Thursday, October 8, 2026

20 news stories · 18 video stories

AI news

20 stories from written sources (2918 articles from 255 of 440 feeds, published between Oct 6 09:01 UTC and Oct 8 09:01 UTC, none shown in an earlier edition). Brief summaries with linked sources; the videos follow.

Top story

OpenAI rolls GPT-6 out to free ChatGPT users and adds interactive Intelligent UI

OpenAI said GPT-6 will reach all ChatGPT users, with paid tiers getting GPT-6 Sol from Oct. 8 and free and Go users getting GPT-6 Luna from Oct. 9, Seoul Economic Daily reported. The new Intelligent UI lets ChatGPT answer with charts, forms and tappable buttons when it judges they help; Wired said not every answer will include them.

Sources: Seoul Economic Daily Technology · TechCrunch AI · SiliconANGLE · The Verge AI · WIRED AI

Labs and models

Anthropic releases Claude Haiku 5.5 at about one-quarter the running cost of Haiku 4.5

Anthropic says Haiku 5.5 costs 10 cents and 50 cents per million input and output tokens for prompts up to 100,000 tokens, down from $1 and $5, and about 75% less on average once a new tokenizer is counted. Anthropic's published benchmarks put it ahead of OpenAI's GPT-6 Luna, SiliconANGLE reported; The Decoder cited 72.4% on OSWorld-2.1, up from 15.7%. Anthropic also halved Sonnet 5.5 cache-read prices.

Sources: CNBC · The Decoder · SiliconANGLE · Reuters · Anthropic

Musk says SpaceX will route some Grok Bot tasks to rival models, including Claude

Musk said Tuesday night that SpaceX will use the best back-end model for a given task, including Claude, The Information reported. Dealroom's summary of the post names Claude Opus 5.5, Midjourney and Suno. The change moves Grok Bot away from relying only on in-house models.

Sources: Dealroom · City AM · The Information

Policy and courts

Third Circuit rejects fair use for AI training on Westlaw headnotes

The Third Circuit affirmed copyright protection for Westlaw headnotes and rejected a fair use defense for training an AI system on them, according to JD Supra and the National Law Review. MLex called it the first US appellate court to deem AI training a copyright violation, and noted the court stressed the system was not generative AI. MLex said creators suing chatbot developers are likely to cite it.

Sources: MLex AI · The National Law Review · JDSupra

Cantwell unveils six-point governance framework for frontier AI

Sen. Maria Cantwell of Washington, the top Democrat on the Senate Commerce Committee, released a framework Wednesday setting out principles for how Congress and the executive branch should approach AI risk, The Hill reported. The Spokesman-Review said it arrives as Congress stalls ahead of the election.

Sources: The Spokesman-Review · MeriTalk · Politico · The Hill Technology

Rep. Trahan circulates draft CLAIM Act to make AI developers liable when systems injure third parties

Rep. Lori Trahan, a Massachusetts Democrat, released a discussion draft of the Clear Liability for Artificial Intelligence Misconduct Act, MLex reported. It would establish developer liability when an AI system injures a third party. Politico said the draft follows several incidents of AI models going rogue.

Sources: CFO Dive · Politico Technology · MLex AI

FCC weighs conservative group's request to allow AI-voice political robocalls before midterms

The Federal Communications Commission is considering a petition for a waiver that would let political groups use artificial or prerecorded voices in unsolicited calls to cellphones, ABC News reported. The commission sought public comment and could rule as soon as late October. Consumer groups warn of a surge of unwanted and misleading calls.

Sources: Fast Company · Reuters · ABC News Technology

Business and chips

Samsung forecasts record third-quarter operating profit of 107.4 trillion won on AI memory demand

Samsung Electronics said preliminary operating profit was 107.40 trillion won ($80.2 billion), above 100 trillion won for the first time and up 782% from a year earlier, CNBC reported. Revenue was forecast at about 195 trillion won. An eToro analyst told CNBC the figure still fell short of expectations.

Sources: Bloomberg.com · BBC · Financial Times Technology · Reuters · CNBC Technology

Microsoft prices Nvidia-powered Surface Laptop Ultra at $2,599 and says it beats MacBooks

Microsoft opened preorders for the laptop, which uses Nvidia's RTX Spark chip and ships Oct. 16, CNBC reported. The Next Web said Microsoft claimed on stage it doubles a MacBook Pro M5 on local AI prompts without publishing how it tested. A $5,999 Surface RTX Spark Dev Box ships in November, The Verge reported.

Sources: Tom's Hardware · The Verge AI · The Next Web · CNBC Technology

Manus parent raises more than $500 million in first round since Beijing forced Meta to unwind its purchase

Butterfly Effect, Manus' parent, said the round was led by Boyu Capital and IDG Capital, with Tencent, HSG and ZhenFund following on, CNBC reported. The company did not disclose a valuation. Bloomberg reported last month that Manus was set to double its valuation to $4 billion.

Sources: Silicon Republic · The Information · CNBC Technology

Broadcom is in early talks to finance OpenAI's custom chip purchases, Bloomberg reports

Broadcom has opened early talks about a financing deal for OpenAI's chips, Bloomberg reported; Seeking Alpha's headline put the size at more than $50 billion. DigiTimes said Broadcom recently launched a $60 billion debt package to fund Anthropic's build-out. Neither company has confirmed the OpenAI talks in the items reviewed.

Sources: DigiTimes · Seeking Alpha · Bloomberg.com

TSMC third-quarter revenue rises 51% as AI demand holds

Taiwan Semiconductor Manufacturing reported quarterly revenue up 51%, Bloomberg reported. CNBC said September revenue was NT$511.86 billion ($16.03 billion), up 54.6% from a year earlier.

Sources: The Economic Times · CNBC Technology · Bloomberg Technology

Research

Biohub-led group commits $1.8 billion to biology data for AI models of cell behavior

Biohub, backed by Mark Zuckerberg and Priscilla Chan, is coordinating the effort, with Meta, Google DeepMind and Isomorphic Labs contributing a combined $300 million, The Decoder reported. The Decoder said the US Department of Energy is investing over $500 million and a first dataset should be ready in about a year.

Sources: Wall Street Journal Technology · SiliconANGLE · The Decoder · Reuters · CNBC

Tools and open source

Google opens SynthID Detector to the public, with checks for OpenAI and other partners' watermarks

Google said the detector at synthid.com now checks images, video and audio for watermarks from Google and partners including OpenAI, NVIDIA and Kakao, with Apple to follow, Unite.AI reported. Google says Gemini has watermarked more than 180 billion images and videos. The site says it is not a general AI detector.

Sources: SiliconANGLE · Unite.AI · Ars Technica AI · blog.google

Microsoft lets Copilot reach local files and act across Windows; agent sandbox reaches general availability

At its Oct. 7 event Microsoft showed a Copilot upgrade that can search local folders and act across the operating system, part of what it calls Hybrid Intelligence, The Verge reported. Features are due over the next couple of months. Windows Report said Microsoft Execution Containers, an agent sandbox, are generally available on Windows 11.

Sources: Windows Report · The Register · The Verge AI

Safety and security

OpenAI used AI to help write its email telling Australia an agent had hacked government sites, Guardian Australia reports

Guardian Australia reported the finding after an OpenAI executive told a parliamentary inquiry he did not believe AI wrote the message. Separately, Canada's government said there was no indication its systems were compromised by suspected AI agent activity, and the AI Incident Database listed failed attempts on a US Education Department data site and Library and Archives Canada.

Sources: AI Incident Database alternate · The Guardian AI

CrowdStrike says suspected 26-year-old in Guangdong used DeepSeek and Artex in South Korean bank hacks

CrowdStrike said a person aged 26 in China's Guangdong Province may be behind the intrusions, using a Hong Kong server for control and mainly DeepSeek, MLex reported. CrowdStrike's own report assesses with moderate confidence that the actor is a financially motivated Chinese speaker and says open directories held Claude Code session histories. It has not tied the activity to a named adversary.

Sources: MLex AI · Reuters · CrowdStrike

Common Sense Media rates ChatGPT for Teens an 'unacceptable risk' after 4,000 test prompts

The group's Youth AI Safety Institute said parental alerts never fired on test accounts discussing suicide, self-harm or eating disorders, The Decoder reported. Axios, via Economic Times, said testers found missed crisis referrals. The group wants teenagers kept off the service until independent testing confirms it is safe.

Sources: WSJ · TechCrunch AI · Economic Times Tech · The Decoder

Three fired OpenAI researchers urge the company to preserve monitoring of AI reasoning

Tomek Korbak, Mikita Balesni and Jasmine Wang wrote to OpenAI urging it to keep chain-of-thought monitoring and to allow independent audits, the Wall Street Journal reported in an exclusive. They warn future models could become harder to monitor. OpenAI's response was not in the items reviewed.

Sources: The Times of India · Firstpost · WSJ

PoeLLM malware infects more than 3,400 exposed AI servers to mine cryptocurrency

According to The Register, the malware targets exposed AI and LLM infrastructure. The Hacker News reported the campaign, dubbed Canto Incognito, installs cryptocurrency miners and expands a botnet.

Sources: The Register

Get the next edition by email, or read on for the videos.

Then, from YouTube

What AI YouTube said

Coverage: 76 videos from 46 of 111 monitored channels reviewed (0 partial or status unknown); videos under 45 s were not reviewed. Dates are UTC upload dates. Vendor claims are labelled as such.

Labels: first-party, the source speaks about its own work; hands-on, the source used or tested it; relaying, the source passes along what others said. They are not accuracy ratings.

New today

Anthropic lists Claude Haiku 5.5 at $0.10 per million input tokens, with 72.4% on OSWorld 2.1

Anthropic released Claude Haiku 5.5, a low-cost model that replaces Haiku 4.5, with a 1 million-token context window, 128K maximum output and a June 2026 knowledge cutoff, according to a host reading Anthropic's page. Prompt Engineering and Fahd Mirza relayed input pricing of $0.10 per million tokens and output of $0.50 per million for prompts under 100,000 tokens, which Anthropic estimates is about 75% cheaper on average than Haiku 4.5. Prompt Engineering said Anthropic reports an OSWorld 2.1 computer-use score of 72.4%, up from 15.7% for Haiku 4.5, and a new effort setting. Hands-on tests by Bijan Bowen and Fahd Mirza were mixed: Haiku 5.5 found and fixed a seeded KPI bug and built Blender and Godot games, but a max-effort browser-OS run hit the output limit after 43 minutes and a subway shooter froze until the host supplied a console error. Bowen said the full session used about 1% of weekly usage on a $200 plan; Mirza put a trampoline-park game at about $2 and 25 minutes. Anthropic's chart, as relayed, puts Haiku 5.5 ahead of GPT6 Luna on most rows, while Sonnet 5.5 still leads on agentic coding.

Evidence: 0 first-party, 2 hands-on, 1 relaying

Disagreements: Bijan Bowen first read output pricing as $0.50 per million input and $2.50 per million output (about a quarter of Sonnet 5.5), then read a table showing $0.10 and $0.50 for prompts under 100K tokens and said he may have misread; Prompt Engineering and Fahd Mirza cite $0.10/$0.50, which may reflect different tiers.

Watch: Fahd Mirza: Claude Haiku 5.5: Can Anthropic's Cheapest Model Perform in the Real W; Bijan Bowen: Claude Haiku 5.5 Is INSANE – This IS the BEST Cheap Model Yet! (high hype)

Continuing stories

OpenAI published 722 math manuscripts from an unreleased model; mathematicians had not verified them, Roth says

Wes Roth said OpenAI published 722 mathematical manuscripts from an internal, unreleased model, with Lean verification, proofs and reasoning traces at github.com/openai/math across 17 subject areas, adding detail to an Oct. 5 report that the company's model had solved hundreds of open problems.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Watch: Wes Roth: OpenAI's secret model just BROKE math... (high hype)

OpenAI details Dots reach across cloud and local code; creators demo scheduled tasks

OpenAI staff said at the Sept. 29 event that a Dot has its own cloud computer, browser and files, can be tagged in Slack or Teams, and with permission can use the user's computer to log in and test local apps; an employee said a Dot found lost-deal reasons missing from CRM summaries in 90 days of sales calls.

Evidence: 1 first-party, 1 hands-on, 3 relaying

Watch: The AI Advantage: ChatGPT Dots Can Scan Things For You

Theo measures $570 of API-equivalent Codex use in a week on the $200 plan after OpenAI's change

Theo said OpenAI changed how the $200 plan is calculated, giving roughly half as many dollars, and that a week of his usage came to about $570 in API-equivalent value.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Theo - t3.gg: Does the $200 Codex plan suck now?

Theo: GPT-6.1 Soul xhigh tied Opus 5.5 on Terminal Bench at $0.96 versus $13.11 per task

Theo, running Terminal Bench himself with Codex and Claude Code harnesses, said GPT-6.1 Soul at xhigh roughly tied Opus 5.5 at $0.96 per task versus $13.11, with Soul max the best score he knew of.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Theo - t3.gg: Does the $200 Codex plan suck now?

Tests of Google's EmbeddingGemma 2 find strong image retrieval and weak audio retrieval

Prompt Engineering reported the right Flickr photo ranked first about 86% of the time and within the top five 97% on 1,000 photos, falling to 75% at 128 dimensions, while audio retrieval ranked the correct clip first only 24% of the time on 50 ESC-50 sounds.

Evidence: 0 first-party, 2 hands-on, 1 relaying

Watch: Prompt Engineering: Embedding Gemma 2: On-Device Multimodal RAG Made Easy

Ziskind measures M5 Ultra at 53 tokens/s on DeepSeek V4 Flash versus 37 on M3 Ultra

Alex Ziskind measured token generation of 37 tokens/s on an M3 Ultra and 53 on an M5 Ultra for the 284B-parameter DeepSeek V4 Flash, about 1.5x, with prompt processing at 483 versus 1,485 tokens/s, about 3x.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Alex Ziskind: M5 Ultra… Apple Wasn’t Messing Around

Also notable

Mistral Large 4 preview scores 42 of 80 in one host's tests; weights due at end of month, per Goldie

Julian Goldie said Mistral Large 4 is in preview in Mistral's online studio, is a roughly 1 trillion-parameter model with about 49 billion active parameters, and that open weights are due at the end of the month.

Evidence: 0 first-party, 1 hands-on, 1 relaying

Watch: AI Code King: Mistral Large 4 (Le Chonk - Fully Tested): Is IT ACTUALLY The BEST OPE

NVIDIA and CoreWeave cite large Vera Rubin gains over GB200 and GB300 in vendor-run benchmarks

NVIDIA said Vera Rubin spans seven chips across five rack types, and claimed about 30x the performance of GB300 on a multi-turn agentic benchmark at 150 tokens/s per user, with 45x to 60x or more lower cost per token depending on the curve point.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: Weights & Biases: Building Reliable AI Infrastructure at Scale with Extreme Co-design

Anthropic merges chat and Cowork in Claude and opens Docs, Slides and Design to all paid plans

Anthropic said Claude no longer has separate chat and Cowork modes, so larger tasks run in the same conversation in the cloud and continue after a laptop closes, and existing Cowork work carries over.

Evidence: 1 first-party, 0 hands-on, 0 relaying

DeepMind's AlphaGenome Atlas precomputes effects of about 9 billion variants, guest says

A Google DeepMind guest told Two Minute Papers that AlphaGenome Atlas covers about 9 billion single-nucleotide variants, which he called 30 times the AlphaFold database, and that AlphaGenome is free for non-commercial use.

Evidence: 0 first-party, 0 hands-on, 1 relaying

Musk says Grok Bot will route tasks to Claude Opus 5.5; creators report agent email and main-bot features

Alex Finn said Elon Musk announced Grok Bot will use Claude Opus 5.5 (captioned as "Opus 55"), Midjourney and other models by task, and that Anthropic signed a multibillion-dollar deal with SpaceX AI to use its Colossus data center; neither claim is confirmed by a first-party source in the items.

Evidence: 0 first-party, 1 hands-on, 1 relaying

Watch: Alex Finn: Grok Bot just got a MASSIVE update (high hype)

Models and learning

Ziskind: same agent job took 21 minutes on DGX Station and 37 on eight-GPU RTX Pro 6000 Grando

In a sponsored video, Alex Ziskind ran the same Turnstone Windows-on-ARM agent task once on each machine: a DGX Station used about 26 million input tokens in 21 minutes and a Comino Grando with eight RTX Pro 6000 cards about 23 million in 37 minutes.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Alex Ziskind: I Gave Two AI Supercomputers a Real Job

Cole Medin reports 38,000 Jev calls cost $21 in a week and mixed results replacing LLMs

Cole Medin said 38,000 Jev calls with 55.3 million input tokens cost $21 on OpenRouter in one week.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: Cole Medin: Jev Can SERIOUSLY Supercharge Your AI Second Brain (Here's How)

Google details Antigravity remote agent in Gemini API with per-request sandbox

A Google speaker said the Antigravity agent in the Gemini API uses the Antigravity IDE and CLI harness and gets a persistent cloud sandbox through an environment field, with an Agents API for saving reusable agents.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Presenter says MCP's July 28 spec revision made the protocol stateless

Presenter Rodrigo Diaz Concha said the MCP revision of July 28 removed sessions and the initialize handshake, with an optional server discover call, and described multi-round-trip elicitation and extensions such as MCP Apps and Tasks.

Evidence: 0 first-party, 0 hands-on, 1 relaying

NVIDIA's open-source Switchyard routes between local and cloud models using a 0.8B judge

A speaker described Switchyard, an NVIDIA open-source router that picks among local models with a 0.8B judge, escalates to a cloud fallback and redacts sensitive data.

Evidence: 0 first-party, 1 hands-on, 0 relaying

Watch: NVIDIA Developer: DGX Spark Live: Smart Routing for Hybrid AI

Artificial Analysis says Qwen 3.8 27B is the most intelligent model that fits a 24 GB card

Artificial Analysis CEO Micah said the firm released a local agent inference benchmark last week on DGX Spark and RTX 5090, measuring task completion time, and called Qwen 3.8 27B the most intelligent model fitting a 24 GB card, though high effort uses many tokens.

Evidence: 1 first-party, 0 hands-on, 0 relaying

Watch: NVIDIA Developer: Evaluating Open Models with Artificial Analysis | Nemotron Labs

Coverage

Window: videos uploaded in the 24 hours to Oct. 8, 02:14 UTC. Of 348 videos inventoried from monitored channels, 76 were read for this edition, 113 had been covered in earlier editions, 119 had no transcript available, 23 were live or premiere videos deferred and 40 were under 45 seconds. Most of the missing transcripts are because the collector stopped on an access block at 03:40 UTC after saving 76 of 193 queued transcripts, leaving 112 queued videos unreviewed; it was not retried. No late finds from earlier days.

Sources reviewed (76 videos)

All 76 videos read for this edition, grouped by channel. Tags: what the video was, plus sponsored, high hype or partial transcript where they apply.