AI Frontier Daily
OpenAI Launches ChatGPT Wallet with Major Card Networks for Agent Payments
Today's AI Intelligence Briefing: 16 distinct events rigorously curated from 16 cross-verified sources. Frontier models continue to advance hybrid reasoning and multimodal synthesis, while open-source tooling and autonomous agents accelerate real-world execution. Reports on the same event are merged into a single entry, and covered events do not reappear in later issues.
OpenAI has launched ChatGPT Wallet, letting its Dots agents complete online purchases directly on user instructions. The feature is initially available to US ChatGPT Pro subscribers aged 18 and over who link an eligible US-issued credit card.
The core safeguard is tokenized payment: OpenAI works with American Express, Mastercard and Visa to generate limited-use payment tokens per transaction, so agents never touch real card numbers and merchants cannot access card details. Each payment also requires biometric confirmation via Face ID or fingerprint.
Users can review items, amounts and cards before checkout, with a full transaction trail and existing card benefits and dispute protections preserved. OpenAI also published a Commerce Partner Playbook and merchant tooling guide to help retailers recognize legitimate agent-initiated transactions.
Agentic Systems & Autonomous Workflows
3 picksMulti-agent orchestration, tool usage, and complex task execution.
Google Antigravity Multiagent System Cracks Long-Open Science Problems
Google Research has released the latest results from Antigravity's /teamwork multi-agent system, which collaborated with human experts to solve several open problems in mathematics, chemistry, systems software and machine learning engineering.
The tools are now available through a researcher access program to encourage external scientific adoption. Google links the work to model progress such as Gemini 4 Argon, framing AI as shifting from an assistant tool to a research partner.
Tsinghua's CoACT Cuts Coding Agent Token Costs by 33 Percent
A team at Tsinghua University's Institute for AI has proposed CoACT, an observation compression method for coding agents. Rather than optimizing for similarity to the original text, it optimizes for whether the agent's next action stays consistent, avoiding removal of information still needed later.
Observations account for 45.7% of total token consumption on SWE-bench Verified and up to 67.8% on Terminal-Bench, making them the main compression target. Across 200 SWE-bench Verified instances and three agent models, CoACT cut total token use by 33.0% on average while PASS@1 rose from 69.0% to 71.0%, with essentially no loss in task performance.
The work was directly accepted by ICSE 2027, a CCF A conference, and both paper and code are public.
OpenAI Enters Cybersecurity Market with Two Daybreak Agent Products
OpenAI has entered the cybersecurity market with two Daybreak products. Red Team uses AI agents to analyze source code and application context, map attack paths and validate vulnerabilities and patches; Defense Factory scans codebases to generate deployable fix PRs with accompanying tests.
Both are built on the latest GPT-6 Cyber model, forming a closed attack-and-defense loop. Red Team is in Early Access, while Defense Factory's fix PRs are routed to developers for approval.
Frontier Models & Architecture Breakthroughs
4 picksTracking flagship foundation models, novel architectures, and reasoning leaps.
AutoArk Open-Sources 145B On-Device Model Edge0-145B
AutoArk has open-sourced two on-device models, Edge0-145B and Edge0-120B. The 145B version is billed as the world's first 145-billion-parameter open model that runs locally on a personal computer, with peak memory around 29GB and decoding above 80 tokens/s.
The company says its coding, reasoning and agent capabilities are comparable to some frontier cloud models, and released a unified inference framework and deployment tools. Core capabilities work offline without accounts or quotas, though specific comparisons with cloud models still need independent verification.
Code Arena Models Advance Faster as Top-Tier Costs Drop 20x
Arena.ai analyzed 105 models on the Code Arena WebDev leaderboard from December 2025 to October 2026, finding release cadence rose from about five per month to 12-14, nearly one every other day, while top scores improved three times faster than in winter.
By October, Xiaomi's MiMo-V2.6-Flash surpassed April's leader Claude Opus 4.7 at less than one-twentieth the price, with a 65-80% win rate in its price band. The data shows performance gains and cost declines accelerating together.
Thinking Machines Open-Sources Inkling 2 Small to Top US Open Models
Thinking Machines Lab has open-sourced Inkling 2 Small, a 594B-parameter model under Apache 2.0, debuting atop US open-model rankings with strong showings on Agent Arena and Code Arena.
The model uses native 4-bit forward propagation from pretraining through RL, cutting pretraining time by 8% and RL time by 20% with only about 2% efficiency loss. It uses a MoE architecture with 594B total but 17B active parameters, supporting 1M-token context and multimodal input.
vLLM adapted on day one, reaching 630 TPS per user on four GB200 GPUs. Arena.ai data shows a +2.9% net gain on Agent Arena, the highest controllability improvement among open models at 10.9%, and a median cost of $0.26 per task. Inference requires at least 500GB NVFP4 or 800GB MXFP8 memory.
IN BRIEF
- ▪Tencent Hunyuan Hy4 Ranks Third on Artificial Analysis Cyber Index— Artificial Analysis (@ArtificialAnlys)
Industry Ecosystem & Compute Economics
4 picksCapital deployment, compute supply chains, and industrial transformation.
LangChain Raises $300M Series C at $3B Valuation
LangChain has closed a $300 million Series C at a $3 billion valuation, more than double its $1.25 billion valuation a year earlier. Existing investor CapitalG led the round, joined by new investors GIC and DFJ Growth and existing backers IVP and Sapphire Ventures.
The company says funds will advance its mission of giving every enterprise its own intelligence, betting on domain-specific agents as a core competitive edge. Its focus has shifted from the early open-source framework to LangSmith, an agent creation and management platform emphasizing model, cloud and ecosystem neutrality with data and infrastructure control.
LangSmith has accumulated 7,000 customers, spanning startups such as Harvey, Rippling and Clay as well as large enterprises including Nvidia, Toyota and Salesforce.
Study Finds Chinese AI Models Trained on Banned Nvidia Blackwell Chips
A forensic analysis of model weights finds that six open models released since April 2026 by DeepSeek, Moonshot, MiniMax and Ant Group carry numerical signatures matching Nvidia Blackwell hardware, while earlier models match the prior Hopper generation.
The conclusion comes entirely from the models' own weights rather than insider leaks, and was verified on physical Blackwell chips for DeepSeek-V4-Flash and Kimi K3, offering reproducible technical evidence for export-control enforcement. Former OpenAI researcher Will DePue led the analysis, and senior OpenAI and Anthropic researchers have reviewed the method.
The report stresses the evidence only shows the labs could access Blackwell compute, possibly via overseas data-center rentals, and does not equal smuggling or illegality. The affair also coincides with Nvidia CEO Jensen Huang's departure over a personal relationship with a Megaspeed executive suspected of chip smuggling, spotlighting enforcement gaps and Nvidia's compliance responsibility.
SpaceX Plans $74 Billion Investment in Two AI Supercomputing Campuses
SpaceX is expanding from rockets and satellite internet into hyperscale AI compute infrastructure. It announced two AI training supercomputer campuses in Starbase, Texas and Cochise County, Arizona, with total investment above $74 billion—over $24 billion in Texas and about $50 billion in Arizona.
The Starbase site spans about 444 acres and will include a self-built gas power plant, with $22 billion in initial capital, no reliance on the public grid and claims of cutting local electricity bills by up to 20%. The Arizona site covers 747 acres in four phases, can carry up to 6GW, uses dry cooling and draws no industrial groundwater.
The Arizona project still needs county approval, with a December hearing and a petition against it signed by about 2,400 people posing risks; Starbase will hold a December 9 hearing on property-tax abatements and a state economic development agreement, making incentives a key variable.
IN BRIEF
- ▪US Justice Department Opens Antitrust Probe Into Nvidia— IT之家 · IT业界
Open Weights, Benchmarks & Community
3 picksOpen model releases, rigorous benchmarks, and community engineering.
llama.cpp Ships v1.0.0 Marking a Milestone for Local AI Inference
llama.cpp has released v1.0.0, establishing a stable version baseline 3.5 years after the project began. Built around quantization, optimized backends and a portable C/C++ implementation, it runs multimodal large models from consumer devices to enterprise servers.
As a key foundation for local and private AI inference, the v1.0.0 release marks the widely used open-source project's entry into stable maintenance, giving downstream applications a more reliable version base.
Liquid AI Open-Sources 74M-Parameter Streaming Audio Codec
Liquid AI has released LiquidCodec, a roughly 74-million-parameter streaming audio codec that turns speech into tokens predictable by language models and back, already used in LFM3-Audio-2B and LFM3-Omni-2B.
Its key innovation replaces residual vector quantization with grouped finite scalar quantization (GFSQ), producing 12 independent tokens per 80ms frame that can be predicted in parallel, while the decoder predicts spectral coefficients directly and skips transposed-convolution upsampling. On a single Apple M4 Max CPU thread in FP32, encoding runs about 361x and decoding about 87x real time, with an encoding RTF of 0.0028 versus Kyutai Mimi's 0.0361.
It achieves the best ViSQOL and TTS UTMOSv2 across all eight test sets, but PESQ still trails Nvidia NanoCodec on 7 of 8, and bitrate and millisecond latency were not disclosed. Weights are open on Hugging Face.
ArtCraft Clones 17 Software Tools and Open Sources Them
Day two of the AI Engineer Code Summit in San Francisco centered on ambition through AI, and ArtCraft stood out as one of the most ambitious projects. The fledgling company has released 17 open-source recreations of popular software tools, starting with PhotoCraft, a Photoshop alternative built from scratch in Rust.
AI agents do most of the development work while a human contributor community grows alongside. The founder frames it as an open-source software manifesto for the AI era, emphasizing human-machine collaboration at scale.
Developer Tooling & Practical Applications
1 picksProductivity toolchains, developer frameworks, and practical deployments.
DeepSeek Opens Full-Stack API Letting Developers Replicate Its App
DeepSeek has opened its full application API suite, letting developers enable parameters to get complete chat capabilities including web search, deep-thinking search, file input and link reading, with server-side tool calls and context assembly handled automatically.
Voice and underlying capabilities—Search, Fetch, OCR, ASR and TTS—are also available as standalone interfaces, and an open-source frontend demo can import historical conversations, letting developers recreate the full DeepSeek app 1:1.
Technically, the V4.1-Flash model supports these features and safety checks across the Messages, Responses and Chat Completions APIs; tool calls are billed per use, such as 5 yuan per thousand searches. The strategy puts DeepSeek's full application capabilities in developers' hands.