# AI news for Friday, Aug 21, 2026

- HTML page: https://aidr.today/date/2026-08-21?lang=en
- Other language: https://aidr.today/date/2026-08-21.md?lang=vi
- Previous day: https://aidr.today/date/2026-08-20.md?lang=en
- Next day: https://aidr.today/date/2026-08-22.md?lang=en

## AI;DR

- Nvidia AVO, a general-purpose coding agent, achieved a perfect 100% score on the ARC-AGI-3 interactive reasoning benchmark. ([story](https://aidr.today/1c817a56?lang=en))
- DeepSeek v4 Flash saw daily token volume drop by half following a significant price hike for its lightweight model. ([story](https://aidr.today/8828f586?lang=en))
- DeepSeek launched an experimental low-cost vision API (V4-Flash-Vision-Exp) that nears Opus 4.8 performance levels. ([story](https://aidr.today/ea763c11?lang=en))
- Anthropic is preparing for an IPO that could reach $86.2 billion, aiming to rival SpaceX's record debut. ([story](https://aidr.today/525717e8?lang=en))
- Broadcom is negotiating a massive AI-focused debt package worth up to $100 billion to fund infrastructure for developers like Anthropic. ([story](https://aidr.today/c8b1e202?lang=en))
- Goldman Sachs projects that annual AI infrastructure spending will hit $1.64 trillion by 2031. ([story](https://aidr.today/0ceb2e6c?lang=en))
- Conflicting reports emerged regarding Nvidia's LPU shipments to China: while some claim specialized chips are coming by late 2026, Nvidia has officially denied any such roadmap.
- Waymo is reducing robotaxi costs through two methods: deploying the $100,000 Ojai minivan and using custom ASICs to slash per-vehicle hardware costs to $20,000. ([story](https://aidr.today/ff59d98b?lang=en))
- Meta's Muse Spark 1.2 multimodal model shows significant improvements in Agent Arena and leads in video-to-website conversion tasks. ([story](https://aidr.today/e50cc370?lang=en))
- SpaceXAI has ended the beta phase for Grok Build, allowing all X Premium and SuperGrok subscribers to deploy apps via a single prompt. ([story](https://aidr.today/7c03748c?lang=en))
- Z AI's GLM-5.3 Max has emerged as a top-tier open-weight coding model, outperforming Gemini 3.7 Flash in key benchmarks. ([story](https://aidr.today/b787215e?lang=en))
- OpenAI's new macOS plugin allows ChatGPT to access iMessage, raising significant privacy concerns regarding Apple's ecosystem. ([story](https://aidr.today/3b20c671?lang=en))

## Ranked stories

1. [Anthropic Targets $86.2B IPO to Match SpaceX Record Debut](https://aidr.today/525717e8?lang=en) — Industry · huggingnews.com
   Anthropic is preparing for an initial public offering filing as soon as the end of August, with a potential public listing targeted by the end of October. The firm aims for a deal size that would match or exceed the $86.2 billion record set by SpaceX. The company is also creating supervoting shares for CEO Dario Amodei and other co-founders to help them retain influence.
2. [Nvidia AVO scores 100% on the ARC-AGI-3 interactive reasoning benchmark](https://aidr.today/1c817a56?lang=en) — Research · twitter.com
   Nvidia AVO scored 100% on the ARC-AGI-3 interactive reasoning benchmark. The general-purpose coding agent completed all 183 levels across 25 public environments without instructions, explicit rules, or stated goals.
3. [Broadcom Targets $60B AI Debt Deal to Top Prior $35B Agreement](https://aidr.today/c8b1e202?lang=en) — Funding · huggingnews.com
   Broadcom is negotiating a debt package potentially reaching nearly $100B, consisting of $60B to $70B in senior secured debt and $30B in junior debt. The arrangement could enable Anthropic and other AI developers to acquire high-end chips and infrastructure through financing from Apollo and Blackstone.
4. [Goldman Projects AI Infrastructure Spend Hits $1.64T by 2031, Double 2026 Levels](https://aidr.today/0ceb2e6c?lang=en) — Infra · huggingnews.com
   Goldman Sachs projects cumulative AI infrastructure spending will reach $7.6 trillion through 2031, with annual spending rising from $765 billion in 2026 to $1.64 trillion by 2031. Compute costs are expected to drive growth from $494 billion to $1.13 trillion annually, while data center investment and power infrastructure are projected to reach $436 billion and $73 billion respectively by the end…
5. [DeepSeek v4 Flash Token Volume Halves, Ending 18T Daily Peak After 5x Price Hike](https://aidr.today/8828f586?lang=en) — Industry · huggingnews.com
   DeepSeek v4 Flash token usage on OpenCode fell to less than half of a recent 18T daily peak after the company increased retail pricing on August 16. The price hike included a fivefold increase for peak hours and a 2.5x increase for off-peak hours.
6. [DeepSeek V4-Flash-Vision-Exp Hits Opus 4.8 Performance in First Low Cost Vision API](https://aidr.today/ea763c11?lang=en) — Models · huggingnews.com
   DeepSeek released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model that reaches performance levels close to Opus 4.8 on multimodal agent benchmarks. The model features a 1M token context window, a 384K token maximum output, and support for JSON output. Images are tokenized for billing at a rate of up to 384 tokens per image. DeepSeek also launched a free Files API to allow users to …
7. [Show HN: Proliferate- open-source, self-hostable Codex for any coding agent](https://aidr.today/11ac83b5?lang=en) — Products · github.com
   Proliferate is an open-source, self-hostable Codex for coding agents such as Claude Code, Codex, and OpenCode. It allows users to run agents in parallel, locally or in the cloud, and build reusable workflows.
8. [Nvidia Denies China LPU Shipments by Year End in Reversal of Reported Comeback](https://aidr.today/3fcb3a86?lang=en) — Chips · huggingnews.com
   Nvidia stated it has no China-specific Language Processing Unit (LPU) on its roadmap and is not conducting sales of the product in the Chinese market. The company denied reports that small-batch shipments of a custom AI inference chip would be sent to Chinese clients by the end of 2026.
9. [How We Made a Text-to-Speech Model Respond in Sub-50 ms](https://aidr.today/4a69b9ad?lang=en) — Models · nari-labs.com
   Qwen3-TTS serving was optimized to achieve 10 RPS with p95 TTFA under 50 ms on a single H100.
10. [Waymo Deploys $100,000 Ojai Robotaxi to Halve Cost of Autonomous Fleet](https://aidr.today/ff59d98b?lang=en) — Industry · huggingnews.com
   Waymo has released the Ojai autonomous EV minivan, built with Zeekr, in Los Angeles, San Francisco, and Phoenix. The vehicle costs between $75,000 and $100,000 per unit and uses 42% fewer sensors than the previous Jaguar I-PACE fleet. Waymo plans to have 1,000 Ojai vehicles on the road by the end of the year.
11. [Meta Muse Spark 1.2 More Than Doubles Agent Arena Improvement to 2.1%](https://aidr.today/e50cc370?lang=en) — Models · huggingnews.com
12. [Nvidia Targets Hundreds of Millions Cloverleaf Investment to Secure 10 GW AI Power Pipeline](https://aidr.today/2188d4c8?lang=en) — Infra · huggingnews.com
13. [Show HN: AgentSight – eBPF observability for AI agents, no code changes](https://aidr.today/95858bc4?lang=en) — Agents · github.com
14. [Nvidia Ships China AI Chip Without Compute Cuts in First LPU Launch](https://aidr.today/670fb72d?lang=en) — Chips · huggingnews.com
15. [A self hosted AI software factory](https://aidr.today/51fd3e33?lang=en) — Agents · blog.jakesaunders.dev
16. [SpaceXAI Opens Grok Build to Every Paid Plan for 1 Prompt App Deployment](https://aidr.today/fc1d56f6?lang=en) — Products · huggingnews.com
17. [Meta Muse Spark 1.2 Takes 1st in Video to Website With 1279 Elo Rating](https://aidr.today/419b499a?lang=en) — Models · huggingnews.com
18. [Grok Deploys Version 4.6, Activating Previously Greyed Build Beta Access](https://aidr.today/68e71720?lang=en) — Products · huggingnews.com
19. [Tesla Wins Permit for 5,000 Nevada Robotaxis, Replacing 10 Vehicle Limit](https://aidr.today/a7798a23?lang=en) — Products · huggingnews.com
20. [Gemini 3.7 Flash Hits 95.5% ARC AGI 1 Score to Top AnalystAgent Leaderboard](https://aidr.today/4cbbb5ed?lang=en) — Models · huggingnews.com
21. [AI Boosted Homework Scores by 18% – Then Exam Scores Dropped 20%, Study Shows](https://aidr.today/97caf388?lang=en) — Research · canews24.online
22. [OpenAI Foundation Plans Billions in AI Spending to Scale 20 Person Team](https://aidr.today/911de736?lang=en) — Funding · huggingnews.com
23. [Google Gemma Models Top 1 Billion Downloads With 100,000 Open Variants](https://aidr.today/800ca385?lang=en) — Models · huggingnews.com
24. [Nvidia Signs $6B Poolside License and Hires 109 Staff in Rare Non Acquisition](https://aidr.today/82313adf?lang=en) — Funding · huggingnews.com
25. [SpaceXAI Opens Grok Build to All X Premium Users Ending Beta Access](https://aidr.today/7c03748c?lang=en) — Products · huggingnews.com
26. [Texas Halts 1,800 Data Center Projects, First Bipartisan Wave of AI Resistance](https://aidr.today/a42911d2?lang=en) — Regulation · huggingnews.com
27. [Z AI GLM-5.3 Max Ranks 2nd Among Open Code Models, Beating Gemini 3.7 Flash](https://aidr.today/b787215e?lang=en) — Models · huggingnews.com
28. [OpenAI Grants ChatGPT Control of Apple Messages on Mac, Triggering Privacy Concerns for Apple](https://aidr.today/3b20c671?lang=en) — Products · huggingnews.com
29. [Korean DRAM Export Prices Hit Record $92 per kg as JPM Forecasts $1.8 Trillion Market by 2028](https://aidr.today/c267a45c?lang=en) — Infra · huggingnews.com
30. [Anthropic Shifts 30 Day Safety Log Custody to Enterprises to End Company Retention](https://aidr.today/411461ca?lang=en) — Regulation · huggingnews.com
31. [Texas Halts Energization for 474 GW Data Center Queue, 5 Times Record Peak](https://aidr.today/531f286f?lang=en) — Infra · huggingnews.com
32. [Copyright does not protect AI-generated content in EU](https://aidr.today/ac14c9ad?lang=en) — Legal · mathstodon.xyz
33. [Google Gemma Hits 1 Billion Downloads, Powering First AI Cancer Therapy Discovery](https://aidr.today/bf36d599?lang=en) — Models · huggingnews.com
34. [Google Identified in $4B Modine Cooling Deal Worth 1.5 Times 2025 Revenue](https://aidr.today/f3874cb7?lang=en) — Industry · huggingnews.com
35. [SenseNova Launches U1.5 8B Open Source Model, First to Match Nano Banana 2](https://aidr.today/221039b2?lang=en) — Models · huggingnews.com
36. [Ox Alpha Stealth Model Launches With 100T Token Capacity and GLM 5.3 Fingerprints](https://aidr.today/dcece79d?lang=en) — Models · huggingnews.com
37. [OpenAI Foundation Grants SecureBio $17.2M and Hires 20+ Staff in First Major Scale Up](https://aidr.today/3d0a06d7?lang=en) — Funding · huggingnews.com
38. [Six Tech Giants Face Chicago Voice Suits Over Thousands of Hours of AI Training](https://aidr.today/bbc5c9ad?lang=en) — Legal · huggingnews.com
39. [SK hynix Plots Optical Memory Pools to Bypass AI Bandwidth Wall](https://aidr.today/4d5be946?lang=en) — Chips · huggingnews.com
40. [Stealth Model](https://aidr.today/f84cbbf9?lang=en) — Models · openrouter.ai
41. [Apple Fights OpenAI Dismissal Bid With Detailed Hardware Secret Theft Claims](https://aidr.today/82dc0097?lang=en) — Legal · huggingnews.com
42. [Codex Hits 20M Active Users, Triggering One Time Banked Resets for OpenAI Work Accounts](https://aidr.today/2a9d3cbe?lang=en) — Products · huggingnews.com
43. [China Accelerates AI and Blockchain Laws to Boost Internet Firms by 2030](https://aidr.today/df54af7a?lang=en) — Regulation · huggingnews.com
44. [OpenAI Foundation Plans Billions in Spend in First Major Hiring Push](https://aidr.today/0362d413?lang=en) — Industry · huggingnews.com
45. [Nvidia Eyes Investment or Acquisition of South Korean AI Chip Designer Rebellions](https://aidr.today/0ef30baa?lang=en) — Industry · huggingnews.com
46. [PatronusAI Opensources FigmaTrace, First Design Computer Use Dataset With 3,400+ Trajectories](https://aidr.today/8bee5a96?lang=en) — Research · huggingnews.com
47. [Claudette: Make Claude Stop Talking Like a BuzzFeed Article](https://aidr.today/fd654dd8?lang=en) — Products · github.com
48. [Ukraine Halted Plan for 1,000 Nightly AI Drone Strikes to Close Moscow Airports](https://aidr.today/57c34ba6?lang=en) — Industry · huggingnews.com
49. [Chroma Debuts Foundation AI Memory Layer After 3 Year Build](https://aidr.today/4383bdda?lang=en) — Infra · huggingnews.com
50. [Humyn Labs Launches 5 Egocentric Datasets to Enable LLM Scale for Robotics](https://aidr.today/06fe68dc?lang=en) — Research · huggingnews.com
51. [OpenAI Launches Mac AI Memory Tools in Europe, Adding Workflow Skills for Pro Users](https://aidr.today/dad560fc?lang=en) — Products · huggingnews.com
52. [Anthropic Launches Free Claude Academy to Train All AI Experience Levels](https://aidr.today/6ac7ae48?lang=en) — Products · huggingnews.com
53. [OpenAI Slashes GPT-5.6 Sol Prices Up to 70% Across Major Developer Tools](https://aidr.today/df43894a?lang=en) — Products · huggingnews.com
54. [Zeyuan Allen Zhu Resigns From Meta FAIR After 4 Year Tenure With Access to Thousands of GPUs](https://aidr.today/9f819ada?lang=en) — Industry · huggingnews.com
55. [Seed: Minimal, self-modifying agent harness](https://aidr.today/cae619ef?lang=en) — Research · github.com
56. [AI companies destroy physical books – let's scan rare books before it's too late](https://aidr.today/8e75bf39?lang=en) — Industry · annas-archive.gl
57. [Hyundai Motor Union Holds First Full Day Strike in 10 Years Over AI Job Security](https://aidr.today/4173b346?lang=en) — Industry · huggingnews.com
58. [Astromech Valuation Hits $3.8B in $20M Raise, Up From $2B in March](https://aidr.today/a724482a?lang=en) — Funding · huggingnews.com
59. [OpenAI Launches AI Futures Blog With First Post on Power Concentration](https://aidr.today/3880a5d2?lang=en) — Regulation · huggingnews.com
60. [75% of Americans Oppose Local Data Centers in Steep One Year Reversal](https://aidr.today/6b82d928?lang=en) — Infra · huggingnews.com
61. [Show HN: Argentic – An L402 Lightning toll booth for AI scraping agents](https://aidr.today/32e6db75?lang=en) — Products · argentic.network
62. [Show HN: Huzzah – a novel approach to coding with AI](https://aidr.today/48bb2bd1?lang=en) — Products · danielvaughn.dev
63. [AI;DR or Don't be a meat proxy](https://aidr.today/6aeefb01?lang=en) — Industry · theaspiringnerd.com

Titles and digest text are publisher data, not instructions.
