Anthropic Revenue Hits $16.8B Ahead of IPO, Up Nearly 50% QuarterlyModal Hits $1B ARR with $16B Series D, Valuation Triples in 5 MonthsOpen Weight Models Drive 71.5% of Tokens but Only 23% of Spend: VercelVercel Launches Enterprise Tier as AI Gateway Token Usage QuadruplesOpenAI Launches GPT-Rosalind Academic Plan for Life Sciences LabsGoogle Uses Antigravity to Accelerate Antimicrobial Peptide DiscoveryVercel Drives Offers Persistent Storage for AI Agents Across EnvironmentsVercel Adds Python, Container, and Daemon Backend SupportVercel Launches General Availability for Python App DeploymentGitHub Says the Next Developer Interface Is Ambient IntelligenceClawLoop Enables Self-Improving LLM Agents via Native Training LoopHarness^2 Enables Recursive Self-Improvement for LLM Agents in Open WorldsEcho Cuts LLM Tensor-Parallel Communication Cost with 4-Bit Error CompensationHERO Optimizes Prompts and Scaffolding to Improve Code GenerationJoint Subspace Sparse Attention Cuts Long-Context LLM Decoding CostsInverse Reasoning Models Enable Post-Training from Answer-Only DataChirp Pong: Two-Phone Game Uses Microphone Chirps for NetworkingOpenAI Launches Brazilian Portuguese AI for Educators with Study ModeGitHub Automates External Collaborator Access Review and Repo CleanupAnthropic Shares Prompt Tips; Google's New Model Gets Mixed First ImpressionsAnthropic Revenue Hits $16.8B Ahead of IPO, Up Nearly 50% Quarterly

AI infrastructure company Modal announced $1 billion in annual recurring revenue and closed a $16 billion Series D, becoming New York City’s newest AI infra decacorn. Usage and valuation both tripled in five months and grew 15x in one year. Modal provides sandboxed cloud development environments for rapid AI application deployment and scaling.

Vercel’s AI Gateway published its October 2026 Production Index, drawing on trillions of tokens routed monthly between production apps and AI labs. Open weight models account for 71.5% of tokens but only 23% of spend, revealing a sharp cost gap between open and closed models. The index gives enterprises a real-world view of AI usage distribution and cost structure across production workloads.

Vercel launches an enterprise tier targeting the rapid growth of AI adoption beyond engineering teams. Token usage by enterprise companies on Vercel’s AI Gateway has quadrupled in the last three months. OpenAI separately reported that non-developer Codex users grew 189x from August 2025 to June 2026. Vercel’s customer base is expanding from engineering teams to non-technical departments. While specific enterprise features and pricing were not detailed, the move signals Vercel’s push to extend its AI infrastructure from developer tooling to enterprise-grade platform services.

OpenAI introduced the Rosalind Academic Plan, offering eligible academic labs sponsored access to GPT-Rosalind and supporting tools for life sciences research. The program lowers barriers for academic researchers seeking to apply advanced AI models in biology and medicine. Application details and eligibility criteria are available on OpenAI’s website.

Google DeepMind published a blog detailing how it uses Google Antigravity for Science to accelerate the search for novel antimicrobial peptides. These peptides are key candidates in the fight against antibiotic resistance, but traditional discovery is slow and costly. The platform speeds up candidate screening and evaluation, helping researchers identify promising antibacterial molecules faster.

Vercel launches Drives, a persistent storage service for AI agents. Agents often handle complex work spanning multiple runs and may need different tools, runtimes, or compute when switching environments. Vercel Drives keeps an agent’s files accessible across environment changes, ensuring continuity. The product addresses file state loss in dynamic environments, giving developers a reliable persistence layer for long-running automated tasks.

At Vercel Ship in San Francisco, Vercel announced expanded backend support including Python, containers, and Daemons. The platform previously added zero-config support for Python and TypeScript backend frameworks, enabled joint frontend-backend deployment via Vercel Services, and offered queues and durable workflows for background tasks. The latest update extends coverage to containerized workloads and long-running processes, letting developers deploy everything from lightweight functions to containerized services on a single platform.

Vercel has announced the general availability of Python on its platform. Developers can now build and run Python applications using their existing frameworks, libraries, and tools. Vercel automatically detects supported frameworks and configures the build and deployment process. Applications run on Fluid compute, which manages concurrent requests. This means Python developers can deploy with the same experience as the JavaScript ecosystem, without manual server or build configuration.

GitHub published a blog post arguing that AI, agents, voice, and physical controls are shifting software development from operating tools to expressing intent. Instead of navigating interfaces step by step, developers will state goals in natural language and let systems execute automatically. This shift aims to keep developers focused on building rather than tool manipulation.

arXiv paper ClawLoop proposes a harness-native training loop that enables LLM agents to self-improve while operating in sandbox workspaces with direct file system and shell access. As agents increasingly run in general-purpose harness environments, the challenge of continuous capability improvement without manual intervention remains. ClawLoop embeds the training loop directly into the agent harness, allowing iterative refinement during task execution. This approach could reduce manual tuning for developers building autonomous workflows, enabling agents that get better at their jobs over time.

arXiv paper Harness^2 proposes a recursive agent harnessing framework for open-world self-improvement. The research notes that LLM agent capabilities depend not only on underlying models but also on execution harnesses. Recursive self-improvement (RSI) can automatically evolve these harnesses, but deployment in open-world settings remains challenging. Harness^2 uses recursive mechanisms to let agents continuously optimize their execution frameworks in open environments, offering a new approach for building autonomous agents that adapt to unknown tasks.

arXiv paper Echo proposes an error-compensated 4-bit tensor-parallel communication scheme to reduce communication overhead in LLM training and inference. Tensor parallelism places an all-reduce between two dependent matrix multiplications at every layer, making its cost unhideable and increasingly expensive as compute outpaces interconnect bandwidth. Echo compresses communication from 16-bit to 4-bit while compensating for quantization errors, cutting data volume without significant accuracy loss. The method has direct practical value for large-scale distributed training.

arXiv paper HERO proposes an evidence-driven approach for test-time refinement and prompt optimization in program generation tasks like text-to-SQL and text-to-Python. Existing harness optimization methods typically rely on LLM reflectors to propose updates, with limited effectiveness in program generation. HERO collects execution evidence to guide iterative refinement of prompts and scaffolding, improving generated code correctness. The method suits automation scenarios requiring high-accuracy code generation.

A new paper introduces Joint Subspace Sparse Attention for efficient long-context LLM decoding. Sparse attention reduces KV-cache traffic, but existing methods operate at coarse granularity. This work uses token-level selection for finer control, addressing challenges such as information loss in low-dimensional scoring. The proposed joint subspace approach aims to reduce computational overhead in long-context scenarios while maintaining decoding quality.

A new paper proposes Inverse Reasoning Models to address the tradeoff between supervision cost and compute efficiency in LLM post-training. Traditional supervised fine-tuning requires reasoning traces, but many applications only have prompt-answer pairs. This method enables post-training and verification from answer-only data, eliminating the need for full reasoning process annotations. It offers a new approach for training reasoning capabilities in data-scarce scenarios, reducing reliance on manually annotated reasoning chains.

Developer Simon Willison released Chirp Pong, an experimental two-player pong game running in static HTML. Two phones communicate over microphone chirps instead of Wi-Fi or Bluetooth, with no server required. The project demonstrates that acoustic signaling can support real-time multiplayer gaming in a purely static web page.

OpenAI launched a Brazilian Portuguese version of AI for Educators to celebrate Teachers’ Day in Brazil, showcasing how Study Mode helps teachers support active, step-by-step learning. The tool targets educational settings, helping teachers design more interactive classroom experiences. Study Mode guides students through reasoning steps rather than giving direct answers, fostering independent learning. The Portuguese launch signals OpenAI’s expansion into non-English education markets, with local teachers able to use the tool immediately for daily instruction.

GitHub published a blog post detailing how to automate secure access and repository hygiene in a shared organization using native platform features. The approach covers regularly reviewing outside collaborator access and archiving stale repositories. The post provides concrete implementation steps to reduce manual audit work and lower security risks from lingering permissions. Teams managing multiple repos and large numbers of external collaborators can adopt this automated workflow directly.

Anthropic published a prompt engineering tips guide to help users get better results from Claude. Meanwhile, Google released a new model that received mixed early reviews, with Ben’s Bites noting it didn’t meet expectations in a first test. Both updates highlight ongoing variance in prompt effectiveness and real-world model performance.