Anthropic Built Custom Version Control to Handle 30,000 Daily Agent Merges

OpenAI's Annualized Monthly Revenue Corrected to ~$50B from $70B Due to Gross vs. Net Reporting Confusion

OpenAI's reported $70B annualized monthly revenue figure was inflated due to confusion between gross and net metrics when accounting for Microsoft and AWS reseller arrangements, with the corrected figure closer to $50B. This matters because accurate revenue metrics are critical for evaluating AI company economics and valuations, especially as the market grows skeptical of "funny numbers" reminiscent of the dotcom bubble.

Meta's Personal Agent Protocol Launches Amid Retailer Pushback Over Fraud and Liability

Meta launched its Personal Agent Protocol with major tech partners, but Amazon and other retailers are blocking agent access over fraud and liability concerns. This signals a critical infrastructure battle around agent interoperability and distribution, with major implications for who controls customer relationships and ad visibility in an agentic future.

Dyna Robotics Deploys Dyna-2.1 Physical Agent with 99.4% Task Success Rate Trained on 1M+ Hours of Video

Dyna Robotics released Dyna-2.1, a semi-humanoid robot powered by the Dyna-2 diffusion transformer pre-trained on over one million hours of egocentric human video. The system achieved near-perfect reliability folding napkins at Din Tai Fung over 24-hour periods and can adapt to new tasks from as little as 45 minutes of demonstrations. The architecture combines future video prediction with robot action generation, providing concrete evidence that scaling laws apply to embodied AI similar to language models.

Google Launches Gemini 4 Argon Globally at $2/$10 Per-Token Pricing, Achieving Frontier Parity with Strong Enterprise Performance

Google rolled out Gemini 4 Argon globally to Gemini API paid tier and Pro/Ultra users, marking the first non-OpenAI/Anthropic model to clear the capability bar set by leading competitors. The model features an industry-leading 1M token output limit (up from 64K), $2 per million input tokens and $10 per million output tokens introductory pricing (doubling post-November 30), and 95% discount on cached input tokens. Argon achieves state-of-the-art performance on enterprise benchmarks: 77.9% on DeepSWE v1.1 (real-world software engineering), #1 on Vals Index (measuring economic impact across finance, coding, legal, and tax), 91.7% on LVBench (long video understanding), and 51.3% on AutomationBench (end-to-end business execution). Early independent testing reveals strong coding ability but a greater-than-50% false-claim rate on codebase analysis, with a 15% hallucination rate that ranks lowest among top-tier models. Google is rolling out Argon through its Fairwind Program with trusted cyber defenders while engaging the U.S. government's voluntary pre-release model access process.

Anthropic Scaled to 30,000+ Daily Git Merges, Built Custom Version Control System to Handle Agent-Driven Development

Anthropic scaled from 2,500 to over 30,000 code changes merged per day between May and September 2026, then built and deployed Scion, a custom in-house version-control system, to replace Git entirely in a five-month migration. The company also built complementary infrastructure including Antique (code review surface), code search, and rich source control IAM. Scion is designed around the assumption that clients (agents and CI jobs) are always online, uses server-side commits with no push step, and enables sparse and virtualized checkouts to avoid unnecessary data transfer—the monorepo's compressed working copy alone exceeds 25 GB and history is nearly 100 GB. This is concrete evidence of how agentic development is reshaping internal tooling at scale; Anthropic notes that "coding is no longer the bottleneck" and that the infrastructure built for human-scale development has become the constraint.

Microsoft's Maia 300 Inference Chip Prioritizes Bandwidth for Large-Scale Agentic Deployments

Microsoft unveiled the "Clea" Maia 300 XPU as a bandwidth-optimized inference accelerator designed for serving large language models and agentic workloads at scale. The chip features 288 GB of HBM4 memory, ~166 TB/sec of aggregate SRAM bandwidth, eight HBM4 stacks, and a chiplet architecture moving to TSMC's 2nm N2P process. It scales coherent clusters to 6,144 devices (up 2.7x from Maia 200) while increasing power consumption only 1.5x, and supports larger agentic workflows consisting of dozens to hundreds of models working in concert. Microsoft signaled ambitions for "multi-gigawatt deployments" on Azure (roughly 575,000 Maia 300 sockets per gigawatt at 1.74 kW per device), with Maia 200s already running Copilot, GitHub Copilot, and OpenAI model inference. This signals Microsoft's strategy to build differentiated silicon for inference serving rather than competing directly on training performance.

Introducing Beam: Reflection's 501B-Parameter Open-Weight MoE Model with High-Compute RL

Reflection AI released Beam, a 501B-parameter sparse MoE open-weight model optimized for coding and agentic workloads. Main takeaways:

• Beam trained on 23.8 trillion tokens with 100+ million rollouts across 10.5K NVIDIA GB300 GPUs over 4 weeks, representing one of the largest-scale RL runs by an open lab
• Achieves competitive performance with GLM 5.2 and approaches Qwen 3.8-Max on coding/agentic tasks while using 3-4x less inference compute on reasoning benchmarks
• Developed asynchronous RL algorithms and infrastructure supporting 110K concurrent rollouts, 170K concurrent sandboxes, and 12-second model weight distribution across fleet
• Implements controllable reasoning effort parameter allowing users to trade off response length against capability; demonstrates generalization beyond training domains (e.g., organic web search and OCR adoption)
• Sustained 1.3 billion sandbox evaluations and sourced one million high-quality coding/agentic/STEM environments through iterative curation; training showed no saturation plateau

Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models

Dyna Robotics published a technical report on world-action models (WAMs) pre-trained on 1M+ hours of egocentric human video. Main takeaways:

• First demonstrated cross-embodiment transfer scaling law: robot performance improves monotonically with human pre-training data scale, with inflection point between 10k-100k hours
• Video co-training (future prediction on unlabeled video) is the primary driver of cross-embodiment generalization; scaling video data alone improves held-out robot performance while action-only training fails
• Dyna-2 achieves 87% production pass rate on unseen customer sites vs. 46% for Dyna-1 at identical post-training budget, demonstrating real-world deployment viability
• Model trained on 1M hours (170 years continuous waking experience) with 39-task zero-shot robot evaluation across bi-manual manipulation tasks (cloth, knots, packing, assembly)

NTT Data to Build First Dedicated Data Center for Anthropic in Japan; 500MW Osaka Facility with ~$6.3B Construction Cost

Nikkei Asia reported on NTT Data Group's plans to construct a dedicated data center for Anthropic in Osaka prefecture, Japan. Main takeaways:

• Facility targets 500 MW power capacity, making it one of Japan's largest data centers, with construction costs potentially reaching 1 trillion yen (~$6.3B)
• Agreement expected by end of fiscal 2026 (March 2027), with partial operations targeted for 2028
• Power sourcing will rely on alternatives including on-site generation and fuel cells due to grid constraints
• NTT Data is world's third-largest data center operator by market share; Anthropic is evaluating multiple Asian countries for expansion (also in talks with Australian partners)
• Japanese government has set goal of mobilizing 32.7 trillion yen in data center and cloud computing investment by fiscal 2035

Redesigning Source Control: Anthropic's Scion System for Agentic-Scale Code Merging

Anthropic published a technical deep dive on Scion, its custom source control system built to handle agentic code generation at scale. Main takeaways:

• Anthropic's code merge rate increased from 2,500 changes/day in May 2026 to 30,000+/day by September 2026, rendering Git's distributed architecture fundamentally unscalable for their workload.
• Scion replaces Git's decentralized model with a server-centric architecture assuming clients are always online; the monorepo history alone is ~100GB with compressed working copies exceeding 25GB, generating 600PB+ daily traffic under naive Git protocols.
• Core design principles include lock-free server-side commits scaled to changeset size (not repo size), derived data structures for blame/history computed on-demand, and sparse/virtualized checkouts to eliminate unnecessary data transfer for CI jobs performing 25M checkouts/day.
• Scion's changeset model stores only file deltas with content-addressed pointers to blob storage, enabling high-throughput acceptance without packfile coordination; horizontally scalable servers cache immutable data with minimal cross-server coordination.
• The system demonstrates how AI-driven development fundamentally breaks assumptions in tools designed for human-scale workflows, with implications for infrastructure buildout across the AI stack.

Microsoft's 'Clea' Maia 300 Inference XPU Is A Bandwidth Beast

The Next Platform published a detailed technical analysis of Microsoft's Maia 300 custom XPU accelerator. Main takeaways:

• Maia 300 moves to TSMC's 2nm N2P process with chiplet architecture (seven compute/IO dies plus eight HBM4 stacks), up from monolithic Maia 200 design at 3nm
• HBM capacity increases to 288GB with 18 TB/sec bandwidth; estimated 166.2 TB/sec aggregate SRAM bandwidth across compute and I/O dies
• Networking architecture scales to 6,144 devices in single coherency domain (2.7x larger than Maia 200) with improved ATL Ethernet implementation; supports multi-gigawatt deployments
• Microsoft targeting 1M+ Maia sockets in field; Maia 200s ramping to production for Copilot, GitHub Copilot, and OpenAI model inference by end of 2026
• Estimated ~280B transistors, ~1,980 mm² die area; architecture optimized for agentic workloads requiring dozens-to-hundreds of coordinated models