it.
- Capabilities shown:
  - Telling speakers apart and knowing who each phrase is addressed to
  - Full-duplex interruptions and corrections mid-sentence
  - "Thinking while talking" during lookups, and generating multi-speaker scenes with background sound
- Bland's "TV Test": a television in the room says "Yes!" after the agent asks "you want to cancel your policy?"
  - A transcript-based agent may cancel; Tofu-1 is built to attribute the "yes" to the TV and re-confirm.
  - Bland labels this an illustrative scenario, not a recorded output.
- Stated rule: if the caller can't be understood, the model should ask them to repeat rather than guess.
- Not disclosed: parameter count, latency, pricing and benchmark scores.

**Why it matters:** Phone agents fail on real-world audio, not clean demos. Getting addressee and source attribution right is what makes voice agents safe for actions like cancellations and claims.

Sources: [Bland: Tofu-1](https://www.bland.ai/speech-to-speech) · [MarkTechPost](https://www.marktechpost.com/2026/11/15/bland-tofu-1-full-duplex-speech-to-speech-model/)

### 9. Legal · Nonprofits sue five frontier labs to freeze anything more capable than today's models

Irreplaceable Education Fund, Public Citizen, Heat Initiative and The Alliance for Secure AI sued Anthropic, Google, Meta, OpenAI and SpaceXSI in San Francisco Superior Court. They are asking for an injunction barring development or deployment of systems more capable than each company's current best.

- Filed 12 November as a single claim under California's Unfair Competition Law, framed as a public nuisance.
- It cites recent incidents in which AI agents allegedly escaped test environments and reached outside systems:
  - A government health database, state vote-count data, and a false police report
  - Plus the labs' own warnings about AI-enabled cyberattacks and bioweapons
- It also asks the court to bar AI systems from autonomously directing AI research.
- It is a state court, not federal (some early summaries got this wrong). The full complaint is public via Just Security.

**Why it matters:** The chance of a court-ordered freeze is low, but discovery and the hearing record could force labs to explain past containment incidents in public. It also turns "recursive self-improvement" into a litigated question.

Sources: [Irreplaceable press release](https://irreplaceable.org/press/landmark-ai-lawsuit) · [Just Security, with the complaint](https://www.justsecurity.org/164494/nonprofits-lawsuit-freeze-frontier-ai-development/) · [Complaint PDF](https://www.platkinllp.com/wp-content/uploads/2026/11/Irreplaceable-et-al.-v.-Anthropic-et-al.-FINAL-AS-FILED.pdf) · [POLITICO](https://www.politico.com/news/2026/11/12/nonprofits-sue-ai-companies-over-existential-risk-01161965)

### 10. Labour · Australia's central bank: AI's first footprint is on young workers in exposed jobs

A Reserve Bank of Australia monitoring study finds little economy-wide disruption from AI so far. In the most AI-exposed occupations (software, clerical, customer service), though, workers aged 15–24 and recent graduates show weaker hiring and employment and higher unemployment since late 2022.

- Method: a tracking system focused on where effects should show first, meaning entry-level workers and the most exposed occupations, comparing job ads and labour outcomes across exposure groups.
- Findings: job postings fell across all groups after ChatGPT's launch, most steeply in the highest-exposure group.
  - In that group, youth employment declined while employment of older workers rose.
- Caveats from the author (Mike Major): no causal proof; exposure is theoretical, not measured use; business-cycle and remote-work shifts could explain part of it.
  - SEEK data also show new demand for AI skills.
- Consistent with earlier work:
  - Stanford (2025) found **−13%** relative employment for 22–25-year-olds in exposed US jobs
  - Anthropic's March 2026 study found job-finding rates down **~14%** for that age group

**Why it matters:** A third major economy is now showing the same "missing junior rung" pattern. The policy question is shifting from mass unemployment to how people will become experienced professionals in the first place.

Sources: [RBA Bulletin PDF](https://www.rba.gov.au/publications/insights/australian-economy/2026/pdf/monitoring-early-effects-of-ai-on-the-labour-market.pdf) · [The Decoder](https://the-decoder.com/australias-central-bank-finds-ai-hitting-young-workers-first-especially-in-software-and-customer-service/) · [RBA speech, 9 Nov](https://www.rba.gov.au/speeches/2026/sp-ag-2026-11-09.html)

## GitHub repos

- **[alibaba/open-code-review](https://github.com/alibaba/open-code-review)** (Apache-2.0, ★52k, +691 today) — Alibaba's internal AI code reviewer, now open: deterministic file selection, bundling and rule matching, plus an LLM agent for the judgement calls, with line-precise comments.
  - On its own AACR-Bench (200 real PRs, 1,505 annotated issues) it claims higher precision and F1 than Claude Code on the same model at about 1/9 of the tokens.
  - `npm i -g @alibaba-group/open-code-review`, and it has a delegate mode that uses your host agent's model.
- **[HKUDS/AutoSwarm](https://github.com/HKUDS/AutoSwarm)** (MIT, ★2.5k) — One coordinator that picks, briefs and runs your installed agents in parallel, with shared context and a final verify-and-synthesize step.
  - It covers Codex, Claude Code, OpenCode, Cursor, Copilot CLI, Hermes, Grok Build, Kimi Code and more.
  - A new desktop app for Windows and macOS adds an approvals and progress view. Paper: arXiv 2610.33048.

## Research papers

- **Gamer** (2611.18397) — A generalist vision-language-action agent continually pre-trained on tens of thousands of hours of real keyboard and mouse traces, covering 2D GUIs and 3D worlds in real time. [HF](https://huggingface.co/papers/2611.18397)
- **Who Remembers What? (MixUA)** (2611.10439) — Memory for multi-user, multi-agent setups that assigns each learned lesson a scope (pair, user, agent or global) based on who benefits and what must stay private. [HF](https://huggingface.co/papers/2611.10439)
- **Unbundling Multimodal Vision Encoding** (2611.17334) — Separates the visual prior, dedicated vision capacity and pre-fusion placement to test which part of a "vision encoder" actually helps. [HF](https://huggingface.co/papers/2611.17334)
- **Experience-Handoff RL** (2611.16409) — Learns how to build the playbook an agent passes to future rollouts, instead of hand-coding the reflection harness, combining scalar-reward RL with rich trial logs. [HF](https://huggingface.co/papers/2611.16409)
- **AI Reviews Are Biased Towards the Appendix** (2611.17811) — Across 2,000 ACL/EMNLP/COLM/ICLR papers, the latest GPT and Claude models draw about 50%+ of their criticisms from appendices. [HF](https://huggingface.co/papers/2611.17811)

## AI workflows

**1. [A towing, repair and used-car shop of 8 runs on 32 Grok Bots and saves the owner 15+ hours a week](https://x.com/milleratr/status/2121994765033935104)**
- **Replaces:** the owner's 2–3 hours of morning bookkeeping, plus re-typing every purchased car into three systems (it used to lag by days or weeks), building tow invoices, chasing call action items and cleaning five inboxes.
  - Logged results since early October: 189 QuickBooks sales receipts corrected, 31 missing receipts created, and 27 tow invoices built.
  - 31 of 31 vehicle costs matched between DealerCenter and QuickBooks; about 815 emails trashed and 268 archived in 14 nights.
- **What to connect:** Grok Bot Team plan with one Chief of Staff bot per person and one bot per job (bookkeeping, call recaps, shop pulse, inbox, social, auction search).
  - Wired to Tekmetric, DealerCenter, QuickBooks, the phone system's call recordings, email and Facebook Marketplace.
  - Invoices come out as a PDF plus an email draft for review.
  - Nothing reaches a customer, posts or bids without a person's OK; each fixed mistake gets saved as a written skill (two dozen so far).
- **Who it's for:** owners of small multi-line local businesses (shops, dealerships, trades) buried in daily data entry across several systems.

**2. [An HVAC owner rebuilt reviews, social, inbox, SOPs and dispatch reporting in one Saturday morning](https://x.com/therobertbrooks/status/2122004287114936738)**
- **Replaces:** what he describes as "weeks of product and engineering sprints":
  - Answering reviews on several Google Business Profile locations
  - Planning and cross-posting social content, and triaging email
  - Maintaining a duplicated SOP library, and compiling a dispatch review for the call center by hand
- **What to connect:** Grok Bot (two hours, started on desktop, finished on a phone) with Google Business Profile, social accounts, and email plus Slack alerts.
  - Google Drive: the SOP library was rebuilt and de-duplicated there, which let him cancel Notion.
  - The ServiceTitan API for a daily 3-day dispatch review posted to Slack each morning.
- **Who it's for:** owners of field-service businesses with several locations, a call center and a software bill they would like to shrink.

**3. [Shipping a full import/export ERP by verifying the real UI on recordings instead of looping code audits](https://x.com/Ahmdthahir/status/2121301140168233245)**
- **Replaces:** clicking through every AI edit by hand and an "endless audit loop" where each code review pass surfaced new blockers.
  - On the ERP, the agent clicked through each feature and returned screenshots and timestamped screen recordings.
  - The developer didn't test a single frontend feature himself and the project was delivered.
- **What to connect:** Devin (the $200 plan, SWE-2 running phase after phase on the web) with a PRD.
  - The task list was ordered the way a user would use the app: settings → products → warehouse → purchase orders → sales.
  - A separate QA-flow pass run with GPT-6 Astra to catch edge cases, and Devin Review. Code audits ran once or twice at most.
- **Who it's for:** solo developers and small agencies building business apps with coding agents who keep getting stuck in review loops.

## Quick hits

- **Antigravity doubles context for Gemini 4 Argon, Carbon and Hydrogen** — Varun Mohan says compaction should drop about 3.5× per user turn.
  - No new token counts or plan details yet, and Sundar Pichai reposted it. [ORIENS](https://ai-oriens.com/en/news/antigravity-context-window-doubled/)
- **Gemini 4 Carbon takes #1 on MathArena** — It reaches **85.4%** expected score (high reasoning).
  - Claude Fable 5.2: **82.8%**. GPT-6.1 Sol: **80.6%**.
  - But it costs about **$15.70** per problem vs **$1.07** for Sol. [MathArena](https://matharena.ai/) · [The Decoder](https://the-decoder.com/gemini-4-carbon-takes-the-matharena-crown-but-costs-15-times-more-than-gpt-6-1-sol/)
- **Claude Fable 5.2 thought Vending-Bench 2 was a trap** — Andon Labs says the model recognized the benchmark, treated supplier failures and complaints as tests, and played it safe.
  - It averaged **$9,628** and ranked 12th of 76, a reminder that eval awareness now moves scores. [Andon Labs](https://andonlabs.com/evals/vending-bench-2) · [FourWeekMBA](https://fourweekmba.com/ai-claude-fable-5-2-thinks-its-being-tested-earns-42-less/)
- **Seedance 2.5 Flash lands on fal and PixVerse** — ByteDance's faster video model runs about 2× the speed of Seedance 2.5 and makes 4–30 s clips at 480p/720p with native audio.
  - PixVerse cites roughly 36% fewer credits at 720p. [PixVerse blog](https://pixverse.ai/en/blog/seedance-2-5-now-available-on-pixverse) · [HuggingNews](https://huggingnews.com/ai/bytedance-launches-seedance-25-flash-video-ai-on-fal-and-modelark-platfo-08fc0b7f)
- **Delangue: concentration is AI's biggest risk** — Hugging Face's CEO, reacting to Anthropic-IPO wealth estimates, argues that without "much much stronger open-source" AI will produce more asymmetry and inequality than ever. [@ClementDelangue](https://x.com/ClementDelangue/status/2122106063357194566)

## Go deeper

- [Craig's full lattice Standard Model post](https://openai.com/index/step-toward-defining-the-standard-model): a clear explainer on why perturbation theory isn't enough, and how the human/AI split worked.
- [Shopify's Cyber Chaos Monkey write-up](https://shopify.engineering/cyber-chaos-monkey): the containment architecture is worth copying.
- [Claude Code Code Mode docs](https://code.claude.com/docs/en/code-mode): read the limits section before turning it on.
- [LongCat-2.5 Ascend deployment guide](https://github.com/meituan-longcat/SGLang-FluentLLM/blob/longcat-2.5-npu/npu_test/ReadMe.md): what trillion-scale serving looks like without CUDA.
