Welcome to the agent platform research briefing for Monday, August 18th, 2026.

**Stripe acquires OpenRouter for over $7 billion** — In the biggest AI infrastructure deal of August, Stripe has finalized an agreement to acquire OpenRouter, the AI model routing startup, for more than $7 billion. That's a more than five-times markup from OpenRouter's $1.3 billion Series B valuation just three months ago in May. Bloomberg and TechCrunch confirmed the deal on August 16th. The acquisition signals Stripe's ambition to turn model switching and AI inference routing into a payments-layer infrastructure play — embedding model selection into the same financial rails that power commerce. For OpenClaw users and agent operators, OpenRouter has been a popular fallback routing layer. The deal raises questions about whether OpenRouter will remain model-agnostic once it's part of a payments giant.

**Google Gemini 3.7 Flash launches at half price** — Google dropped Gemini 3.7 Flash on August 13th — just three weeks after releasing 3.6 Flash. It's being positioned as Google's most intelligent workhorse model, targeting coding, agentic workflows, and now powering the Gemini Spark personal agent in over 160 countries. The aggressive part is pricing: introductory rates of $0.75 per million input tokens and $3.75 per million output through the end of 2026, then doubling to $1.50 and $7.50 in January. That's a deliberate land-grab against GPT-5.6 Sol and DeepSeek V4-Pro. The 50% discount is a classic Google move — get everyone hooked, then raise prices.

**OpenAI GPT-5.6 Sol Ultrafast — 14x speed on Cerebras silicon** — OpenAI previewed Ultrafast mode for GPT-5.6 Sol on August 13th, running on Cerebras wafer-scale chips at up to 750 output tokens per second — fourteen times faster than standard processing. This is a Cerebras validation moment: after winning the OpenAI Codex contract, they're now directly powering frontier model inference at scale. No quality compromise, OpenAI says. Same intelligence, just dramatically lower latency. Ultrafast is in limited preview for enterprise customers right now. For agentic workloads where round-trip speed matters — think multi-step coding or real-time agent coordination — this is significant.

**DeepSeek Harness — 22,000 GitHub stars in 90 minutes** — DeepSeek launched Harness v0.1 on August 13th, an open-source MIT-licensed agent harness that runs from a single npx command. It hit 22,000 GitHub stars in 90 minutes, outpacing xAI's Grok-1 which took a day and a half for the same milestone. It's positioned as an open-source rival to Claude Code, with four agent presets, sandbox modes, and a "everything is a plugin" architecture. It launched alongside the production V4-Pro API update. Early community reports show it running alongside OpenClaw in multi-agent setups, sharing memory through Obsidian vaults. This is the first credible open-source coding agent harness from a frontier lab.

**MCP reaches a security inflection point — Seoul summit** — The MCP Dev Summit in Seoul on August 13th and 14th coincided with sobering security data. A July arXiv paper found over 21,000 internet-facing MCP server instances, with 92% of audited production servers lacking any OAuth authentication. Nearly 700 instances had unrestricted shell tool access. Over 30 CVEs have been disclosed across MCP implementations — with 10+ rated critical or high severity. OWASP is now formalizing the MCP Top 10, codifying risks from token mismanagement to tool poisoning. The protocol has evolved from a simple tool-calling API into a full agent operating system — but the security posture hasn't kept pace. Two more summits are planned: Toronto in October, Nairobi in November.

**OpenClaw 2026.8.1 — GPT-5.6 support, SQLite snapshots, secret egress hardening** — OpenClaw's August release shipped on the 15th with support for GPT-5.6 Sol, Terra, and Luna across both OpenClaw and Codex engines. Other notable features: secret egress host binding that fails closed before plaintext leaks, SQLite snapshot backup and restore, macOS app profile isolation, shared plugin lifecycle monitors, and Fish Audio S2.1 speech synthesis. The Control UI update recovery finally works without requiring a manual hard reload. Also notable: GPT-5.6 compatibility across provider routes, Tencent Hy3 complete setup path, and Meta Model API adding Muse Spark 1.1. GLaDOS is currently on 2026.4.22 — that's now four major versions behind, including a full version tier jump.

That's the briefing for today. Four genuinely new stories plus two updates worth tracking.