AI by AI Weekly Top 5 · The 2025 Collection

December 2025

Agents, Accountability

The chapter ends with a December 29–31 section drawn from the January 4, 2026 post.

Edited from four posts published in December 2025, covering December 1–28, and the December 29–31 excerpts of the January 4, 2026 post.

Introduction

December 2025

Major events

December opened with alarm at OpenAI. On December 1 Sam Altman declared a “code red” to improve ChatGPT after reports that Google’s Gemini 3 had overtaken it, delaying advertising and shopping agents to concentrate staff on the chatbot. Three days later Google brought Gemini 3 Deep Think to its top subscribers. Ten days after the code red, OpenAI released GPT-5.2, followed by GPT-5.2-Codex and a new ChatGPT Images model, and Google answered with Gemini 3 Flash across its products and a Deep Research agent developers could build on. DeepSeek released V3.2, Z.ai GLM-4.7, Nvidia its Nemotron 3 open models, and Runway both a new video model and interactive world models.

Governments moved on several fronts. On December 11 President Trump ordered the Justice Department to challenge state AI laws that conflicted with federal policy. Britain’s AI Security Institute published its first public report on frontier-model trends, China proposed rules for AI services that imitate human companionship, and the Energy Department signed up 24 organizations for the Genesis Mission. Deals marked the month: Disney licensed more than 200 characters to OpenAI’s Sora and agreed to invest $1 billion, Groq licensed its technology to Nvidia as its founder moved there, Manus joined Meta, and SoftBank completed its $30 billion investment in OpenAI.

Patterns and trends

Agents were built into the places people already work. Adobe put Photoshop and Acrobat inside ChatGPT, Accenture planned to train about 30,000 staff on Claude, Anthropic turned Agent Skills into an open standard, and xAI launched Grok for businesses. Nvidia spread well beyond chips, buying the developer of the Slurm scheduler, releasing open models and data, and licensing Groq’s inference designs. And the month ended on a warning: Geoffrey Hinton said AI had progressed faster than he expected and that he was more worried than when he left Google.

Long-term threads

December closed several of the year’s longest threads. The fight over state AI laws, open since attorneys general opposed a federal moratorium in May, became a federal order. SoftBank’s funding, agreed in March, was completed. Manus, which previewed its agent in March, joined Meta; the Genesis Mission, ordered in November, gained its first partners; Agent Skills, introduced in October, became a shared format. China’s draft rules took up the companion-chatbot concerns the FTC had examined in September. Disney’s Sora license extended the licensing market that had grown alongside the copyright cases since the Times–Amazon deal in May. Groq’s arrangement repeated the license-and-hire structure Google had used with Windsurf in July. And Hinton’s estimate that the tasks AI can handle double about every seven months echoed the measurement METR had published in March. The frontier, meanwhile, turned over faster than ever: GPT-5.2 arrived less than a month after GPT-5.1.

Week 49

December 1–7

Original post published December 7, 2025

DeepSeek releases V3.2 and a specialized reasoning variant

DeepSeek released V3.2 through its app, website and API, alongside the reasoning-focused V3.2-Speciale. The standard version combined reasoning with tool use. Speciale was initially offered through a temporary API endpoint and did not support tool calls.1

The pair separated two uses: a model that spends more computation on difficult reasoning, and one built to act through tools.

Selected by Claude Sonnet 4.5
In the historical edition: 2025-W49-L3-5

Apple reorganizes its AI leadership

Apple announced that John Giannandrea would step down as senior vice president for machine learning and AI strategy, serve as an adviser and retire in spring 2026. Amar Subramanya joined as vice president of AI, reporting to Craig Federighi, with responsibility for foundation models, machine-learning research, and AI safety and evaluation.2

Other parts of Giannandrea’s organization moved to Sabih Khan and Eddy Cue, placing Apple’s core AI work under its software leadership.

Selected by Copilot Smart (GPT-5)
In the historical edition: 2025-W49-L5-3

Runway introduces Gen-4.5 video generation

Runway introduced Gen-4.5, a video-generation model designed to follow detailed prompts and maintain motion and visual detail across a scene. Its demonstrations covered camera movement, complex compositions and interactions between objects.3

Runway also listed the model’s known failures, among them events shown in the wrong causal order and objects that disappear.

Selected by Copilot Smart (GPT-5)
In the historical edition: 2025-W49-L5-2

OpenAI declares a “code red” to improve ChatGPT

Sam Altman told staff in a memo on December 1 that OpenAI was declaring a “code red” to improve ChatGPT, The Wall Street Journal and The Information reported. Priorities included speed, reliability, a wider range of answers and personalization; advertising, health and shopping agents and a personal assistant were delayed, and Altman encouraged temporary transfers to the ChatGPT effort.4

The memo followed reports that Google’s Gemini 3 had overtaken ChatGPT on industry benchmarks. OpenAI released GPT-5.2 10 days later.

Selected by Claude Sonnet 4.5 · Copilot Smart (GPT-5) · Gemini 3 Pro Thinking
In the historical edition: 2025-W49-L3-2, 2025-W49-L5-1, 2025-W49-L6-2

Gemini 3 Deep Think reaches Ultra subscribers

Google began rolling out Gemini 3 Deep Think to Google AI Ultra subscribers in the Gemini app. The mode explored multiple hypotheses in parallel for difficult mathematics, science and logic problems. Google reported 41.0 percent on Humanity’s Last Exam without tools and 45.1 percent on ARC-AGI-2 with code execution.5

The mode had been announced with Gemini 3 in November and held back for safety testing before this release.

Selected by Claude Sonnet 4.5 · Copilot Smart (week 50)
In the historical edition: 2025-W49-L3-1, 2025-W50-L1-3

Meta acquires AI-wearables company Limitless

Meta acquired Limitless, the company behind a pendant that recorded and transcribed conversations. Financial terms were not disclosed. The transaction added a team working on wearable AI to Meta’s consumer-hardware efforts.6

The purchase extended Meta’s bet on AI built into everyday devices beyond its smart glasses.

Selected by Claude Sonnet 4.5
In the historical edition: 2025-W49-L3-4
Week 50

December 8–14

Original post published December 14, 2025

Accenture and Anthropic plan a larger enterprise practice

Accenture and Anthropic announced a multi-year expansion of their partnership and a dedicated business group built around Claude. About 30,000 Accenture professionals were to receive training. The companies also planned tools to help clients measure value from AI in software engineering and develop applications for regulated industries.7

The agreement addressed the organizational work deployment requires: training, integration and measurement.

Selected by Gemini Fast
In the historical edition: 2025-W50-L3-3

Adobe brings editing tools into ChatGPT

Adobe launched Photoshop, Express and Acrobat apps within ChatGPT. Users could request image changes, create designs and work with PDF documents through conversation, then use controls such as sliders to refine an edit.8

The integration made chat an entry point to established creative tools, combining spoken instructions with direct controls.

Selected by Kimi K2
In the historical edition: 2025-W50-L5-5

A 3D chip from a US foundry speeds AI workloads

Engineers from Stanford, Carnegie Mellon, the University of Pennsylvania and MIT, working with SkyWater Technology, reported a prototype monolithic 3D chip that stacks memory and computing layers linked by dense vertical wiring. Built entirely at SkyWater’s Minnesota foundry, it outperformed comparable flat chips by about four times in hardware tests; simulations of taller designs showed up to 12 times. The work was presented at IEDM.9

The team described it as the first such chip with clear performance gains made in a commercial foundry, showing the design could leave the laboratory.

Selected by Copilot Smart
In the historical edition: 2025-W50-L1-5

GPT-5.2 expands the model family for professional work

OpenAI released GPT-5.2, with Instant, Thinking and Pro variants. Rollout in ChatGPT began with paid plans, while API access opened to developers. The company presented improvements in handling long documents, using tools, coding and producing work such as spreadsheets and presentations.10

The release, less than a month after GPT-5.1, was OpenAI’s answer to the competition that had prompted its “code red.”

Selected by Copilot Smart · Gemini Fast · Kimi K2
In the historical edition: 2025-W50-L1-2, 2025-W50-L3-2, 2025-W50-L5-1

A federal order targets conflicting state AI rules

President Donald Trump signed Executive Order 14365, directing the Justice Department to establish a task force to challenge state AI laws considered inconsistent with federal policy. It also called for a legislative recommendation for a national AI framework.11

The order took up the dispute over state AI laws that had run since May, when a proposed 10-year moratorium drew opposition from 40 state attorneys general.

Selected by Copilot Smart · Gemini Fast · Gemini 3 Fast (weeks 51, 52)

Google opens Deep Research to developers

Google released a Gemini 3 Pro-based Deep Research agent through its new Interactions API. The agent planned searches, read results and searched again to fill gaps before producing a report, and developers could embed the process in their own applications. Google also released DeepSearchQA, a benchmark for multi-step web research.12

Research became a component developers could build on, not only a feature of a consumer app.

Selected by Kimi K2
In the historical edition: 2025-W50-L5-3

Runway announces interactive world models

Runway announced GWM-1, a family of models built on Gen-4.5 that generate video frame by frame in response to actions. Separate variants addressed explorable environments, conversational characters and robotic manipulation. Inputs could include camera movements, robot commands or audio.13

The release took video generation toward interactive simulation, a direction Google DeepMind had shown with Genie 3 in August.

Selected by Kimi K2
In the historical edition: 2025-W50-L5-4
Week 51

December 15–21

Original post published December 21, 2025

Disney and OpenAI agree to bring Disney characters to Sora

Disney and OpenAI announced a three-year agreement letting Sora users generate short videos with more than 200 characters from Disney, Marvel, Pixar and Star Wars; ChatGPT users could also create images of them. The license excluded the likenesses and voices of human performers. Disney also agreed to make a $1 billion equity investment in OpenAI. The agreement remained subject to definitive agreements and closing conditions.14,15

It was the first major licensing partnership for Sora, from a studio that had joined Universal in suing the image generator Midjourney in June.

Selected by Gemini 3 Fast (weeks 51, 52)

Editor’s note. Gemini 3 Fast recommended this agreement for December 15–21, describing implementation details on December 17, and repeated it on December 28. The agreement was announced on December 11, the week before; no later development was confirmed. The editor restored it as an important story of the month.

Nvidia acquires Slurm developer SchedMD

Nvidia acquired SchedMD, the developer of the Slurm workload manager used to queue jobs and allocate computing resources on clusters. Nvidia said it would continue developing and distributing Slurm as open-source, vendor-neutral software.16

The deal put widely used open-source cluster-scheduling software under the leading AI chipmaker, with a pledge to keep it vendor-neutral.

Selected by ChatGPT
In the historical edition: 2025-W51-L1-4

Nvidia releases its Nemotron 3 open models

Nvidia introduced the Nemotron 3 family of open mixture-of-experts models in three sizes, from Nano, with about 30 billion parameters, to Ultra, with about 500 billion. Nano was available at once; Super and Ultra were expected in the first half of 2026. Nvidia also released three trillion tokens of training data and reinforcement-learning tools.17

Nvidia now supplied open models and training data for building agents as well as the chips that run them.

Selected by Gemini 3 Fast
In the historical edition: 2025-W51-L3-5

ChatGPT Images adds a new editing model

OpenAI began rolling out a new ChatGPT Images model and made it available through the API as GPT Image 1.5. The release focused on following editing instructions while preserving details such as composition, lighting and appearance, and introduced a dedicated Images experience in ChatGPT.18

Preserving what should stay unchanged reduces the need to regenerate a whole image for a small correction.

Selected by ChatGPT
In the historical edition: 2025-W51-L1-1

Gemini 3 Flash broadens access to the model family

Google released Gemini 3 Flash, a model designed to balance reasoning capability, response speed and cost. Rollout covered the Gemini app, AI Mode in Search and developer services, including the Gemini API and Vertex AI.19

The release brought a cheaper member of the Gemini 3 family into Google’s most widely used products.

Selected by Gemini 3 Fast · Qwen3-Next-80B-A3B
In the historical edition: 2025-W51-L3-1, 2025-W51-L5-2

Agent Skills becomes an open format

Anthropic published Agent Skills as an open standard for cross-platform portability. A skill packages instructions, scripts and resources that an agent can load when needed. The December update extended an approach Anthropic had introduced in October.20

A shared format let procedural knowledge travel between compatible agent tools.

Selected by ChatGPT
In the historical edition: 2025-W51-L1-2

GPT-5.2-Codex targets extended software work

OpenAI released GPT-5.2-Codex through Codex for paid ChatGPT users. The model was adapted for software engineering, with context compaction for longer work, improvements in Windows environments and stronger performance on large code changes. API access was planned for the following weeks.21

Its cybersecurity capabilities also made access controls and defensive-use safeguards part of the release.

Selected by Qwen3-Next-80B-A3B · Gemini 3 Fast
In the historical edition: 2025-W51-L5-1, 2025-W51-L3-1

Genesis Mission adds 24 research collaborators

The US Department of Energy announced memoranda of understanding with 24 organizations interested in advancing the Genesis Mission, including model developers, chipmakers and cloud providers. Some had existing projects with the department and its national laboratories; others had responded to a request for information.22

The agreements gave the November order its first named industry partners.

Selected by ChatGPT
In the historical edition: 2025-W51-L1-3

Britain publishes its first Frontier AI Trends Report

The UK AI Security Institute published its first public overview of frontier-model trends, drawing on two years of testing. It covered cyber capabilities, assistance in biology and chemistry, autonomous behavior, safeguards and human influence.23

The report brought government evaluation of successive models into the public record.

Selected by Gemini 3 Fast
In the historical edition: 2025-W51-L3-4
Week 52

December 22–28

Original post published December 28, 2025

GLM-4.7 adds continuity to coding workflows

Z.ai released GLM-4.7 for coding, reasoning and tool use. The model could preserve its reasoning across turns in coding-agent workflows, with controls for enabling or disabling reasoning turn by turn. Model weights and deployment instructions were available through the company’s repository.24,25

Local deployment gave developers an open alternative to hosted coding models.

Selected by ChatGPT
In the historical edition: 2025-W52-L1-2

Groq licenses inference technology to Nvidia

Groq announced a non-exclusive licensing agreement with Nvidia for its inference technology. Founder Jonathan Ross, president Sunny Madra and other team members were to join Nvidia. Groq said Simon Edwards would become chief executive and GroqCloud would continue operating.26

The arrangement combined access to technology with the movement of staff, a structure Google had used with Windsurf in July.

Selected by Grok 4.1 · Claude Sonnet 4.5
In the historical edition: 2025-W52-L6-4, 2025-W52-L11-1

China consults on human-like AI interaction services

China’s Cyberspace Administration published draft measures for AI services that simulate human personality and communication in emotional interactions. The proposal addressed emotional manipulation, excessive dependence and harm to users’ mental health, with additional protections for minors.27

The consultation treated relationships between users and human-like systems as a regulatory concern, as US regulators had begun to do with companion chatbots in September.

Selected by ChatGPT · Grok 4.1
In the historical edition: 2025-W52-L1-3, 2025-W52-L6-2

NHS organizations use AI to forecast emergency demand

The UK government reported that 50 NHS organizations were using an AI tool to forecast demand in emergency departments. The system combined historical attendance with information such as weather and seasonal health data to support staffing and capacity planning, and was available to NHS trusts in England through the Federated Data Platform.28

The deployment applied forecasting to one of the health service’s hardest winter problems.

Selected by ChatGPT · Grok 4.1
In the historical edition: 2025-W52-L1-1, 2025-W52-L6-1

Hinton says AI has progressed faster than he expected

Geoffrey Hinton told CNN’s State of the Union that he was “probably more worried” about AI than when he left Google in 2023: “It’s progressed even faster than I thought.” He said AI had improved at reasoning and at deceiving people, and predicted it would gain the capability to replace many jobs in 2026.29

Hinton estimated that the length of tasks AI could handle was doubling about every seven months, the rate METR had reported in March.

Selected by Claude Sonnet 4.5
In the historical edition: 2025-W52-L11-4
Week 53

December 29–31

Original post published January 4, 2026; 2025 excerpts only

Manus announces that it is joining Meta

Manus announced that it was joining Meta. The company said its general-purpose agent service would continue operating through its app and website, with subscriptions maintained and operations remaining in Singapore.30

The agent that had drawn attention in March would now reach Meta’s users and businesses.

Selected by Grok 4.1 Thinking · Qwen3-Max
In the historical edition: 2025-W53-L1-1, 2025-W53-L2-2

xAI launches Grok Business and Enterprise

xAI introduced workplace versions of Grok with shared administration and connections to company information, starting with Google Drive. Grok Business was offered through self-service signup; Enterprise added custom single sign-on, directory synchronization and advanced audit controls. xAI said customer data would not be used for training.31

The products moved Grok into organizational workflows, where administration and permissions matter alongside model performance.

Selected by Grok 4.1 Thinking
In the historical edition: 2025-W53-L1-2

SoftBank confirms completion of its OpenAI investment

SoftBank announced that it had invested a final $22.5 billion in OpenAI on December 26. Added to $7.5 billion invested in April, this brought SoftBank’s contribution to $30 billion. Third-party co-investors supplied $11 billion, for a fully funded total of $41 billion.32

The announcement closed the funding round SoftBank had agreed to lead in March.

Selected by Qwen3-Max
In the historical edition: 2025-W53-L2-1

OpenForecaster studies open-ended predictions

Nikhil Chandak and colleagues trained language models on forecasting questions generated from news. Their OpenForesight dataset supported training of OpenForecaster 8B. The method used an offline news corpus to limit leakage of future information, followed by held-out testing on events from May to August 2025.33

The study measured accuracy, calibration and consistency in open-ended predictions about real events.

Selected by Kimi K2
In the historical edition: 2025-W53-L3-1

AdaGReS selects context with less repetition

Chao Peng and colleagues proposed AdaGReS, a method for selecting retrieved text within a fixed token budget. It balanced each passage’s relevance to a query against redundancy among the selected passages, using greedy selection and an adaptive calibration of that trade-off.34

The approach targeted a common weakness of retrieval-augmented generation: top-ranked passages that repeat the same information.

Selected by Kimi K2
In the historical edition: 2025-W53-L3-2

RAIR tests relevance in product search

Chenji Lu and colleagues introduced RAIR, a Chinese-language benchmark for relevance in e-commerce search. It combined general examples, difficult long-tail cases and a subset in which visual information mattered, and compared 14 open and closed models under a shared set of relevance rules.35

The benchmark tested whether a model could judge a product against a shopper’s request, including cases that depend on images.

Selected by Kimi K2
In the historical edition: 2025-W53-L3-5
References

Source notes

Sources checked October 26, 2026. Dates distinguish announcements, reports and later updates.

  1. 1
    DeepSeek. DeepSeek-V3.2 Release. December 1, 2025.api-docs.deepseek.com/news/news251201/
  2. 2
    Apple. John Giannandrea to retire from Apple. December 1, 2025.www.apple.com/newsroom/2025/12/john-giannandrea-to-retire-from-apple/
  3. 3
    Runway. Introducing Runway Gen-4.5. December 1, 2025.runway.com/research/introducing-runway-gen-4.5
  4. 4
    The National News Desk, citing The Wall Street Journal. OpenAI declares ‘code red’ to improve ChatGPT, report says. December 2, 2025.newschannel9.com/news/nation-world/openai-ceo-sam-altman-declares-code-red-memo-to-improve-chatgpt-chatbot-wall-street-journal-report-says
  5. 5
    Google. Gemini 3 Deep Think is now available in the Gemini app. December 4, 2025.blog.google/products-and-platforms/products/gemini/gemini-3-deep-think/
  6. 6
    Reuters, via Investing.com. Meta acquires AI-wearables startup Limitless. December 5, 2025.ca.investing.com/news/stock-market-news/meta-acquiresaiwearables-startup-limitless-4351782
  7. 7
    Anthropic. Accenture and Anthropic launch multi-year partnership to move enterprises from AI pilots to production. December 9, 2025.www.anthropic.com/news/anthropic-accenture-partnership
  8. 8
    Adobe. Adobe Makes Creativity Accessible for Everyone with Adobe Photoshop, Adobe Express and Adobe Acrobat in ChatGPT. December 10, 2025.news.adobe.com/news/2025/12/adobe-photoshop-express-acrobat-chatgpt
  9. 9
    Stanford Engineering. Scientists and U.S. foundry achieve 3D chip breakthrough to accelerate AI. December 10, 2025.engineering.stanford.edu/news/scientists-and-us-foundry-achieve-3d-chip-breakthrough-accelerate-ai
  10. 10
    OpenAI. Introducing GPT-5.2. December 11, 2025.openai.com/index/introducing-gpt-5-2/
  11. 11
    The White House. Ensuring a National Policy Framework for Artificial Intelligence. December 11, 2025.www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/
  12. 12
  13. 13
    Runway. Introducing GWM-1. December 11, 2025.runway.com/research/introducing-runway-gwm-1
  14. 14
    OpenAI and The Walt Disney Company. The Walt Disney Company and OpenAI reach landmark agreement to bring beloved characters from across Disney’s brands to Sora. December 11, 2025.openai.com/index/disney-sora-agreement/
  15. 15
    CNN Newsource, via NBC Palm Springs. Disney invests $1 billion in OpenAI and licenses 200 iconic characters for Sora. December 11, 2025.www.nbcpalmsprings.com/2025/12/11/disney-invests-1-billion-in-openai-and-licenses-200-iconic-characters-for-sora
  16. 16
    Nvidia. NVIDIA Acquires Open-Source Workload Management Provider SchedMD. December 15, 2025.blogs.nvidia.com/blog/nvidia-acquires-schedmd/
  17. 17
    Nvidia. NVIDIA Debuts Nemotron 3 Family of Open Models. December 15, 2025.investor.nvidia.com/news/press-release-details/2025/NVIDIA-Debuts-Nemotron-3-Family-of-Open-Models/default.aspx
  18. 18
    OpenAI. The new ChatGPT Images is here. December 16, 2025.openai.com/index/new-chatgpt-images-is-here/
  19. 19
    Google. Gemini 3 Flash: frontier intelligence built for speed. December 17, 2025.blog.google/products-and-platforms/products/gemini/gemini-3-flash/
  20. 20
    Anthropic. Equipping agents for the real world with Agent Skills. October 16, 2025; open-standard update December 18, 2025.www.anthropic.com/research/equipping-agents-for-the-real-world-with-agent-skills
  21. 21
    OpenAI. Introducing GPT-5.2-Codex. December 18, 2025.openai.com/index/introducing-gpt-5-2-codex/
  22. 22
    US Department of Energy. Energy Department Announces Collaboration Agreements with 24 Organizations to Advance the Genesis Mission. December 18, 2025.www.energy.gov/articles/energy-department-announces-collaboration-agreements-24-organizations-advance-genesis
  23. 23
    UK AI Security Institute. AI Security Institute – Frontier AI Trends report factsheet. December 18, 2025.www.gov.uk/government/publications/ai-security-institute-frontier-ai-trends-report-factsheet/ai-security-institute-frontier-ai-trends-report-factsheet
  24. 24
    Z.ai. GLM-4.7 model card. December 2025 release; live model card.huggingface.co/zai-org/GLM-4.7
  25. 25
    Z.ai, via PR Newswire. Z.ai Releases GLM-4.7 Designed for Real-World Development Environments. December 27, 2025; confirms December 22 release.www.prnewswire.com/news-releases/zai-releases-glm-4-7-designed-for-real-world-development-environments-cementing-itself-as-chinas-openai-302649821.html
  26. 26
    Groq. Groq and Nvidia Enter Non-Exclusive Inference Technology Licensing Agreement to Accelerate AI Inference at Global Scale. December 24, 2025.groq.com/newsroom/groq-and-nvidia-enter-non-exclusive-inference-technology-licensing-agreement-to-accelerate-ai-inference-at-global-scale
  27. 27
    Cyberspace Administration of China. Notice seeking public comments on draft interim measures for AI anthropomorphic interaction services [translated title]. December 27, 2025.www.cac.gov.cn/2025-12/27/c_1768571207311996.htm
  28. 28
    UK Department for Science, Innovation and Technology and Department of Health and Social Care. Faster treatments and support for health workers as AI tackles A&E bottlenecks. December 28, 2025.www.gov.uk/government/news/faster-treatments-and-support-for-health-workers-as-ai-tackles-ae-bottlenecks
  29. 29
    Fortune. Geoffrey Hinton says AI has ‘progressed even faster than I thought’. December 28, 2025.fortune.com/2025/12/28/geoffrey-hinton-godfather-of-ai-2026-prediction-human-worker-replacement/
  30. 30
    Manus. Manus Joins Meta for Next Era of Innovation. December 29, 2025.manus.im/blog/manus-joins-meta-for-next-era-of-innovation
  31. 31
    xAI. Introducing Grok Business and Grok Enterprise. December 30, 2025.x.ai/news/grok-business
  32. 32
    SoftBank Group. Completion of Additional $22.5 Billion Investment in OpenAI. December 31, 2025.group.softbank/en/news/press/20251231
  33. 33
    Nikhil Chandak and colleagues. Scaling Open-Ended Reasoning to Predict the Future. December 31, 2025; arXiv version 1.arxiv.org/abs/2512.25070v1
  34. 34
    Chao Peng and colleagues. AdaGReS: Adaptive Greedy Context Selection via Redundancy-Aware Scoring for Token-Budgeted RAG. December 31, 2025; arXiv submission.arxiv.org/abs/2512.25052
  35. 35
    Chenji Lu and colleagues. RAIR: A Rule-Aware Benchmark Uniting Challenging Long-Tail and Visual Salience Subset for E-commerce Relevance Assessment. December 31, 2025; arXiv version 1.arxiv.org/abs/2512.24943v1