// Frontier · weekly

AI weekly · 2026-W41

2026-10-05 – 2026-10-11 (Beijing time) · Final
AI-for-science went into mass production but exposed a verification gap; GPT-6 went global and small models got cheaper, while an OpenAI revenue gap made markets re-examine AI valuations.

Read this week’s hotspots ↓ Past issues →

2,548 items collected1,692 events10 hotspots

Some older feeds carried titles only; descriptions stick to titles and visible summaries—see the originals for details.

01Hotspots this week

01OpenAI releases ~700 maths manuscripts; its Navier–Stokes proof is soon questioned
Maths·6 outletsVendor claimDisputed
WHAT HAPPENEDOn 6 October OpenAI published results on open maths problems from an internal frontier model, with Lean formalisations and details on GitHub. Commentators count about 722 manuscripts but differ on how many major problems were solved. On 8 October New Scientist reported that the human-readable and machine-checked versions of its Navier–Stokes proof do not match.
WHY IT MATTERSIt is the largest single AI maths release yet, and it shows formal verification can itself be mistranslated—raising the question of how to review AI output at volume.
WHO IT AFFECTSMathematicians and theoretical computer scientists, journals and reviewers, and the formal-proof community.
ORIGINAL SOURCES
OpenAI News · Sharing AI progress in mathematicsNew Scientist Tech · OpenAI mistranslated mathematics into code for its Navier-Stokes proof
4 related reportsComputational Complexity · Open No MoreGil Kalai · Updates: Sharing AI progress on mathematics (amazing!); and my lecture plansLatent Space · [AINews] Quasi-Riemann-Hypothesis: OpenAI publishes 722 math papers solving 90 of the top 500 open math problems; “the most significant moment” in >100 years of mathematicsRisk.net · From Navier-Stokes to manuscripts by the kilo: AI’s impact on quant research
02GPT-6 rolls out to all ChatGPT users with Intelligent UI
General·3 outletsVendor claim
WHAT HAPPENEDOn 7 October OpenAI rolled GPT-6 out globally in ChatGPT with Intelligent UI, which composes answers from text, visuals and interactive components. The same week an Ultrafast mode for GPT-6.1 Sol arrived in the API, Codex and ChatGPT Work.
WHY IT MATTERSA flagship model reaching every user, with a new answer format, changes daily use for hundreds of millions and sets a design bar for rival chat apps.
WHO IT AFFECTSChatGPT users, developers and rival product teams.
ORIGINAL SOURCES
OpenAI News · GPT-6 and Intelligent UI for everyone
3 related reportsOpenAI Forum · GPT-6 and Intelligent UI in ChatGPTOpenAI Forum · Ultrafast is rolling out today for GPT-6.1 Sol in the API, Codex, and ChatGPT Work雷峰网 · 再裁员5000人！保时捷明确回归燃油车，高端车型计划涨价20%；腾讯拟发行50亿美元离岸债券，或筹钱加码AI；OpenAI全面上线GPT-6
03Mistral releases open-weight multimodal flagship Mistral Large 4
Models & software·8 outletsVendor claim
WHAT HAPPENEDOn 6 October Mistral released Mistral Large 4, an open-weight, natively multimodal reasoning model for text and images, pitched at cyber defence, manufacturing and finance. Ars Technica reports Mistral says the model, nicknamed 'Le Chonk', can challenge the best.
WHY IT MATTERSEurope's leading model maker sticks with open weights, giving European firms that need on-prem deployment and data sovereignty a flagship option.
WHO IT AFFECTSEuropean enterprises and public sector, and self-hosting teams.
ORIGINAL SOURCES
Mistral AI News · Introducing Mistral Large 4
6 related reportsars-ai · Mistral says "Le Chonk" can challenge the best AI modelsVercel Changelog · Mistral Large 4 now available on AI GatewayMozilla.ai Blog · Mistral Large 4 is live in OtariSimon Willison · Mistral Large 4Tom's Hardware AI · Mistral’s new Large 4 trails some Chinese open models in independent testsSimon Willison · Introducing Mistral Large 4: Le chonk

Anthropic launches Claude Haiku 5.5 with 1M context from $0.10 per Mtok

Models & software·8 outletsVendor claim

WHAT HAPPENED: On 7 October Anthropic released Claude Haiku 5.5 for high-volume, cost-sensitive work such as summarisation, context compaction, database queries and classification, and as a coding sidekick. Claude Code release notes list 1M context at $0.10/$0.50 per million input/output tokens. Sonnet 5.5 cache-read prices were halved at the same time.
WHY IT MATTERS: Agent systems hand many subtasks to small models, so their price and speed set total cost; GitHub Copilot, AWS and Devin shipped it on day one.
WHO IT AFFECTS: Claude API developers, coding tools and agent platforms.

ORIGINAL SOURCES

Claude Code releases · v2.1.293 SiliconANGLE AI · Anthropic releases Claude Haiku 5.5 small model and halves Sonnet 5.5 cache read prices

6 related reports

Vercel Changelog · Claude Haiku 5.5 now available on AI Gateway GitHub Changelog · Claude Haiku 5.5 in GitHub Copilot AWS Machine Learning Blog · Introducing Claude Haiku 5.5 on AWS Windsurf Blog · Claude Haiku 5.5 is now available in Devin Simon Willison · Claude Haiku 5.5 TLDR Tech · Claude Haiku 5.5 🤖, mathpocalypse 🧮, designer influence 👩🏻‍🎨

Google introduces the Gemini agent for work at Gemini at Work 2026

Business·6 outletsVendor claim

WHAT HAPPENED: On 8 October Google Cloud CEO Thomas Kurian unveiled the Gemini agent at Gemini at Work 2026—a single universal work agent that answers questions, handles knowledge work, creates images and media, and writes and runs code, starting with business customers.
WHY IT MATTERS: Google puts an agent straight into the enterprise workplace, competing head-on with Microsoft Copilot and OpenAI's business products.
WHO IT AFFECTS: Google Workspace and Cloud customers and rival office-software vendors.

ORIGINAL SOURCES

Google Cloud AI/ML · Welcome to Gemini at Work 2026: Introducing the Gemini agent

5 related reports

TechCrunch Enterprise · Google brings agentic AI to Gemini, starting with businesses SiliconANGLE AI · Google Cloud introduces Gemini agent to change enterprise work CBS News Tech · Google unveils Gemini AI agent to handle workplace tasks geekpark · 谷歌云发布 Gemini Agent；Manus 官宣五亿美元融资；小鹏上线 Robotaxi 打车小程序 Product Hunt · Gemini Agent for Google Cloud

Report says OpenAI's annualised revenue is ~$20bn below earlier signals; AI stocks fall

Finance·5 outletsUnconfirmed

WHAT HAPPENED: On 8 October CNBC, TechCrunch and others reported OpenAI's annualised revenue was about $20bn below what had been signalled, and SiliconANGLE said AI stocks fell. Later reports said OpenAI projects a $20bn jump in annualised revenue by year-end and is seeking about $30bn in new capital.
WHY IT MATTERS: OpenAI's revenue anchors the AI-infrastructure investment story—NVIDIA, Oracle and CoreWeave valuations are tied to it. The HN thread passed 430 points.
WHO IT AFFECTS: AI-stock investors, OpenAI's compute suppliers and AI firms raising money.

ORIGINAL SOURCES

Hacker News · OpenAI annualised revenues $20B less than previously signalled

4 related reports

TechCrunch AI · OpenAI’s revenue is reportedly $20 billion less than previously projected SiliconANGLE AI · AI stocks crumble on report that OpenAI’s annualized revenue is much lower than believed PYMNTS · OpenAI Projects $20 Billion Jump in Annualized Revenue Before Year End The Decoder · OpenAI revenue keeps surging as company seeks $30 billion in fresh capital

Manus raises over $500m after China blocked Meta's acquisition

Business·5 outlets

WHAT HAPPENED: On 8 October Butterfly Effect, parent of AI agent Manus, said it had raised more than $500m led by Boyu Capital and IDG Capital, with Tencent, HSG and ZhenFund joining. It is Manus's first round since China blocked Meta's roughly $2bn acquisition.
WHY IT MATTERS: A landmark case of cross-border AI M&A constrained by regulators; Manus turning to domestic capital shows exit paths for Chinese agent companies are changing.
WHO IT AFFECTS: Chinese AI start-ups and their investors, and foreign acquirers of Chinese AI teams.

ORIGINAL SOURCES

TechNode · Manus parent Butterfly Effect completes more than $500 million funding round

4 related reports

CNBC · AI startup Manus raises $500 million in first funding round since Meta breakup Phys.org · AI startup Manus raises $500 million after Meta acquisition blocked by China Decrypt · AI Startup Manus Raises $500 Million After China Nixed Meta’s $2 Billion Acquisition geekpark · 谷歌云发布 Gemini Agent；Manus 官宣五亿美元融资；小鹏上线 Robotaxi 打车小程序

China AI capital: DeepSeek reportedly nears RMB80bn round; Kling AI picks banks for HK IPO

Finance·2 outletsUnconfirmed

WHAT HAPPENED: Citing Bloomberg on 6 October, TechNode said DeepSeek is close to raising at least RMB80bn (~$12bn), with Tencent and CATL among the largest investors, up from an initial target of about RMB50bn. Bloomberg also reported Kuaishou's Kling AI has picked banks for a Hong Kong IPO of at least $1bn.
WHY IT MATTERS: If DeepSeek, long self-funded, takes outside money it reshapes China's model-funding landscape; a Kling listing would test AI-video commercialisation.
WHO IT AFFECTS: Chinese model makers, Hong Kong investors, and Tencent and Kuaishou shareholders.

ORIGINAL SOURCES

TechNode · DeepSeek reportedly nears $12 billion funding round backed by Tencent and CATL TechNode · Kuaishou’s Kling AI reportedly picks banks for Hong Kong IPO targeting at least $1 billion

1 related reports

cnBeta · DeepSeek被曝拟融资超800亿元 OpenAI立下连更28天承诺

Anthropic offers free AI security scans to open source; Google pauses OSS bug bounty over AI slop

AI safety·6 outlets

WHAT HAPPENED: On 8 October Anthropic launched free AI security scans for open-source projects, which can apply for vulnerability reports from its strongest models including Claude Mythos; crypto projects were among the first to apply. The same week Google paused its Open Source Vulnerability Rewards Program after a surge of AI-generated submissions.
WHY IT MATTERS: AI bug-finding both helps defenders and floods maintainers with noise—open source gets new tools and a deluge of reports at once.
WHO IT AFFECTS: Open-source maintainers, security researchers and bug-bounty platforms.

ORIGINAL SOURCES

The Verge AI · Anthropic launches free AI security scans for open-source projects TechCrunch Security · Google froze its open source bug bounty program due to a ‘significant rise’ in AI submissions

4 related reports

The Hacker News · Anthropic Launches Free AI Vulnerability Scanner for Open-Source Projects Cointelegraph AI · Crypto projects apply for Anthropic’s new frontier AI security scanner Infosecurity Magazine · Google Suspends Open-Source Bug Bounty Due to AI Vulnerability Reports The Decoder · Anthropic launches a free AI scanner for open-source projects

South Korea probes possible AI use in bank hacks that exposed ~25,000 customers

Finance·4 outlets

WHAT HAPPENED: On 6 October South Korean President Lee Jae Myung told his cabinet there were signs hackers used AI against several banks, and police opened an investigation; Tom's Hardware reports data on about 25,000 customers was exposed. Insurance Journal says both South Korea and Japan have been hit as AI lowers the bar for attackers.
WHY IT MATTERS: A rare head-of-state acknowledgement of AI in attacks on banks, which may tighten financial regulators' AI-security demands.
WHO IT AFFECTS: Banks and financial regulators, Asian security teams and bank customers.

ORIGINAL SOURCES

NYT · South Korea Investigates Possible Use of A.I. in Hackings on Its Banks

3 related reports

TechXplore · South Korea warns of possible AI use in banking hacks Tom's Hardware AI · Hackers suspected of using AI agents for cyberattacks on South Korean banks, exposing data from about 25,000 customers Insurance Journal · South Korea, Japan Buffeted by Hacks as AI Lowers Bar for Cybercriminals

02Also worth a look

03By field

Models & software724

Science547

General222

AI safety168

Finance133

Business123

Health102

Robotics101

Chips & compute96

Biology76

Items collected this week by field

Models & products
GPT-6 for everyone, Claude Haiku 5.5, Mistral Large 4, Nano Banana 2.1 and Qwen-Image-2.1-Turbo all shipped in one week.
AI × maths & science
OpenAI's batch of maths manuscripts drew excitement and scrutiny; Ai2's work on byte-level language models appeared in Nature.
Safety & security
South Korea's AI bank-hack probe, Anthropic's free scans and Google's paused bounty; Anthropic investigated unintended model actions; Wikimedia reported suspected rogue OpenAI agent activity.
Policy
EU text-provenance rules took effect in practice; USA Today's parent sued OpenAI for over $250m.
Business & funding
OpenAI revenue reports moved AI stocks; Manus, DeepSeek and Kling made capital moves; Crunchbase counted a record number of billion-dollar rounds in Q3; SAP agreed to buy TechWolf.
Chips & compute
Surface Laptop Ultra debuted RTX Spark; SpaceX reportedly sought $40bn in debt for GPUs; Samsung forecast record Q3 profit; Amazon ended secret data-centre pacts and pledged $1bn to host towns.

04Trends of the week

AI research output has gone bulk; review and verification are now the bottleneck.
Cheap small models such as Haiku 5.5 keep cutting the cost of agent subtasks.
Markets are starting to test the gap between AI revenue and valuations.
AI lowers the bar for attackers, straining defences from Korean banks to bug bounties.

05Next week

RTX Spark laptops' retail launch and first reviews.
Per ifanr, Apple's new products could launch as soon as next week—watch for AI hardware.
Follow-ups on Anthropic's IPO and OpenAI's new round.

Method

同一事件聚类（实体加权 Jaccard + 4 天窗口 + 同 URL 合并）→ 打分 = 报道量 3·log2(1+信源数) + 信源权威 + 社区热度(HN/PH/HF 票数) + 影响面(发布/金额/监管/头部厂商) + 多样性(中英双语/多类信源)

Past issues

2026-W43in progress The US banned DeepSeek from federal systems while DeepSeek open-sourced its inference engine and reportedly raised $15B; Anthropic launched the proactive Claude Duet agent, and OpenAI brought its models to Google Cloud and bought Orchestra Bio as labs race for enterprise and industry entry points (week in progress).US orders federal civilian agencies to remove DeepSeek10 2026-W42final Gemini 4 Argon went GA, Meta returned to the frontier with Muse and open weights, Anthropic showed revenue ahead of its IPO, and agents started hitting platforms' permission walls.Google opens Gemini 4 Argon to developers, enterprises and paying users10 2026-W40final The three frontier labs refreshed flagship models in the same week, while a rogue-agent incident led OpenAI to pause training of its strongest model—safety now sets the pace of R&D.Google announces Gemini 4 Argon for coding, enterprise knowledge work and cyber defence10