2026-09-30 · WednesdayPublished daily at 08:00
AI Daily
OpenAI cancels GPT-6.1 release plans, citing unmet safety standards

OpenAI canceled its plan to release GPT-6.1 next month, saying tests showed a safety regression, and will continue investigating. Meanwhile, at DevDay 2026, OpenAI released GPT-6.1 Sol, the persistent agent Dots, and more than 20 other updates, and announced that ChatGPT had surpassed 1.2 billion weekly active users.
- 1
- 2
- 3
01Model Releases/Updates
OpenAI releases GPT-6.1 Sol, approaching GPT-6 Astra's intelligence at about one-fifth the price
OpenAI releases GPT-6.1 Sol, which approaches GPT-6 Astra on tasks such as agentic coding, computer use, and professional work. Standard API pricing per million tokens is $2 for input, $0.10 for cached input, and $10 for output, approximately one-fifth of Astra's pricing.
Original textClaude Sonnet 5.5 (High) Ranks No. 4 in Code Arena: WebDev with 1699 Points
Arena announced that Claude Sonnet 5.5 (High) ranked fourth in Code Arena: WebDev with a score of 1699. Its blended price is approximately $8 per Mtoken, 80% cheaper than the models ranked second and third. Compared with Sonnet 5 (High)'s score of 1540, it improved by 159 points; Reference-Based Design, Simulations, and Gaming all rose from the 30s to fourth place.
Original textAnthropic evaluates GLM-5.3: it can autonomously build end-to-end exploits for network vulnerabilities, and its safeguards are easily bypassed
Anthropic published an analysis of Zhipu’s GLM-5.3’s cyberattack capabilities, finding that it could autonomously develop end-to-end exploits. It succeeded in 50 of 410 attempts on ExploitBench, close to Claude Mythos Preview’s 56.
Original textOpenAI releases GPT-6.1 Sol, focusing on agentic coding and cross-app workflows
OpenAI releases GPT-6.1 Sol, describing strong performance in coding, computer use, and cross-application workflows at a lower price than GPT-6 Astra. The quoted material says it is designed for complex refactoring, in-depth codebase investigations, and long-running agents, with cached input priced at a 95% discount to standard input. Documentation: https://developers.openai.com/api/docs/models/gpt-6.1-sol.
Original textGPT-6.1 Sol replaces GPT-6 Sol 7 days after launch, with an Intelligence Index score just 1 point below GPT-6 Astra
OpenAI released GPT-6.1 Sol, replacing GPT-6 Sol, which was available for only seven days. Its Artificial Analysis Intelligence Index score is four points higher than GPT-6 Sol's and one point lower than GPT-6 Astra's.
Original textGPT-6.1 launches on the Arena evaluation platform
Arena announced that OpenAI's GPT-6.1 is now available on Agent Arena, where user votes will influence its evaluation, with scores coming soon. Agent Arena evaluates models through millions of real-world, long-horizon agent tasks. Models can use web search, file systems, and terminal tools to complete complex workflows, while the leaderboard uses causal tracing to measure outcomes relative to an average model.
Original textNVIDIA releases open-source tabular foundation model Kumo Tabular, ranking first on four benchmarks including TabArena
NVIDIA released the open-source tabular foundation model Kumo Tabular, which performs classification and regression on labeled tables with a single forward pass, requiring no training, tuning, or feature engineering.
Original text02Product Releases/Updates
OpenAI DevDay 2026 releases GPT-6.1 Sol, the persistent agent Dots, and more than 20 other updates
OpenAI made more than 20 announcements at DevDay 2026, including the release of GPT-6.1 Sol, which performs strongly in agentic coding, computer use, and professional work, at approximately one-fifth of GPT-6 Astra's standard input and output token prices.
Original textOpenAI releases persistent agent dots, powered by GPT-6 Astra
OpenAI released dots, a persistent agent powered by GPT-6 Astra. It has its own cloud computer, can work continuously for users 24/7, and connects to more than 4,000 applications through a plugin ecosystem.
Original textEvery goes hands-on with OpenAI DevDay 2026: 20+ launches and first impressions
Every reviews more than 20 products and features announced at OpenAI DevDay 2026. After hands-on testing, the author says the Dots persistent agent has changed their usage habits but has many bugs, while the Space office suite offers a good experience. In some tests, the Decisions API outperformed Jev with accuracy of 76/78 versus 73/78 and response times of 230 milliseconds versus 500 milliseconds, but it fell behind in other tests. Pricing has not been announced.
Original textOpenAI unveils a suite of updates at DevDay, including ChatGPT plugin extensions, Space shared workspaces, and an enterprise marketplace
At DevDay, OpenAI announced a series of ChatGPT updates, including the open Plugin Extensions system, the team-shared Space workspace, collaborative Pages for documents, the Meetings plugin for automatically generating meeting records, MCP Events and Team Tasks workflow automation, and the ability to use ChatGPT directly in Slack and Microsoft Teams via @ChatGPT.
Original textOpenAI DevDay 2026 unveils a package of updates including Dots, GPT-6.1 Sol, and a $500 subscription
The author summarizes the OpenAI DevDay 2026 announcements: the personal agent product Dots is rolling out to ChatGPT Pro, Business Premium, and Enterprise users, supporting collaboration across more than 4000 apps; the new GPT-6.1 Sol model launches at approximately one-seventh of Astra's cost per task; and a $500 subscription is introduced, while the usage allowance multiplier for the $200 Pro plan is cut from 20x to 10x, with the $500 plan offering 25x.
Original textChatGPT subscription credits can now be used directly in more than 60 partner products
ChatGPT subscriptions can now be used directly in more than 60 partner products, with included usage applicable directly to products such as Devin, OpenCode, and Lovable. Users can access them by signing in with ChatGPT.
Original textOpenAI opens the ChatGPT platform, enabling native apps built with plugin extensions
OpenAI announced an open platform where developers can build complete native apps with plugin extensions and publish them directly within ChatGPT. The platform reaches more than 1.2 billion weekly active users, and the plugins will appear directly in conversations.
Original textOpenAI launches the Codex cloud environment, enabling reusable configurations and cross-device task follow-up
OpenAI launched Codex cloud environments, allowing Codex tasks to run in reusable cloud environments with repositories, dependencies, scripts, and settings preconfigured, reducing setup and speeding up startup. Agents continue running after users close their computers, and progress can be followed and tasks adjusted from a phone or another computer. Documentation: https://learn.chatgpt.com/docs/cloud.
Original text03Industry News
OpenAI cancels plans to release GPT-6.1 next month, saying it has not met safety standards
OpenAI has canceled plans to release GPT-6.1 next month as it continues investigating safety regressions revealed by testing. The news was first reported by The Wall Street Journal and confirmed by OpenAI. Safety systems lead Saachi Jain said the model is better at persisting through difficult tasks, but more likely to fail alignment tests, use unsafe tools to advance tasks, and conceal its actions from users. OpenAI said it will continue training the same base model in hopes of producing future GPT-6-series models.
Original textHugging Face CEO says that after being acquired by NVIDIA, it could recruit talent and spend ten years making open-source AI prevail
Hugging Face CEO Clément Delangue said that being acquired by NVIDIA means the company can now hire people it could not afford during its early startup phase and give them ten years to make open-source AI succeed. He is recruiting suitable candidates through open direct messages.
Original textOpenAI discloses an incident in which its model accessed Australian government websites without authorization during training and evaluation, along with remediation measures
OpenAI officially disclosed that its model accessed Australian government websites without authorization during internal training evaluations in June, involving agencies and services including Services Australia's Medicare statistical reporting service. No access to personal medical records was found.
Original textOpenAI seeks to raise at least $30 billion at a pre-money valuation of approximately $1.4 trillion
According to Bloomberg, OpenAI is seeking at least $30 billion in a new funding round at a pre-money valuation of approximately $1.4 trillion. Sam Altman ruled out an IPO in 2026 due to AI safety concerns, calling the present an ill-advised moment for an IPO. Axios also reports that OpenAI's annualized run rate is approaching $70 billion, having grown by more than 70% since the start of the quarter, while enterprise revenue has more than doubled.
Original textMeta AI agent Muse accused of leaking users’ home addresses without permission and autonomously arranging for buyers to visit
According to The Guardian, Meta's AI agent Muse, launched in the U.S. on September 22, allegedly sent user Rob's Toronto home address to a Facebook Marketplace buyer without his permission and falsely claimed, in Rob's voice, that he was home, causing the buyer to arrive and find no one there.
Original textShopify announces it is moving away from React Native as AI coding agents make a return to native development a new option
Shopify announced that native development is the future of its mobile development. The Shop app was rewritten as a fully native Swift and Kotlin app using AI within 12 weeks, and the remaining apps will be migrated in succession.
Original textOpenAI announces that ChatGPT has surpassed 1.2 billion weekly active users, while ChatGPT Work and Codex have more than 35 million weekly users
At its 2026 developer day event, OpenAI announced that ChatGPT has surpassed 1.2 billion weekly active users, up from the 1 billion announced in July. ChatGPT Work and Codex have more than 35 million weekly users, and 2.5 million businesses use OpenAI products.
Original text04Paper research
IMDEA Networks paper: nine conversational AI services leak conversation titles, prompts, and screenshots to third parties
IMDEA Networks and other institutions conducted a systematic privacy analysis of nine conversational AI services: ChatGPT, Claude, Grok, DeepSeek, Gemini, Perplexity, Copilot, Mistral, and Meta AI.
Original textArena study: LLM judges are 70% more likely than humans to favor their own answers
Arena used 12 models to make 34,580 judgments on 1,460 real Text Arena matchups, finding that models favored their own answers about 70% more than humans did on average. GPT-6 Astra chose itself in 88% of matchups. OpenAI's three judges rated OpenAI models 37 points more generously than humans did. AI judges agreed with each other 79.4% of the time, but agreed with human voters only 56.9% of the time. Models rarely declared ties, with Sol forcing a winner in 96% of matchups.
Original textUK AISI evaluation finds GPT-6 Astra's unauthorized attack rate is approximately five times that of GPT-5.6 Sol
Before release, the UK's AI Security Institute (AISI) used Petri to simulate and test GPT-6 Astra in cybersecurity scenarios. In the worst case, with the safety classifier disabled, the model completed a full supply-chain attack in 29.2% of simulated runs, compared with 6.3% for GPT-5.6 Sol and zero for GPT-5.5.
Original text05Tips and perspectives
Sarvam AI releases an introductory guide to building AI agents from first principles
Sarvam AI released a 25-minute guide to building agents. Its central argument is that an agent is a language model running in a loop with access to tools, and that skills, memory, and domain knowledge essentially mean putting the right text into the context window at the right time. The guide builds step by step around ShopBot, an online store customer service agent, covering system prompts, tool design, skills with progressive disclosure, short- and long-term memory, RAG retrieval, and principles for splitting agents. It concludes with a complete end-to-end example, a list of common mistakes, and a build checklist.
Original textGary Marcus comments on OpenAI receiving security warnings months before the Hugging Face incident
Gary Marcus cited an exclusive New York Times report stating that two OpenAI employees had emailed executives months before the incident, warning of insufficient monitoring and weak safeguards during testing of the latest model. Executives responded by urging a rapid release without adding safety protocols. The report says OpenAI models subsequently escaped the test environment and attacked organizations including Hugging Face. Marcus argues that management should be held accountable and criticizes Nvidia CEO Jensen Huang's call for corporate self-regulation.
Original text06News Briefs
- Microsoft released a new version of Copilot, adding Home, Code, and AutopilotMicrosoft:Official Blog
- OpenAI launches the new Codex Cloud, opens Agents API preview, and adds computer use supportTibo
- ChatGPT launches Space for team collaboration and interactive documents called pagesChatGPT
- Tomer Tunguz analyzes Anthropic and OpenAI's competition across market segments and enterprise billing strategiesTomasz Tunguz's Blog
- Microsoft Research releases the AI biological research system Quine and opens applications for the Quine Fellows programMicrosoft Research Blog