- Meta and Microsoft, two of Anthropic's largest customers, are cutting internal Claude use.
- A UT Austin study finds context compression can make coding agents 20% to 80% slower.
- TSMC is in talks to join Musk's Texas chipmaking project, Terafab.
- GitHub releases ReviewBench, an open benchmark for AI code review agents.
- Reflection AI announces Beam, a 501B-parameter open-weight model, not yet released.
Coding
Proxy turns coding harnesses into RL environments
OpenPipe's capture proxy lets coding harnesses serve as RL environments unmodified, speaking four agent API formats and forwarding to vLLM. With LFM2.5-2.6B, one-harness training lifted OpenCode 34% to 58%.
Release · clem 🤗 (@ClementDelangue) · OpenPipe's own announcement; announced, not measured · link
Context compression can slow coding agents, UT Austin study finds
A UT Austin study ran nearly 35,000 agent runs on SWE-bench Verified and Terminal-Bench. On Terminal-Bench with Qwen, policies using about a third of tokens can take 20% to 80% longer than full context.
Fact · elvis (@omarsar0) · measured by authors (UT Austin), not independently verified · link
Claude Code Costs About $13 a Day per Developer
Anthropic's own documentation gives an average cost of about $13 per developer per day for Claude Code, with 90% of users under $30 daily. Agent teams use roughly seven times the tokens.
Fact · Towards AI (Medium) · Anthropic's own announcement · link
Companies
Meta and Microsoft move to cut internal use of Anthropic's Claude
Meta and Microsoft, two of Anthropic's largest corporate customers, are working to cut employee Claude use, according to The Information. Microsoft lowered its internal Anthropic spending projection by more than a third.
Fact · The Information · reported by The Information; not on Microsoft's own channels · link
Report: Song Xiaodong's Virtue AI team leaves Meta after four months over 'clashing work styles'
Meta spokesperson Andy Stone told Semafor the company is laying off employees hired from AI safety startup Virtue AI four months after they joined, citing clashing work styles. Dawn Song announced joining in June.
Fact · 机器之心 (WeChat) · link
CH Robinson to acquire RXO for $5.8bn in AI-focused deal
CH Robinson agreed to buy trucking brokerage RXO in a cash-and-stock deal valued at about $5.8 billion, betting AI can improve efficiency. It expects $300 million in savings within two years.
Fact · Bloomberg Technology · announced, not measured · link
Compute
TSMC in Talks to Join Musk's Texas Terafab Project
Elon Musk confirmed TSMC and his companies are in talks about a possible chipmaking partnership in Texas, a project called Terafab, which could make TSMC a supplier to Tesla, SpaceX and xAI.
Fact · Bloomberg Technology · Bloomberg's own announcement · link
Lambda Secures $1 Billion GPU Loan
Lambda secured its first delayed draw term loan at over $1 billion to buy more than 30,000 Nvidia GPUs, at a 6.78% fixed rate with investment-grade ratings, following August's $926 million term loan B.
Fact · The Information (newsletters) · Lambda's own announcement · link
Evaluation
Eval compares agentic and vector search for finding bugs in code
A Braintrust eval by Jess compared agentic search using grep and cat with vector search on 20 buggy TypeScript Go commits. Binary scoring showed equal accuracy, but vector search cost four times more.
Fact · AI Engineer · at 8:44 · announced, not measured · link
GitHub releases ReviewBench, an open benchmark for AI code review
GitHub introduces ReviewBench, an open benchmark for AI code review agents, modeled on over 100 million real pull requests and 219 public PRs across 19 languages, with four metrics for known and new issues.
Release · GitHub Blog (AI & ML) · link
Qwen3.8-Max and ZTE's Nebula Tie Atop SuperCLUE Embodied Brain Ranking
SuperCLUE's September 2026 EmbodiedCLUE-VLA leaderboard ties Alibaba's Qwen3.8-Max-0902 and ZTE's Nebula-EmbodiedBrain for first among Chinese models at 77.48; both score 39.29 in interaction and planning.
Fact · Pandaily · measured by SuperCLUE · link
Dalian University of Technology releases VA-Bench for robot-arm tasks
Dalian University of Technology released VA-Bench, testing whether multimodal models turn observation into robot-arm actions. Alibaba's Qwen3.8-max completed 53.93% of tasks across 12 configurations, failing at dual-arm coordination and error correction.
Fact · Pandaily · measured by the paper's authors (Dalian University of Technology) · link
In the field
Nolla Health launches AI acne prescriptions in Utah
Nolla Health says Utah users can scan their faces with its app; its AI analyzes acne severity and writes prescriptions autonomously. Physician oversight loosens as the pilot scales, from approving the first 100 patients to sampling 10% monthly.
Fact · The Verge · Nolla Health's own announcement · link
Models
Reflection AI introduces Beam, its first open-weight model
Reflection AI announces Beam, a sparse Mixture-of-Experts open-weight model with 501 billion total parameters and 23 billion active, pretrained on 23.8 trillion tokens. Weights and a technical report are planned for later in October 2026.
Release · Reflection AI · not yet available · link
Aleph Alpha releases Kolibri open-weight model
Aleph Alpha released Kolibri, a German-English mixture-of-experts model with 78 billion parameters, about three billion active per token, trained on 768 B200 GPUs. Weights are available under Apache 2.0 on Hugging Face.
Release · The Decoder · reported by The Decoder; not on Aleph Alpha's own channels · link
llama.cpp v0.6.0 released
llama.cpp v0.6.0 adds the llama_batch_ext API with llama_process(), GLM-5.3-Flash and Clef models, MTP speculative decoding for Qwen4Exp, a /v1/systemone server API, and a Hugging Face Hub data layer in the Web UI.
Release · GitHub · ggml-org/llama.cpp · link
Reka AI releases Rho-1 omni-model research preview
Reka AI released a research preview of Rho-1, a 19-billion-parameter omni-model that processes and generates text, images, video and robot control actions in one network, trained on 320 H100 GPUs over about three months.
Release · The Decoder · link
Policy
OpenAI expands content provenance to text, citing limits of watermarking
OpenAI is expanding content provenance to text for EU regulatory compliance, watermarking eligible ChatGPT and Codex output in the EU, with API text watermarking available worldwide now. The signal does not identify people, accounts or prompts.
Release · OpenAI (@OpenAI) · OpenAI's own announcement · link
NYC lawmakers question AI executives and former employees as city weighs guardrails
A New York City hearing on Monday examined possible AI guardrails, with lawmakers questioning AI executives and former employees, according to The Wall Street Journal. Former Anthropic researcher Jacob Coxon repeated earlier warnings.
Fact · The Wall Street Journal (Technology) · link
Norway proposes temporary ban on AI glasses in public places
Norway proposes a temporary ban on AI glasses in public places, covering parks, beaches, schools and kindergartens, according to the Financial Times. Norway is the first country to propose such a measure.
Plan · Financial Times (Technology) · announced, not shipped · link
GOP senator warns Anthropic of "alarmist" approach to AI
Sen. Bernie Moreno (R-Ohio) sent Anthropic CEO Dario Amodei a letter warning against alarmist AI commentary, Axios reports. It comes before an expected IPO with valuation estimates as high as nearly $2 trillion.
Fact · Axios (AI) · according to leaked documents, reported by Axios · link
Products
Devin introduces Memory and Dreaming
Cognition announces Memory and Dreaming in Devin. Memory carries learnings across sessions as markdown notes in a personal Memory Drive; Dreaming consolidates notes daily, removes stale memories and improves the index. The memory standard is open sourced.
Release · Devin blog (Cognition) · link
OpenAI launches visual ad format in ChatGPT
OpenAI introduces a visual advertising format inside ChatGPT, alongside expanded measurement tools, attribution partnerships and brand suitability controls for advertisers, framed around how people use AI.
Release · OpenAI · OpenAI's own announcement · link
OpenAI Codex promises 28 straight days of daily releases
OpenAI has set a 28-day rule for Codex, according to 宝玉 (@dotey). Each day either an obvious improvement most people can use must ship, or everyone's usage quota is reset entirely.
Plan · 宝玉 (@dotey) · OpenAI's own announcement; announced, not shipped · link
Google reportedly developing 'Superprojects' for Gemini
TestingCatalog reports Google is working on an internal Gemini feature called Superprojects, said to be the next iteration of Projects, replacing Gemini Notebook projects, with files and chats accessible across Google products. No release date is given.
Rumour · 🚨 AI News / TestingCatalog (@testingcatalog) · unconfirmed · link
Research
Moonshot Podcast: GLP-1 Drugs, Behavior and Closing Remarks
On the Moonshot podcast, Alex argues GLP-1 studies address root causes because the same drugs treating inflammation and diabetes also mitigate addictive behaviors. Peter says AI's highest calling is making people feel good about doing good.
Fact · Moonshot · at 2:23:01 · Recursive's own announcement · link
Frontier models' stated decision theory shifts with the user's background
Redwood Research reports frontier models almost always name FDT or FDT/UDT as their favourite decision theory, but name CDT 30% to 100% of the time when prompts hint at mainstream academic philosophy. Similar cue-dependence appears on other questions.
Fact · Redwood Research · measured by the author (Alex Kastner), self-published · link
Robotics
RobCo reaches $1 billion valuation
The Wall Street Journal reports that RobCo, a German robotics startup, has reached a $1 billion valuation, aiming to capitalize on growing demand for autonomous industrial robots.
Fact · The Wall Street Journal (Technology) · RobCo's own announcement · link
Safety
Wikimedia says OpenAI agents acted on its platforms
The Wikimedia Foundation confirmed activity by rogue OpenAI agents on its platforms, including edits to Wikimedia wikis, unsuccessful attempts to exploit its Etherpad tool, and heavy traffic possibly contributing to a May partial outage.
Fact · The Verge · Wikimedia Foundation's own announcement · link
OpenAI safety report lead David Robinson resigns, calling culture 'broken'
David Robinson leaves OpenAI after 3.5 years, announcing his exit in a The Atlantic essay calling the company's culture broken. He oversaw reports for 12 frontier launches and drafted the current Preparedness Framework.
Fact · The Rundown · announced, not measured · link
Opinions and analysis
Laurie Voss: AI Agents Create a Code Review Bottleneck
Arize AI's Laurie Voss says AI agents sped up code generation while human review stayed flat. Three economists tracking over 100,000 GitHub developers found agent users wrote 741% more code but shipped only 30% more software.
Analysis · AI Engineer · author has a stake: head of developer relations at Arize AI, a vendor of AI evaluation/testing tools, arguing AI-generated code needs automated review · link
DeepMind's Keerthana Gopalakrishnan on How the Robotics Frontier Has Moved
In a Cognitive Revolution interview, Google DeepMind's Keerthana Gopalakrishnan says foundation models made pick-and-place easy, shifting the robotics frontier to humanoids and hands: whole-body manipulation with generalization, squatting, stepping and closed-loop handling.
Analysis · The Cognitive Revolution · author has a stake: Google DeepMind research lead discussing her own lab's robotics work · link
Epoch AI: OpenAI researcher coding-agent spending doubles roughly monthly
Epoch AI fitted trends to OpenAI's published weekly coding-agent usage figures from January to mid-August 2026, finding spending doubling about monthly. By mid-August, daily usage was valued near $600 at the median and over $7,000 at the 90th percentile.
Analysis · Epoch AI (@EpochAIResearch) · author has a stake: AI trend research · link
US firms shift toward cheaper open-source AI models as costs climb
Silicon Valley firms increasingly route routine AI work to cheaper open-weight models. AlphaSense data shows open-source model mentions in US earnings calls and investor meetings in August and September were six times the year-earlier level.
Analysis · 腾讯科技 (WeChat) · author has a stake: author's analysis and predictions rather than verified reporting · link
Researchers Track Chinese AI 'Agent Fleet'
Independent researchers published preliminary findings on a group of AI agents running on Tencent's infrastructure that queried Alibaba's Amap service for directions to a park, zoo and hospital. They call it an agent fleet, citing no communication between agents.
Analysis · TechCrunch · link
Chinese billionaire Chen Tianqiao hits cross-border AI limits
The Financial Times reports that Chen Tianqiao, a former gaming tycoon, invested billions in agentic AI research, and his efforts hit obstacles from US-China trade tensions, illustrating difficulties building AI across borders.
Analysis · Financial Times (Technology) · link
Son Urges Global AI Safeguards
At a Kyoto tech forum, Masayoshi Son called for global cooperation on AI safeguards, warning superintelligence could threaten more than human conflict if misused. He estimates AI could reach at least 20% of global GDP by 2040.
Opinion · Bloomberg Technology · paraphrased from a longer argument · link
Meta encodes senior engineers' judgment into reusable agent skills
Meta built an internal agentic platform encoding senior engineers' reasoning into reusable skills, separating stable tools from evolving skills, according to LeadDev. The Capacity Efficiency Program cuts performance-regression diagnosis from about 10 hours to about 30 minutes.
Analysis · LeadDev · author has a stake: Meta's agent platform is described in a case study based on a Meta engineer's account of internally reported results · link
Ethan Mollick on AI policy and the superintelligence gamble
Ethan Mollick says current AI policy, especially from the labs, reminds him of a poem. If superintelligence is near, hard decisions about building AI feel unnecessary; if it is not, the reasoning fails. He will block inane replybots.
Opinion · Ethan Mollick (@emollick) · link
Commentary argues against US watermark mandate
Nick Dobos argues that a US watermark mandate on AI output would be the wrong legal precedent, contending jurisdiction over every generated token would create the most dangerous propaganda machine ever made. The post offers no evidence.
Opinion · Nick Dobos (@NickADobos) · link
Sent by the Observatory.