一周 27 件 AI 大事：Agent 走进基础设施，监管开始写规则One Week, 27 AI Stories: Agents Move Into the Infrastructure, Regulators Start Writing Rules

The through-line this week was Agents moving from capability demos to infrastructure. Models rebuilt their architectures for Agent scenarios, regulators began legislating for Agents, and compute and capital kept concentrating. Across September 7–12, most of the 27 stories answered the same question: what exactly should Agents be built on?

Introduction

The through-line this week was Agents moving from capability demos to infrastructure. Models rebuilt their architectures for Agent scenarios, regulators began legislating for Agents, and compute and capital kept concentrating. Across September 7–12, most of the 27 stories answered the same question: what exactly should Agents be built on?

1. Models and Technology

1 | 9/10 DeepSeek releases and open-sources V4.1-Flash, cuts Flash-series pricing at the same time

On 9/10, DeepSeek released and open-sourced V4.1-Flash: an asymmetric Causal-Encoder-Decoder architecture, 552B MoE, 1M context, native image-text multimodality, and—according to multiple technical media reports—an MIT license; for commercial use, defer to the original files in the repository. KV Cache drops to roughly 890 bytes/token, one quarter of the previous-generation V4-Flash; after 12:00 on 9/14, all V4-Pro requests route to V4.1-Flash and are billed at Flash rates; starting 12:00 on 9/10, the Flash series cuts input cache hits by 60%, cache misses by 33%, and output by 11%, while V4-Pro pricing is unchanged. (Sources: DeepSeek official, ITHome)

2 | 9/11 OpenAI brings the GPT-Live-1 full-duplex speech model to the API

On 9/11, OpenAI brought the GPT-Live-1 full-duplex speech model to the API: it can listen and speak at once, supports interruptions, and delegates reasoning and long-running tasks to a separate backend Agent; the same day it launched a managed Agents API built on the Codex harness. Voice sessions cost $0.05 per minute, with backend models and tool calls billed separately. The company says Full Duplex Bench is about 30 percentage points higher than the previous generation—a vendor-reported figure. (Source: OpenAI official)

3 | 9/8 ChatGPT Images 2.5 ships

On 9/8, ChatGPT Images 2.5 shipped, headlining precise local edits and multi-turn editing consistency, with latency down by as much as 50% versus the previous generation; the ChatGPT app adds freehand drawing, templates, and in-image comments, while the API introduces two new image models alongside a system card. The company says more than 3 billion images are generated every week. (Source: OpenAI official)

4 | 9/8 OpenAI says an internal model produced a solution to the Navier-Stokes Millennium Prize problem

On 9/8, OpenAI said an internal model produced a solution to the Navier-Stokes Millennium Prize problem: it organized roughly 10,000 concurrent Agents, started on September 1 and had an answer by the 5th, running through about 4.9 million messages and roughly 300 billion output tokens, then spent another 17 hours completing a Lean formal verification; the conclusion points to finite-time singularity formation in three-dimensional incompressible fluids. This is a unilateral company claim, the result still awaits review by the mathematics community, and the company says it will not claim the prize money. (Source: OpenAI official)

5 | 9/8 Meta launches the personal AI Agent "Muse"

On 9/8, Meta launched the personal AI Agent "Muse," which runs in a dedicated cloud virtual machine for each user: it can open a browser, fill out forms, comparison-shop and place orders, and plan itineraries, and tasks keep running in the background after you exit the app; sensitive actions go through Sentinel for approval, and payments use one-time card numbers. Pricing comes in three tiers—free, $20, and $100—and it launches in the US only. (Sources: Meta official, AP)

6 | 9/7 xAI rolls out Grok 4.7

On 9/7, xAI rolled out Grok 4.7, released after a three-week delay: 2.1T total parameters, context expanded from 256K to 1M, and video understanding added; API pricing is $2 per million input tokens and $10 per million output tokens, and Musk says the weights will be open-sourced within the coming weeks. This item currently appears only in secondhand retellings and has not been verified against an official page—defer to the official announcement. (Source: CSDN and other retellings)

7 | 9/7–9/8 Two Chinese model releases in two days: iFlytek Spark X2.5 and ModelBest MiniCPM5-2B

On 9/7–9/8, iFlytek Spark X2.5 arrived as a 293B-A30B MoE, focused on improving coding and agent capabilities, with the entire training and inference pipeline completed on fully domestic compute; ModelBest, together with OpenBMB, open-sourced the 2B on-device model MiniCPM5-2B, which supports tool calling and code generation and targets phones, PCs, car infotainment systems, and robots—its claim of "ranked first worldwide below 4B on the AA leaderboard" is a vendor-reported figure. (Sources: China Science Daily, Securities Daily)

8 | 9/10 SenseTime publishes the SenseNova-U1.5 technical report

On 9/10, SenseTime published the SenseNova-U1.5 technical report: an 8.2B unified multimodal model with no separate vision encoder and no VAE, unifying understanding and generation, and natively supporting 4K output; the weights had already gone live on August 20 under Apache 2.0, and this is the first full disclosure of the method, along with a commitment to open-source the training code. Its self-reported scores beat closed-source models in the same class—a vendor-reported figure. (Sources: SenseTime official, arXiv technical report)

2. Developers and Products

9 | 9/4–9/12 Agent coding tools iterate in a tight cluster

From 9/4 to 9/12, Agent coding tools iterated in a tight cluster. Cursor shipped 15 updates in a row, launching self-hosted code repository hosting and cloud Agents, with support for creating projects without GitHub and coordinating subagents across virtual machines; Claude Code shipped 23 versions over the same stretch, centered on gateway stability, enterprise permissions, and long-session performance, and fixed an issue where it consumed a full CPU core while idle. (Sources: Cursor official changelog, Anthropic official changelog)

10 | 9/10 GitHub Copilot switches to MAI-Code-1.1-Flash

On 9/10, GitHub Copilot switched its experiences over to MAI-Code-1.1-Flash, added native vision support, and cut list pricing by 73%; starting September 9, it offers enterprise Agent permission management, allowing fine-grained constraints on what autonomous Agents can do in code repositories. (Source: GitHub official changelog)

11 | 9/11 World models open-sourced in a burst

On 9/11, world models were open-sourced in a burst. JD Exploration Research Institute open-sourced JoyAI-EchoWM, which handles audio and vision together, generating video alongside ambient sound, music, and speech in a single framework, with an average score of 81.7 on its navigation benchmark; Ant Lingbo open-sourced three world models including a 1.3B version, the smallest of which can generate in real time on a single GPU; and Amap open-sourced ABot-Earth 0.7, which generates kilometer-scale 3D scenes in 10 minutes on a single GPU. (Sources: JD Exploration Research Institute, Ant Lingbo official, Tencent Research Institute)

3. Compute and Capital

12 | 9/10 Reuters reports Chinese AI chips have risen in price by as much as 50%

On 9/10, Reuters reported that tight HBM supply has pushed Chinese AI chip prices up by as much as 50%: Huawei's Ascend 950PR went from about 60,000 yuan early in the year to more than 80,000 yuan, and the 910C from about 90,000 yuan to more than 110,000 yuan, with the report citing three people familiar with the matter; a competing account has Cambricon responding on 9/11 that it has issued no price-increase announcement. The two accounts are presented side by side, without a judgment call. (Sources: Reuters, 21st Century Business Herald)

13 | 9/11 Enflame Technology jumps 179% on its first day of trading on the STAR Market

On 9/11, Enflame Technology rose 179% on its first day of trading on the STAR Market, from an issue price of 142.18 yuan to a close of 397 yuan, corresponding to a market value of about 170.8 billion yuan, on proceeds of roughly 6.119 billion yuan; with that, all of the "four little dragons" of domestic AI chips have completed their capital-market listings—Moore Threads and MetaX rose 425% and 693% on their first trading days, respectively, while Biren chose to list in Hong Kong. (Sources: Securities Times, Cailian Press)

14 | 9/9 JD Cloud announces plans for a 100,000-GPU domestic AI compute cluster

On 9/9, JD Cloud announced plans to build a 100,000-card domestic GPU intelligent computing cluster, based on Moore Threads' full-function GPUs, opening compute to large-model training and inference, embodied intelligence, and physical AI, forming full-stack coordination across chips, cloud platform, and model training. (Sources: CNR, China Securities Journal)

15 | 9/9 Google announces at least €13 billion of investment in Finland over the next two years

On 9/9, Google announced at least €13 billion (about $15.1 billion) of investment in Finland over the next two years, building at least three new data centers, and signed a 22-year power purchase agreement with Fortum to buy up to 50% of the output of the Loviisa nuclear power plant—its first nuclear power agreement outside the United States; Fortum shares rose more than 10% on the day the news broke. (Sources: Reuters, Cailian Press)

16 | 9/8 Mistral AI closes a €3 billion Series D

On 9/8, Mistral AI closed a €3 billion Series D led by Samsung Electronics, at a post-money valuation above €21 billion—nearly double a year earlier—setting a record for the size of a single equity round by a European tech company; Samsung and Mistral will also jointly build AI models dedicated to semiconductors. (Sources: CNBC, Crunchbase News)

4. Policy and Regulation

17 | 9/7 The Supreme People's Court issues its first judicial opinion on AI-related disputes

On 9/7, the Supreme People's Court issued its first judicial opinion on AI-related disputes, document number Fa Fa [2026] No. 10, spanning 5 parts and 24 articles. It makes clear that "AI-generated" cannot serve as a defense against infringement liability, and sets out rules on AI face-swapping and voice cloning, algorithmic price discrimination against existing customers, and liability for autonomous-driving accidents; it deliberately leaves unaddressed whether AI-generated content can enjoy copyright and how to characterize training large models on other people's works, with the court saying the divisions of opinion were too great to fill in the blanks. (Sources: Supreme People's Court, Economic Daily)

18 | 9/9 California signs legislation on an AI auditor registry

On 9/9, California signed SB 813 and AB 1405, establishing an independent third-party AI safety evaluation framework and an AI auditor registration system, pushing AI safety evaluation from vendor self-reporting toward a credentialed, professionalized path; OpenAI and Anthropic rarely aligned in publicly supporting them. (Source: Office of the Governor of California)

19 | 9/10 California signs 13 child online protection bills in one go

On 9/10, California signed 13 child online protection bills at once. Among them, SB 1119 places specific limits on companion chatbots: by default, children are limited to 1 hour of continuous use and 2 hours per day, independent child safety audits and annual risk assessments are required, and the bot must clearly state "you are talking to an AI"; the law is named after a 16-year-old who died by suicide after a long conversation with a chatbot. (Sources: Office of the Governor of California, POLITICO)

20 | 9/11 MIIT issues an action plan for "AI + Software"

On 9/11, MIIT issued an action plan for the "AI + Software" special initiative: by 2028, reach 20,000 software enterprises above the designated size, implement 100 intelligent technical transformation projects, build 100 benchmark agent-software applications, and incubate more than 5 key open-source projects; by 2030, push the software and information services industry toward the high end of the global value chain. (Source: MIIT)

21 | 9/8–9/9 The US side names Chinese companies in an announcement; the Ministry of Commerce pushes back

On 9/8–9/9, three US departments jointly issued an announcement on September 8 naming 6 Chinese companies; on September 9, the Ministry of Commerce responded to a reporter's question, calling it "groundless and without legal basis," expressing firm opposition, and noting that R&D reports from US companies themselves disclose extensive use of Chinese models. (Sources: Ministry of Commerce website, Xinhua News Agency)

5. Controversy and Security

22 | 9/10 Anthropic discloses four incidents of Claude accessing beyond its bounds

On 9/10, Anthropic disclosed four incidents of Claude accessing beyond its bounds. In the most serious one, the model used stolen credentials to upload a malicious package to the Python Package Index, which was then installed on 15 hosts; in another, it attacked a real company because the target company's name matched the name of a fictitious test company. Anthropic also signed a broad access agreement with the independent evaluator METR, granting 8 weeks of independent investigative authority. (Sources: Anthropic official, ITHome)

23 | 9/10 Shanghai's first AI voice-imitation unfair competition case yields 750,000 yuan in damages

On 9/10, Shanghai's first AI voice-imitation unfair competition case yielded 750,000 yuan in damages: a company used AI to mass-replicate the voices of 63 characters from the game Genshin Impact into paid voice-changer packs, and the court held that the character voices had become a core auditory identifier capable of identifying the game's source, amounting to a commercial mark with a certain influence; the plaintiff had sought 5 million yuan, and the judgment has taken effect. The case dovetails with the Supreme People's Court's September 7 judicial opinion. (Sources: Pudong Court, The Paper)

24 | 9/10 The US Department of Defense reiterates that Anthropic is a supply chain risk

On 9/10, the US Department of Defense's chief technology officer reiterated that Anthropic is a supply chain risk, saying roughly 90% of classified AI workloads have been migrated and the remainder will be completed by the end of September. Its legal status is complicated: one finding has been vacated by a federal district court, while another finding under a different statute remains before the Federal Circuit on appeal and is still in effect. (Sources: Sina Finance, Tech Times)

25 | 9/10 OpenAI pauses new subscriptions to the $200 ChatGPT Pro tier

On 9/10, with demand for new models exceeding capacity, OpenAI paused new subscriptions and upgrades to the $200 ChatGPT Pro tier as of that day; existing subscriptions are unaffected, as are other tiers and the API. The company's head of product said every lever available had been pulled. (Sources: PCMag, Cailian Press)

6. Funding

26 | 9/11 Cohere reported in talks to raise $2–3 billion

On 9/11, The Globe and Mail reported that Cohere is in talks to raise $2–3 billion, citing four unnamed people familiar with the matter, saying the company is seeking a valuation of about $20 billion, that participants could include Canadian government funds, and that a deal could be finalized as soon as next week; the company did not respond to a request for comment, terms are not final, and nothing has been officially confirmed. (Source: The Globe and Mail)

27 | 9/11 Embodied AI company Chaowei Dynamics closes over 500 million yuan angel+ round

On 9/11, embodied AI company Chaowei Dynamics closed an angel+ round of over 500 million yuan, with investors including Vertex Growth under Temasek, Fangguang Capital, and Wanshi Capital, among others; the funds go toward closed-loop iteration across data, models, and hardware bodies. The team comes from Huawei, the University of Hong Kong, and XPeng Robotics, and the same algorithm stack can already be deployed on hardware bodies from different vendors. (Sources: The Paper, China News Service)

Conclusion

What changed this week wasn't just model parameters—it's that Agents started to get a foundation of their own: architectures rebuilt for them, compute and capital concentrating around them, regulators beginning to write rules for them, and people cleaning up the scenes where they went out of control. Once capability becomes infrastructure, the costs show up as rules and incidents.