Updates

Moonshot AI's Kimi-K3 topped the Arena frontend code leaderboard with 1679 points, surpassing Claude Fable 5, while Kimi K3.1 is teased to focus on professional engineering tasks, as Chinese large models keep up the pressure in the coding and Agent tracks [20][2][17]; on the capital side, OpenAI's target valuation for its roughly $30 billion funding round was cut from $1.4 trillion to $1.25 trillion, while 13 U.S. public pension funds have booked $138 billion in paper gains from AI stocks...

Model competition shifts from "stacking compute" to "compute efficiency and post-training": Microsoft MAI-Base-2 claims to surpass the GPT-5.6 backend model using one quarter of OpenAI's compute, DeepSeek V4.1-Flash forms a Pareto improvement with "stronger, faster, shorter outputs," and GLM 5.3 is...

OpenAI is reportedly in talks to raise up to $30 billion at a $1.25 trillion valuation [16], while Grok Bot is rapidly integrating with native mobile apps and 1Password login capabilities, as agents move from "chatting" to "getting things done" [0][17]; on the technical side, Microsoft's MAI-DxO achieved 83% accuracy on complex NEJM cases and was published in Science [2], while Intel's CTO warned that the realistic upper limit for leading-edge wafer output is only about 2x, constraining compute expansion by supply chain hard limits [5].

Google is accelerating iteration across the Gemini ecosystem. Pichai confirmed that Gemini 5 will deliver a major leap over Argon, while the Gemini app side continues to ship creative capabilities such as web page building and video soundtracking [15][2][10]. On the open-source side, Ant Ling and Recursive respectively released the medical-enhanced MoE model Ling-3.0-flash-Sante and the automated research model Flint, while the long-context inference optimization solution TokenGate reported...

This week, security and model iteration accelerated simultaneously: a critical PHP core vulnerability triggered a global emergency shutdown, exposing the double-edged sword effect of AI-assisted vulnerability discovery [0]; at the same time, iFlytek released Spark X2.6 and SparkAudio 1.0 based on fully domestic computing power [10][5], Anthropic launched the proactive agent application Claude Duet [11][14], and Mirendil completed pre-training of a hundred-billion-parameter code foundation model at a cost of less than ten million USD [4]. The threshold and paradigm of AI R&D are being rewritten.

Grok Bot has received a major update, adding support for real-time artifacts and collaborative editing, and it can even let you play DOOM directly on the page [0][4][5]. Meanwhile, Anthropic launched Claude Duet, a proactive agent app built on Claude Opus 5.5, and released the first official version of the Claude API [7][22]. Walmart also launched an MCP agent gateway that allows AI to place orders on users' behalf, marking an acceleration in the commercialization of agents [2]...

Kling 4.0 completed intensive deployment across multiple platforms today and opened API applications, with its native multimodal video capabilities in the spotlight [2][3][13]. Meanwhile, OpenAI led the fix for a high-severity PHP vulnerability that had lain dormant for 14 years, further validating AI's practical value in cybersecurity offense and defense [16][24].

SpaceX achieved its first quarterly profit in its Q3 financial report, with its AI compute leasing business generating $7.6 billion in revenue, surpassing Starlink to become its largest revenue source, and it plans to build 20GW of new power and cooling facilities next year [22][20]. Meanwhile, OpenAI acquired Orchestra Bio to enter biopharmaceutical R&D collaboration, while Mistral AI and ElevenLabs each released new families of speech models, as AI infrastructure and vertical applications accelerate in tandem [3][9][13...

Agents are becoming the core growth engine of the AI ecosystem: weekly agent usage on the OpenRouter platform has reached 8x human usage, surging 34x since February 2026 [18]. Meanwhile, DeepSeek-V4.1-Flash tops Arena's Agent and Code leaderboards as the #1 open-source model [2], and vendors such as Mistral, VAST, and Eagle Corp have densely released new voice, world model, and inference acceleration products, with industry competition extending from model capabilities...

This week the AI field saw a dense burst of activity in the open-source ecosystem: VAST open-sourced a real-time world model and renderer, DeepSeek released the production-grade inference framework DeepInfer, and IBM Research topped the τ²-bench customer service leaderboard with a learned rubric. Meanwhile, Jensen Huang visited Europe and called for avoiding over-regulation of AI, while Spotify removed over 50 million AI spam songs within a year, showing that governance and innovation are accelerating in tandem [1][3][4][7][12].

The mismatch between AI capital expenditure and revenue growth has drawn market attention. Bank of America research shows AI capex this year beat expectations by 50% while revenue exceeded by only 20% [11], yet Morgan Stanley still forecasts that hyperscaler and neocloud AI spending will reach $1.58 trillion by 2027 [12]. On the technology side, ModelBest has for the first time achieved real-time, lossless on-device operation of a 35B-A3B-scale MoE on smartphones [10], while Marvell's photonic module allows HBM and GPUs to be separated by up to ...

This week the AI industry advanced on two fronts—"security governance" and "model cost-effectiveness": Microsoft launched Guardian Agents to monitor AI agents as insider threats [0][10], while MiniMax M3.1 Flash, in the guise of a "stealth model," topped OpenRouter for 10 consecutive days, processed 56.7T tokens before being officially revealed [23], and on Cline benchmarks against Opus and GPT at roughly 1/25 the cost [21]. Meanwhile, ...

Google's Gemini 4 Argon agent has, for the first time, autonomously completed Minecraft in 27 hours without human assistance, marking a key threshold crossed in AI Agent long-horizon task capabilities [8]; meanwhile, DeepSeek V4.1 Flash completed 3D web and game development at a cost of 4 yuan, pushing the cost-effectiveness of Vibe Coding to new heights [1][14]. On the enterprise side, Meta uses a 1,300-line thin orchestrator to schedule 14 sub-agents, providing a reusable engineering paradigm for Agent reliability...

AI Agents are moving from demos into the deep waters of real business scenarios: ModelBest released a grand supervision agent covering "five-in-one" supervisory capabilities [4], and Volcano Engine is incentivizing AI short drama creation with up to 200,000 yuan per title [6]; but at the same time, an Amazon paper points out that evaluating memory systems solely by QA scores carries an average bias of 16.8 points [7], and an experiment in which 6 Agents handled refunds failed entirely [9], reminding the industry that Agent deployment still requires more rigorous evaluation and boundary design.

The focus of "secondary development" in the open-source ecosystem is unexpectedly shifting toward image models. With 43,000 derivative models, FLUX.1-dev has become the most-modified base on Hugging Face, surpassing Llama and Qwen [0]. Meanwhile, DeepSeek-V4.1-Flash tops the Code Arena WebDev leaderboard as the No. 1 open-source model, at about 1/10 the cost of its competitors [19]. On the commercialization side, Anthropic has committed 5% of its shares to a newly established philanthropic fund...

This week the AI industry showed a pattern of "technological leaps and risk repricing" running in parallel: on one hand, Nvidia boosted Kimi K3 inference speed by 1.68x through vLLM optimization, and Google DeepMind's four Gemini 4 Argon agents defeated Minecraft for the first time without human intervention [6][11]; on the other hand, CoreWeave's stock plunged 25% and its junk bond yields soared to 15%, making it the epicenter of the "dual sell-off in stocks and bonds" in this round of AI investment frenzy, with the market...

Anthropic has released multiple reports outlining the full picture of AI's economic impact: its internal employees expect AI to complete 75% of their own work by 2030, and co-founder Jack Clark further predicts AI will drive nearly 50% growth in US GDP [1][14]; meanwhile, AWS says the payback period for AI server investment is under three years, while DeepSeek opens search and scraping APIs at 5 yuan per thousand calls, further lowering the cost of building Agents [6][12].

Scaled access to agentic AI is forcing governments and enterprises to reweigh their service strategies, while reliability issues with long-horizon agents reveal the stark reality that a 99% per-step success rate can plummet to 37% over a 100-step task [0][19]. Meanwhile, Anthropic revoked affected Claude API keys due to an OpenCode distillation attack, and NVIDIA is bringing 1 petaflop of compute to the edge with the RTX Spark laptop, starting at $2,399 [14][15].

DeepSeek releases V4.1-Flash-1017, achieving stronger benchmarks, faster speed, and shorter outputs through a "Pareto improvement in performance and cost," which evaluators say creates competitive pressure on GPT 5.6 Sol [15][16][19]; meanwhile, Anthropic's Claude cracked the 25-year-old NSA ECDSA seed mystery in under 30 minutes, reigniting discussion about cryptographic backdoors and the boundaries of model reasoning [4].

DeepSeek has officially launched its Search/Fetch web search API, entering the search infrastructure space at a low price of ¥5 per thousand calls [4]. Meanwhile, StepFun has released Step 5 Preview, a 600B MoE model with open weights and 1M context support, escalating competition among open-source large models [7]. On the application side, Anthropic disclosed its Claude Code code review system, achieving notable results of doubled PRs while rollbacks dropped by 40%+ [23]...

Anthropic's latest report reveals that AI agents' working hours have reached 6 times those of human employees, with the share of entry-level technical job openings plummeting from 27% to 9%. Meanwhile, the AI regulation draft it pitched to the U.S. Congress sets an extremely high threshold for halting deployment [3][4][9]. At the same time, Musk announced the development of an enterprise negotiation bot called Hagglebot, claiming it can improve cost efficiency by more than 5%, translating to annual savings of up to $6 trillion based on global GDP [6].

This week the AI industry saw a dual explosion in model capabilities and infrastructure: Google Gemini 4 Argon became the second lab to cross the 95% threshold on ARC-AGI-3 with a score of 97.5% [5], while Reflection released its first open-weight model Beam, shaking up the open-source ecosystem with a 501B-parameter MoE architecture under the Apache 2.0 license [14][23]. Meanwhile, real-world laptop testing of Nvidia's RTX Spark chip and Doubao Work's one-click AI asset migration feature signal that on-device AI and user data portability are becoming new competitive focal points [0][10][11].

Apple confirms that Apple Intelligence for China will launch this month, with Cook personally endorsing it; meanwhile, Mistral's CEO writes to oppose Europe abandoning frontier AI, Nvidia's software licensing model is exposed, and Gemini 4 Argon opens to Pro users—global AI competition has entered a critical phase of ecosystem positioning and business model validation [0][2][22][16].

Microsoft released its self-developed foundation model MAI-Base-2, claiming to surpass performance with one-quarter of OpenAI's compute [3]; meanwhile, Anthropic announced a new generation of anti-distillation defense results, with inference token theft dropping sharply by 90%–99% [6]. Domestically, Alibaba Cloud Bailian launched Coding Plan integrating models such as Qwen, GLM, Kimi, and MiniMax [9], while Xiaomi open-sourced the first end-to-end spatial audio understanding large model [21].

Microsoft's in-house model MAI-Base-2 achieves performance surpassing the current GPT-5.6 backend model using only one-quarter of OpenAI's compute; Anthropic's new-generation anti-distillation defense causes inference token theft to plummet by 90%–99%, elevating the threat to the national security level [2][4].