01OpenAI Told Investors Annualized Revenue Was About $50 Billion at the End of September, Below the Previously Circulated $68 Billion Measure
OpenAI told investors that its annualized revenue reached roughly $50 billion at the end of September, clarifying a key financial measure as the company faces pressure to support its $852 billion valuation and prepares for a possible public offering. Annualized revenue is a run-rate estimate based on the current pace of business, not revenue recognized over a completed year.
The figure is about $18 billion below the $68 billion number widely reported in late September. A person familiar with the presentation said the higher measure included gross revenue generated by OpenAI’s partners, allowing investors to compare the company more directly with rival Anthropic. The source did not disclose which partner revenues were included or provide audited figures.
OpenAI’s investor presentation also highlighted growth during the third quarter. The company said its total revenue run rate grew 77% during the period, while the run rate for its enterprise business rose 107%.
Several AI-related stocks declined after the new details emerged on Thursday. Nvidia shares fell 3%, Oracle dropped nearly 6%, and cloud-computing provider CoreWeave lost nearly 8%. AMD and Broadcom each fell 4%, Intel declined 5%, and Super Micro Computer slipped nearly 5%. The timing establishes that the declines followed the disclosure, but does not show that OpenAI’s revenue figures were their sole cause.
The distinction between the two revenue measures matters as OpenAI seeks to demonstrate that its business can support its valuation and enormous capital needs. The company confidentially filed a prospectus with regulators in June, and executives have indicated that they are considering a 2027 market debut. However, no listing date has been formally confirmed, and CEO Sam Altman said in September that the present moment would be ill-advised for going public, partly because of continuing AI safety concerns.
OpenAI is also in early discussions with investors about another funding round that could raise about $30 billion. That amount may change, the talks are being driven by investor demand, and no term sheet has been finalized. The newly disclosed $50 billion run rate therefore gives prospective investors a narrower measure of OpenAI’s own revenue, but leaves both the financing terms and IPO timetable unresolved.
02Anthropic Updates Claude Policy, Banning Sustained, Needless Abuse and Tightening Rules on Weapons, Surveillance and Political Propaganda
Anthropic has updated its usage policy for the first time in more than a year, setting clearer limits on how Claude may be used for weapons, surveillance, political campaigns and autonomous hardware. The company also prohibited “sustained and needless abusive or cruel behavior” toward its AI model, extending an experiment announced last August that allowed Claude to end conversations with persistently harmful or abusive users.
The model-abuse rule is intended only for extreme cases in which users repeatedly act cruelly toward Claude without a discernible purpose, Anthropic said. It does not cover ordinary frustration, disagreement, dark themes in creative work, or model testing and research. Ending the conversation remains Anthropic’s primary enforcement mechanism; the company has not said whether it might impose further penalties such as user bans. Anthropic is studying questions around potential model welfare but has not claimed that Claude is conscious.
The broader update consolidates existing restrictions into a ban on deceptive commercial or political campaigns. Users may not hide who is behind a message or amplify material through fake accounts or posts. Election-related prohibitions include deceiving voters about candidates or voting, impersonating candidates or election officials, disrupting elections and attempting to suppress turnout.
Anthropic also expanded its weapons restrictions after observing multiple attempts to obtain weapons-development guidance and control software from Claude. The ban now expressly covers software and components that make weapons function, as well as activities such as arming drones and other autonomous vehicles.
The surveillance rules were clarified after Anthropic’s threat intelligence reporting indicated that Claude-powered surveillance was increasingly being used to identify and track political dissidents. The policy prohibits tracking people without consent, whether live or through previously collected data; building or improving surveillance tools; and using Claude to decide or recommend whom law enforcement should investigate, arrest or charge.
Models connected to autonomous hardware capable of causing injury must now be observable and stoppable by a qualified operator, although the policy does not clarify whether that operator must be human. Rules may also be modified for certain government contracts if Anthropic judges that contractual restrictions and safeguards adequately mitigate potential harms, but the company has not disclosed its standard for making that determination.
03Long-WAM Extends Robots’ Visual History to 19.2 Seconds, Raising RoboCasa GR-1 Success From 63.3% to 78.7%
Robots need visual history to judge motion and determine how far a task has progressed, but processing more video can delay their next action. Long-WAM, a framework for causal world-action models—systems that predict what will happen next while selecting robot actions—aims to use longer histories without breaking real-time control.
Its central claim is that giving a model access to past observations does not guarantee that it will use them effectively. The researchers first pretrained Long-WAM to predict future video from past robot and first-person footage, without action labels. They then preserved that causal, history-to-future structure while adapting the model to produce actions.
On the RoboCasa GR-1 benchmark, extending visual context from no history to 19.2 seconds increased reported task success from 63.3% to 78.7%. A version initialized with bidirectional pretraining, which can process information in both temporal directions during training, showed no net improvement from the longer context. The authors also report that robot-domain autoregressive pretraining further raised peak success on GR-1 and LIBERO-Long.
Across comparisons with other methods, Long-WAM achieved the best reported results on the LIBERO-Long, RoboTwin 2.0 and DOMINO long-horizon manipulation benchmarks. The source does not provide every benchmark’s individual score or trial count, limiting comparisons beyond those headline results.
To keep the larger context practical for live control, the system combines streaming observation encoding, asynchronous execution and hardware-specific acceleration. It was deployed on an RTX 5090, DGX Spark and Jetson AGX Thor while retaining future-video prediction. On the RTX 5090, each action chunk, including prediction of future-video latent representations, took 107.4 milliseconds.
The researchers also tested Long-WAM in real time on Unitree G1 and YAM robots for dynamic and long-horizon manipulation. They report 95% success on dynamic cup stacking, while Pi0.5 and Fast-WAM recorded no successes in the reported 20-trial comparison. Success rates for the other physical-robot tasks were not disclosed, and the source reports no independent replication or production deployment.
Google Unveils a Cross-App Gemini Agent for Enterprise Work Google’s cloud-based agent can work across Workspace, Slack, and Microsoft 365 while maintaining context across devices. It is currently available to enterprise customers in private preview. theverge.com
Anthropic Launches Cyber Program for Critical Infrastructure and Open-Source Software Anthropic’s Cyber Mission will provide frontier models, engineers, and threat research to infrastructure defenders, while its free OSS Scanner will periodically inspect participating open-source projects. Anthropic cautions that scanner reports are model-generated without human review and may be invalid. anthropic.com
Arena Raises $200 Million at a $3.1 Billion Valuation Arena, a crowdsourced AI-model evaluation platform that began as a UC Berkeley research project, raised a Series B led by Lightspeed Venture Partners and Khosla Ventures. The company says it reached $100 million in annualized revenue in June and has added an alignment leaderboard. techcrunch.com
USA Today Seeks More Than $250 Million From OpenAI USA Today Co. and several of its local newspapers sued OpenAI, alleging that it copied hundreds of thousands of articles without permission to train AI models. OpenAI had not responded to The Verge’s request for comment. theverge.com
Fired OpenAI Researchers Contest Misconduct Allegations Jasmine Wang, Tomek Korbak, and Mikita Balesni denied mishandling sensitive information and said their dismissals could discourage safety work and external collaboration. OpenAI said an investigation found a pattern of policy violations and denied that the firings were retaliation for raising safety concerns. techcrunch.com
Anthropic Commits $150 Million to a Federal AI Science Initiative Anthropic will provide Claude products, API credits, training, and technical support to the Genesis Mission, a US federal initiative involving more than 15 agencies. The three-year commitment is intended to support several hundred research projects, including work in fusion energy and quantum computing. anthropic.com
Google Releases an Offline AI Meeting-Notes App for Macs Google AI Edge Foresight transcribes and summarizes meetings, answers questions, and references local files without sending audio or notes to the cloud, according to Google. The free experimental app uses EmbeddingGemma 2 and is optimized for Apple Silicon Macs. theverge.com
OpenAI Says It Disrupted Two “False Front” Influence Operations OpenAI reported disrupting two AI-enabled influence operations that used purported journalists and a think tank to distribute geopolitical messaging. openai.com
OpenAI Withdraws Three Mathematics Manuscripts OpenAI updated its mathematics repository with six new Lean formalizations, 19 modifications, and three withdrawals. The withdrawn manuscripts concerned Weil classes, Kuga–Satake correspondences, and the rational Hodge conjecture. twitter.com
STEPQuant Cuts Recurrent-State Serving Memory by Up to 68.7% Researchers introduced STEPQuant, a post-training quantization method for the persistent states used by linear-attention models. They report that its six-bit configuration preserved near-FP32 accuracy, compressed recurrent states by more than fivefold, and reduced total serving memory by as much as 68.7% in an SGLang implementation. huggingface.co