OpenRouter closes a $113M Series B at a $1.3B post, Anthropic acquired Stainless, 'Anthropic surpasses OpenAI to become the most val

OpenRouter closes a $113M Series B at a $1.3B post-money valuation — led by Google's CapitalG with NVIDIA's NVentures, plus ventures arms from Snowflake, Databricks, MongoDB, and ServiceNow — having grown weekly volume from 5 trillion to 25 trillion tokens in six months across 8M+ developers and 400+ models. The valuation more than doubled from ~$547M a year ago. A multi-provider router funded by both Google and NVIDIA is the clearest sign yet that the investable layer of 2026 is the chokepoint between models, not the models themselves.

Anthropic acquired Stainless — the SDK-generation startup whose software ships production client libraries for OpenAI, Google, and Cloudflare from their API specs. Owning the tool that builds your competitors' developer SDKs is a quiet distribution-layer power move, stacked on top of being the most valuable AI startup in the world after the $65B Series H. The frontier-model war is increasingly being fought through the plumbing rivals depend on, not the weights.

'Anthropic surpasses OpenAI to become the most valuable AI startup' rides the HN front page into the weekend alongside the OpenRouter raise (item 48338660) — the day's two top business stories are both about value migrating away from raw model capability toward valuation supremacy and aggregation infrastructure. The signal: capital markets are now pricing position in the stack over benchmark leadership.

'MCP is dead?' (~227 pts / 195 comments) remains the most-engaged AI tooling debate of the week and is still unresolved into May 31 — the Model Context Protocol is undergoing its first real credibility test, repricing from 'the future of AI tooling' to 'one integration pattern among several.' No verdict has formed; watch for a 'post-MCP' proposal to surface rather than an outright abandonment.

'SQLite is all you need for durable workflows' (~527 pts) and Runpod's new open-source 'Flash' SDK — which turns a local Python function into an auto-scaling inference endpoint in minutes — are the same instinct from two ends of the stack: collapse it, remove the operational overhead. The 2026 solo/small-team default is boring substrate (SQLite, Postgres) over Redis+Celery+Temporal, and one-command deploy over orchestration sprawl.

Anthropic's 'Mythos' model is reportedly finding critical vulnerabilities in legacy systems while GPT-5.5 reaches parity on threat-detection tests — and Mandiant's M-Trends 2026 confirms exploits now routinely arrive before patches, with the exploit window shrinking faster than patch cadence can compress. AI is now a first-class actor on both offense and defense, and the structural problem of May 2026 stays a deployment problem: the flaws get found faster than fixes get shipped.

GitHub trending velocity is still owned by the agent-skills/harness economy into late May — OpenClaw (Peter Steinberger) past 210k stars as one of the fastest-growing OSS projects ever, Hermes Agent (Nous Research) at 105k+, and at least 5 of the top 20 fastest-growing repos carrying 'skills' in the name with 3+ more from the Claude Code ecosystem. Fifth consecutive week the instruction layer compounds faster than the model layer.

The May 31 research run shipped with zero scraped sources — the agent-browser scraper did not produce a scraped/ directory, forcing AI-native web search only and dropping confidence on same-day HN point counts and exact GitHub stars/day. A reminder that the grounding pipeline is a soft dependency: when the timestamped scrape is missing, weekend-cadence signal leans on web-indexed trackers that lag the live front page.