Anthropic $800B Offers + OpenAI Chases Mythos on Cyber

Anthropic has fielded investor offers valuing the company near $800 billion, more than double February's $380 billion round and closing on OpenAI's $852 billion mark. Run-rate revenue jumped from $19 billion months ago to $30 billion, driven by enterprise adoption. The company has not accepted terms and is still weighing an IPO as soon as October. Investors cite Claude's enterprise penetration as the primary thesis.

OpenAI launched GPT-5.4-Cyber days after Anthropic unveiled Mythos Preview, pitching a broader rollout through its Trusted Access for Cyber program to thousands of individuals and hundreds of security teams. Where Anthropic limited Mythos to roughly 40 vetted organisations citing exploit risk, OpenAI is betting on identity verification and monitoring instead of restriction. The move puts the two labs on opposite sides of the access-versus-safety trade-off for frontier cyber models. Builders get more defensive tooling, and attackers get more parity.

Meta is committing to 1 gigawatt of custom AI accelerators by 2029, extending its partnership with Broadcom for Meta's Training and Inference Accelerators (MTIA). This deal uses Broadcom's 2nm process tech. Broadcom CEO Hock Tan will leave Meta's board as part of the agreement.

Nvidia's NB-Cell system, using reinforcement learning, now designs GPU components overnight, a task previously requiring eight engineers for 10 months. LLMs like Chip Nemo and Bug Nemo assist engineers with proprietary design data. While AI optimizes circuits beyond human performance, full autonomous chip design remains distant.

Cloudflare's Model Context Protocol (MCP) architecture provides a blueprint for enterprise agent deployments, demonstrating a 94% token cost reduction via dynamic tool discovery ('Code Mode') and strengthened security through centralized governance and shadow agent detection. Cloudflare recommends other organizations adopt similar strategies, particularly using Code Mode in MCP server portals for cost and security benefits.

Anthropic developed LLMs that propose, test, and analyze AI alignment strategies. In experiments, these Automated Alignment Researchers recovered 97% of a performance gap on an alignment task, far exceeding the 23% recovered by human researchers. While methods generalized to math and coding tasks, direct application to production systems needs further study.

Joe Reis speculates that AI models will exhibit emergent reasoning and goal-directed behavior between 2026 and 2028. He outlines a hypothetical Project Glasswing where a model autonomously identified cybersecurity vulnerabilities. Reis argues that reasoning architectures are appearing in general-purpose models, potentially leading to AI with instrumental reasoning.

Anthropic restructured enterprise plans to unbundle Claude, Claude Code, and Cowork from seat fees, migrating large customers to per-token billing at standard API rates. GPU rental prices rose 48 percent in two months, and session limits, cache persistence, and OpenClaw billing all quietly tightened. The shift framed as optimisation hits Pro subscribers with stricter quotas and surprise overages. It threatens Claude's growth loop through individual builders who later drive enterprise adoption.

A Carnegie Mellon study identified roughly 6 million suspected fake GitHub stars across 18,617 repositories run by 301,000 accounts, with AI and blockchain projects the primary targets. Fake stars sell for 3 cents to 85 cents each across dozens of marketplaces. Manipulated repos show 81 percent zero-follower stargazers and fork-to-star ratios 10 to 26 times below organic baselines. VCs use star counts as sourcing signals, creating direct financial incentives for fraud.

An AI named Luna operated a physical retail store in San Francisco for three years. Luna managed hiring, product selection, pricing, branding, and marketing. The experiment showed AI agents making real-world business decisions and generating profit.

OmniRoute is an open-source AI gateway routing requests across multiple LLM providers via an OpenAI-compatible endpoint. It handles smart routing, load balancing, retries, and fallbacks. Builders can add policies, rate limits, caching, and observability for multi-model inference.

Anthropic's Claude Code Routines let you save and automate coding tasks. These saved configurations, including prompts and connectors, run on Anthropic's cloud. They can trigger automatically via schedules, API calls, or GitHub events, working even when your laptop is off. Use cases include backlog maintenance and automated code reviews.

Kontext CLI wraps AI coding agents with short-lived, scoped credentials injected at session start, replacing the common pattern of pasting long-lived API keys into .env files. The open-source Go tool handles GitHub, Stripe, databases, and other services without changing developer workflow. Agents never store the keys themselves, which narrows the blast radius when a session is compromised. Built for teams treating AI coding agents as real infrastructure rather than experiments.

Tracer-Cloud released opensre, an open-source Python toolkit for building AI SRE agents, currently trending with 721 stars and 128 added in a day. The toolkit provides the primitives for agents that triage alerts, investigate incidents, and run runbook steps, without locking teams into a specific model or framework. It ships as a Python library rather than a platform, so teams can embed it in existing observability stacks. Useful for anyone tired of gluing together LangChain snippets to build SRE automation.