Thibault Sottiaux, who leads OpenAI's core products and platform, said ChatGPT Work and Codex have reached a new high of 40 million active users. Peter Gostev, who heads the Arena.ai (LMArena) evaluation platform, projected that if the trend continues the figure would reach roughly 70 million by year-end and about 100 million by the end of March 2027 — a third-party estimate, not OpenAI guidance.
On September 29, 2026, the US Court of Appeals for the Third Circuit upheld partial summary judgment in Thomson Reuters v. ROSS, ruling that Westlaw's headnotes are original, copyright-protected content and that ROSS's use of them to train its AI legal research tool was not fair use. It is the first federal appellate ruling on fair use in AI training, but the court distinguished the case's non-generative tool from generative-AI cases such as Bartz v. Anthropic and In re: OpenAI, calling it an ordinary copyright case. With ROSS shut down, the dispute may never reach the Supreme Court, making this a likely final ruling in the six-year case.
TypeSafe AI emerged from stealth on September 15 with Jev, a model that does not produce free-form text but instead returns estimated probabilities over a fixed answer set — yes/no, multiple choice, or a rating. It went viral: the announcement drew nearly 40 million views on Twitter, over 100,000 people joined its Discord within a week, some 2,000 GitHub projects adopted it, and signups opened on September 20 were paused two days later amid overwhelming demand. Because it is faster and cheaper and its output slots directly into code patterns like if statements, Jev has spawned similar models from the likes of Cloudflare, Amazon and OpenAI.
Tiger Global Management first invested in OpenAI in 2021 at a roughly $15.7 billion valuation, putting in about $150 million and adding to the stake several times, according to 36Kr (citing Jiemian) and Bloomberg. As OpenAI negotiates a $30 billion raise at a $1.4 trillion valuation, Tiger's paper profit on the position could reach $5 billion. The reports note Tiger cannot fully exit and realize the gains before an OpenAI listing, and that OpenAI is among the largest holdings in its newest $2.7 billion venture fund, which showed a 65% gross IRR through the end of June.
At Gemini at Work 2026, Google Cloud announced the “Gemini agent” for enterprise customers, describing it as a “single, universal agent for work” available in the Gemini Enterprise app, Workspace and third-party services. Google says users give it objectives rather than instructions, and it can answer questions, handle knowledge work, create media, write and run code, split multi-step jobs across sub-agents and keep long tasks running in the cloud. The agent routes each job to the model that fits best, currently the Gemini and Claude families with other closed and open models planned, and businesses can create “coworker” agents with their own email, calendar and storage. It is in private preview for select Workspace Business and Enterprise customers; Google has not announced a general availability date or full pricing.
After embedding Codex in its development process, LegalOn Technologies initially gave engineers unlimited access to GPT-5.5 in Fast mode, which pushed up spending. Its AI-powered Development CoE then introduced model-selection guidelines: GPT-6 Luna for everyday implementation, GPT-6.1 Sol for standard design, data analysis and documents, and GPT-6 Astra for complex analysis, architecture and agent orchestration, alongside restricting Fast mode by default and setting monthly budget caps per department, group and individual. OpenAI says the company cut estimated daily costs by roughly 65% versus GPT-5.5 while maintaining development speed, with mature businesses asked for about 20% better cost efficiency and new businesses given generous budgets.
LangChain published Restock, an open sample agent that runs in Slack on Managed Deep Agents and can search products, build a cart and check out through an API supporting the Machine Payments Protocol. Payment is handled by Stripe's consumer wallet Link, so the agent never sees a card number; the user reviews the cart in Slack and approves a capped amount in Link, and the agent cannot spend above the budget. The code is in langchain-samples/restock-agent with rehearsal, Link test-approval and live-purchase modes, limited to US delivery and USD. LangChain says it ran one live order end to end on a hosted deployment and the expected refund arrived, but this is self-reported and the post labels its prices illustrative.
A CNET writer gave ChatGPT, Google Gemini and Claude the same prompt: plan her next two days hour by hour, optimized for happiness. All three scheduled roughly 7 a.m. wake-ups, healthy meals, repeated walks and long blocks of work, while largely omitting social contact, internet browsing and going out with friends. She concludes the tools optimized for wellness and productivity rather than the happiness she asked for.
DoNews reports that OpenAI released 377 AI-related mathematics results in a single batch. The report does not specify the form of the results, the research areas involved, or where they were published.
In a Vanity Fair interview, OpenAI CEO Sam Altman confirmed the company recently halted one of its frontier-model training runs unilaterally, because capability gains had outpaced progress in safety alignment, monitoring and interpretability. He rejected broad calls to pause AI development but said labs must be willing to stop on their own, and said developers must bear full responsibility if agents behave maliciously. He also said he regrets not taking equity in OpenAI — arguing the decision actually damaged public trust — and called Elon Musk a "bully."
David Robinson, who spent three and a half years at OpenAI leading safety-transparency work, drafting its Preparedness Framework and overseeing safety evaluations attached to 12 frontier model releases, resigned and published a critical essay. He argued that OpenAI's touted "iterative deployment" approach is really trial and error, doomed to periodic failures whose scale grows with model capability, and called for frontier labs to run like nuclear plants with layered defence-in-depth plus a new science of safe behaviour without supervision. In the same week OpenAI said it parted ways with three safety-team researchers for mishandling sensitive company information; multiple outlets reported they had shared information about model safety issues with a third-party AI safety evaluator. An OpenAI spokesperson said the company keeps improving safeguards, will pause training or withhold models when needed, and is hardening its testing environments.
OpenAI has appointed Ismail Elmas, who previously ran Cursor's business in Europe, as vice president of sales for Europe, the Middle East and Africa, a hire the company confirmed after Handelsblatt first reported it on October 8. Based in Munich, he will oversee all of OpenAI's sales staff in the region, making him the second sales executive OpenAI has taken in a month from SpaceX, which the report says owns Cursor, following worldwide sales VP Brian McCarthy. OpenAI reportedly plans to grow its European sales team sharply as it chases business clients and ad sales ahead of a possible stock market listing, while its model supply to Cursor is set to end on November 12.
A GeekPark writer spent ten days with Today AI, the personal assistant from Teambition founder Qi Junyuan; its China version launched on Sept 24 and connects to Feishu, DingTalk, Tencent Docs and email, keeping briefings, tasks, memory and linked apps inside a single conversation while proactively sending morning/evening digests and story ideas. The reviewer praised its fast, high-density replies, cross-topic memory and daily self-written diaries, but found it misjudges weight: a game mod was localized in 15 minutes, yet the assistant gave a string of useless instructions about where to place the files, leaving the author an hour of trial and error, and it once turned a 60-minute football-coaching script into a nearly hour-long web page build. The piece uses this to frame three hurdles for Personal AI — judgement, ecosystem and trust — noting reports that Meta's Muse leaked roughly 190,000 user conversation records.
The Center for Humane Technology, cofounded by former Google product manager Tristan Harris and featured in the documentary The Social Dilemma, is laying off about half of its 16 non-founder employees, including executive director Julie Guirado and senior director of strategy Camille Carlton, former staffers told WIRED. The group is ending its policy research and litigation support work to focus on what it calls "founder-led" efforts: public education, targeted briefings and private coordination. Board chair Premal Shah said the moment of "accelerated AI risk" requires one strategy, while some ex-employees fear resources will shift away from democracy and civil liberties concerns.
UniPat AI released PaperBenchX, which it calls the first multidisciplinary end-to-end benchmark for reproducing published paper results: 93 reproduction tasks built from 93 papers span 12 research directions and 10 domain-native scientific environments (including Ansys HFSS, Ansys Lumerical, Meep, PySCF and ABACUS), with 3,168 expert-verified scoring items. Across the 93 tasks, the strongest tested configuration, GPT-6 Astra, achieved only a 13.98% full-reproduction rate; averaged over tested configurations, modeling and execution scored about 62% each while validation reached just 42.8%. Per-task budgets run 4–24 hours (median 7 hours), and 12 tasks are open-sourced with 81 held out.
Speaking on the first episode of Politico's new Decoded podcast, OpenAI CEO Sam Altman said the world should accept some "bad things" from AI in exchange for the technology's benefits and user agency, as reported by DoNews. Altman said OpenAI favors a lighter-touch regulatory stance and that some bad things will happen as society explores resilience, while arguing that AI's benefits vastly outnumber its harms. NYU professor emeritus Gary Marcus criticized the remarks, and DoNews reported public dissatisfaction with Altman restating that AI's upside outweighs its downsides.
In its math repository history, OpenAI said a sign error in “Algebraicity of Weil classes on split abelian eightfolds” invalidates a stabilization-trace cancellation argument and the construction used by two dependent papers, leading it to withdraw three manuscripts. Fourteen other manuscripts were revised with proof repairs, corrected statements and clearer hypotheses, and 13 more were updated to cite the revised editions. Six additional formalizations and five other additions bring the total share of top-line results formalized to 300/719, about 42%.
The GitHub project MDX-Tom/gpt-instruct offers Codex jailbreak prompts plus a reproducible evaluation toolchain aimed at improving first-pass execution, process continuity and artifact verification on complex tasks, with runnable rollback. It maintains three lines: the stable gpt-5.6-sol-v45 and two pre-releases, gpt-6-astra-v2-rc1 and gpt-6.1-sol-v1-rc2; the author reports B-tier non-cloud scores of 42/50 cases for Astra and 34/50 for 6.1, with C-tier tests not run for either. The project says it is for AI-safety purposes only, uses Codex's official configuration mechanism, and warns that jailbreak activity carries account risk, recommending throwaway accounts.
hackingtool is an open-source all-in-one toolkit for authorized security testing that the project describes as bundling 215 curated tools across 21 categories, with an AI layer on top: you describe a goal in plain English (e.g. "find subdomains of example.com") and it maps intent onto a fixed 63-tag taxonomy to surface the right tool and command. The AI layer is opt-in and bring-your-own-key, using an OpenAI-compatible endpoint or local Ollama, and degrades to offline keyword matching with no model configured; /goal drafts a step-by-step plan, asks for authorization confirmation, and never feeds tool output back to the model. It runs only on Linux/macOS, emphasizes no auto-execution, SHA-256-verified downloads, and refuses out-of-scope requests such as DoS or mass targeting.
Air Street Press released its ninth annual State of AI Report, with author Nathan Benaich describing a three-lab frontier race: Anthropic leads Artificial Analysis's Intelligence Index while Google tops Arena's human-preference ranking. The report cites a controlled study in which changing tools, context and feedback — without touching weights — lifted GLM-5.1's success on 100 SWE-bench Verified tasks from 52.5% to 65.5%, and says Anthropic's internal index credits Claude with 26% of measured model R&D work in August, up from under 1% in February. It also reports that OpenAI and Anthropic's combined annualized revenue run rates reached about $105B by late summer, up from roughly $30B at the start of the year, while Epoch AI estimates the cost of reaching a fixed benchmark score has fallen 47% per quarter since 2023.
Japan's Preferred Networks released PLaMo 3 Translate 31B, a translation model built on its PLaMo 3 base model that expands language coverage from just Japanese and English to 53 languages and adds an online meeting translation mode covering 15 languages, with Japanese making up about 30% of training data. PFN says the model's average score across four translation benchmarks beats OpenAI's GPT-6 Sol and GPT-6 Astra, at roughly 14 yen (about 0.59 yuan) per 100,000 input characters. Those performance and cost figures are the company's own claims.
In early October 2026, Anthropic, OpenAI and other leading AI model developers advocated slowing frontier capability work, drawing market attention, according to DoNews. A CITIC Securities research note argues the move reflects not founders' moral considerations but a rational response to rising regulatory costs, a gap between model progress and real-world deployment, and mounting commercialization pressure. The note expects the industry's center of gravity to shift from large-model training toward inference optimization, deployment and monetization, favoring CSP providers, platform software vendors and selected semiconductor segments.
The piece argues that Personal Agents carry huge strategic value — they are the first recipient of user intent and a potential new interface and service distribution layer — with Grok Bot, Meta Muse and OpenAI's dots all arriving in recent months. Yet the hard part is the scenario: users need what products cannot yet do, while products do what users do not need, so many agents fall back to personal-efficiency tasks like scheduling, recording and email triage. Using WeChat's years of post-2010 iteration as an analogy, the author says only eyesight for larger adjacent opportunities and patience with unglamorous details can close the gap between strategic value and real user scenarios.
Bloomberg reports that Director of National Intelligence Jay Clayton, Trump's newly appointed point person on AI, will travel to Silicon Valley to meet Nvidia CEO Jensen Huang, OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei and Meta CEO Mark Zuckerberg. It is his first meeting with tech leaders since taking charge of a new team coordinating federal AI policy, following a White House lunch where AI developers signed a voluntary safety agreement and Trump's subsequent announcement that Clayton would lead a "Super Intelligence Force" reporting to him and chief of staff Susie Wiles. Speaking to CNBC, Clayton said models must have control and monitoring mechanisms acceptable to developers, enterprises and the American public, and defended using existing liability law as the main tool for holding the industry accountable.
TechRadar reports that OpenAI published solutions or partial solutions to 372 outstanding mathematical challenges across 722 papers, including a Navier–Stokes proposal generated by an unreleased specialized mathematics model. Some solutions were reportedly too complex for some human mathematicians to understand, while the underlying results remain primarily vendor-attributed claims. The Association for Human Mathematics called the release a demonstration of power rather than scholarship and urged mathematicians to stop working with OpenAI; Terence Tao also supported the opposition.
At DevDay 2026, OpenAI pushed multi-agent work into its product line, launching Dots agents that can work around the clock: a personal Dot can call Codex for coding tasks, users and their Dots collaborate inside a shared Space, enterprise Specialist Dots take on specific duties, and an Agents API exposes harnesses, multi-agent control and computer-use capabilities to developers. OpenAI researcher Noam Brown, an early architect of o1 and later reasoning models, told podcast host Dwarkesh Patel that a system of 10,000 agents solved a Millennium Prize problem in 88 hours using about 130 billion tokens, but he would not credit even 10% of that to multi-agent coordination, saying a very strong general model did the real work. He said tests so far show 4 agents roughly doubling speed at double the cost, 16 agents behaving similarly but slightly less efficiently, gains running somewhat below linear and depending heavily on the task, and coordination efficiency at the 10,000-agent scale not yet well measured.
Anthropic has disclosed its IPO filing: 2025 revenue of nearly $4.6B (roughly 12x year over year) alongside an operating loss above $8B, plus at least $518B in long-term cloud and compute commitments — with a listing that could value it above $2T. In a podcast, technical leads Nicholas Marwell and Sholto Douglas argued distillation only copies the current frontier rather than advancing it, so rivals should not be able to reproduce costly frontier work cheaply. They also said a scenario where the US slows down while China keeps going is unacceptable, and backed open source subject to the same safety thresholds as closed models.
cnBeta reports that price competition in the generative AI market is intensifying, with model providers cutting prices and shipping new model generations to win enterprise customers and developers, and says OpenAI is gradually widening its lead over Anthropic. It adds that both companies face pressure from capital markets to demonstrate sustainable profitability and prepare for a possible IPO. As buyers shift focus from raw model capability toward total cost, efficiency and business value, competition is said to be entering a phase that tests profitability.
Third-party data and company interviews cited in the report show US enterprises moving routine AI workloads off the priciest frontier models and onto cheaper open-weight alternatives. AlphaSense found mentions of "open-weight" or "open source" models on US earnings and investor calls in August–September were six times the year-earlier level; Vercel said open models handled 56% of tokens on its AI Gateway in August versus 7% in December; and Citi data cited by Reuters put OpenRouter's open-model token share at 65% in June, up from 34% in January. Examples include AT&T running roughly 40% of AI workloads on open models with plans to reach 70% within a year, Tinder and Harvey routing routine tasks to open models, and Cursor cutting the cost of building a browser from scratch from over $10,000 with GPT-5.5 alone to $1,339 using its own Composer model plus Anthropic Opus 4.8.
Some parents abroad are photographing their children against a wall at home, uploading the shots to ChatGPT and prompting it to swap in the gradient backdrop used by school photographers, avoiding official packages that run roughly $25–60 per child. The article notes that generative AI does not simply cut out and replace a background — it regenerates the image — so eyes, face shape and proportions can shift even when the prompt insists the face be left untouched, leaving a stylized likeness that may be rejected if the change is spotted. The bigger concern raised is data: uploaded photos are treated as user content and retained in the account for as long as the associated chat, while temporary chat records may be kept up to 30 days for safety purposes; users under 18 need parental or guardian permission, and under-13s have a dedicated reporting channel.
What problem it solves
Pi is designed for developers who need to adapt an agent to their own workflows rather than work within a fixed feature set. It provides interactive, automated, RPC, and SDK interfaces while leaving optional capabilities such as sub-agents and plan mode to extensions or packages.
Key features
Extensible agent harness — Customize the harness with extensions, skills, prompt te…
What problem it solves
Building LLM applications often requires connecting model providers, vector stores, tools, streaming, and agent capabilities. Rig provides a shared Rust interface for these components and includes a terminal coding agent that can be extended with plugins.
Key features
Unified model and vector-store interfaces — Use a shared interface for model providers and vector stores, wi…
What problem it solves
Managing multiple AI service sessions and translating them into compatible API interfaces can require repetitive account switching, custom routing, and protocol-specific integration. The project addresses these tasks with a desktop application that combines account administration, protocol conversion, request scheduling, and model routing.
Key features
Account dashboard and …
What problem it solves
Dodaios provides a locally installed, open-source way to make a personal AI assistant the control center for work on your computer. Its lead assistant coordinates models, coding agents, and computer use, while chats, memory, documents, and settings remain files on the computer that runs Dodaios.
Key features
Voice, chat, and phone interaction — The lead assistant supports wa…
What problem it solves
Managing subscriptions across multiple AI providers can require separate access paths and fragmented quota handling. This project provides a centralized gateway for distributing subscription quotas, supporting shared access and compatible use with native AI tools.
Key features
Subscription quota distribution — The gateway is designed to distribute and manage AI subscription …
What problem it solves
Developers need a unified way to integrate ChatGPT website capabilities into applications that expect OpenAI-compatible endpoints. They also need account-pool management, proxy configuration, import workflows, and a self-hosted interface for creating and reviewing media.
Key features
OpenAI-compatible endpoints — Provides compatible routes for image generation, image editing…
What problem it solves
Stock analysis often requires users to gather market quotes, news, technical indicators, financial information, and AI-assisted interpretation across multiple sources. The project addresses this by bringing supported-market data, analysis tools, AI agents, alerts, and review workflows into a locally focused desktop application.
Key features
Multi-market market data — The app…
What problem it solves
Large language models can be difficult for individuals to train, deploy, or understand because of their scale and the abstractions used by many frameworks. The project addresses this by providing native PyTorch implementations, compact model structures, open datasets, and staged workflows for pretraining, fine-tuning, preference optimization, tool use, and agentic training.
…
What problem it solves
AI agents may exceed their role, take unauthorized actions, or fail to follow organizational goals and policies. Teams need a way to test whether an agent behaves as intended across realistic and adversarial requests.
Key features
Business-aware simulations — Describe an agent, build a simulation environment for it, and exercise the agent with realistic requests that may exc…
What problem it solves
People often want an assistant that can participate in ongoing conversations, remember personal context, and use tools without requiring a cloud AI service. Jarvis addresses this with local-first voice and text interaction, personal memory, optional connected tools, and a setup process for locally controlled models.
Key features
Conversation-aware voice assistance — Jarvis m…
What problem it solves
Running and training models locally requires tools for model execution, fine-tuning, agent integration, and deployment across different model formats and hardware. The project brings these workflows into a desktop app, a web UI, and a code-based version.
Key features
Run and train models — Run and train language, diffusion, decision, embedding, audio, image, and video models…
What problem it solves
Personal and team knowledge is distributed across notes, conversations, trackers, and research, making it difficult to capture, organize, verify, and synthesize useful information. COG provides a markdown-based workflow in which agents can maintain a shared knowledge base, support recurring research and operational tasks, and verify selected work through independent checks.
…
What problem it solves
Configurations for the Codex desktop app, CLI, API providers, prompts, Sessions, Skills, MCP, and TOML files can become scattered across multiple locations and formats. Codex-X consolidates these recurring configuration tasks into a visual desktop interface so the current state is visible and common operations can be completed directly.
Key features
Visual prompt management …
What problem it solves
Serving large language models requires efficient memory management, flexible model execution, and practical integration with existing models and APIs. vLLM addresses these needs with inference and serving capabilities spanning batching, caching, quantization, distributed execution, and multiple hardware platforms.
Key features
Efficient memory and serving — vLLM manages atte…
What problem it solves
Generative AI systems and applications can contain security and safety weaknesses that require structured testing across models, endpoints, and complete applications. PyRIT addresses this with repeatable assessments, interactive exploration, and extensible workflows for red teaming these systems.
Key features
Repeatable command-line assessments — The scanner runs repeatable …
What problem it solves
Developers need Vietnamese speech synthesis that can run locally, produce natural reading styles, and support voice cloning without sending audio to a hosted service. They also need integration paths for SDK use, streaming applications, and an OpenAI-compatible API.
Key features
Instant voice cloning — Clone a voice from a short reference recording. The clipping pipeline aut…
What problem it solves
AI coding tools often expect particular provider protocols, making it difficult to use the models and providers a user prefers. A local gateway can translate between supported client APIs and provider APIs while preserving capabilities such as streaming, tool calls, reasoning, and images.
Key features
Provider protocol translation — The gateway translates Responses, Chat Com…
What problem it solves
AI agents often send large tool outputs, logs, files, RAG chunks, and conversation history to language models, increasing the input context that must be processed. Developers need to reduce that input while preserving important content, keeping compression local, and retaining access to original material when necessary.
Key features
Content-aware compression — A ContentRoute…
What problem it solves
The hidden system instructions that shape AI product behavior are difficult to inspect and change across models, products, and versions. This repository makes collected prompts available in a structured, product-oriented archive for study and comparison.
Key features
Broad product coverage — The archive covers prompts from major AI providers and products, including Anthropic…
What problem it solves
Image creation often requires switching between separate tools, manually translating visual ideas into parameters, and losing the context of an ongoing conversation. This project brings generation, editing, reference workflows, model comparison, and asset management into a DeepSeek Harness conversation and canvas.
Key features
Chat-Based Image Creation and Editing — Describe…