Stripe–OpenRouter & Anthropic $65B: AI News Aug 22, 2026

1. Today's Headlines

• Stripe confirms a $7.5B OpenRouter acquisition. Payments giant Stripe officially confirmed its acquisition of OpenRouter, the AI model-routing platform that sends developer prompts across many model providers. While terms were not released, the New York Times reports the deal is valued at $7.5B — a surge from OpenRouter's $1.3B valuation in May 2026. Founders receive $1.5B, with the remaining $6B going to investors; Stripe outbid Databricks. (Source: The New York Times; subject to vendor confirmation)
• Anthropic routes Claude Mythos into Claude Security. Per a Sina Finance 7x24 flash on Aug 22, Anthropic said Claude Mythos is now deployed inside Claude Security, and ClaudeSecurityScans can now run on Claude Mythos 5. (Source: Sina Finance; subject to vendor confirmation)
• OpenAI fully open-sources the Codex Harness. OpenAI published the core Codex runtime — the agent loop and execution logic behind its coding agent — on GitHub (github.com/openai/codex), with a long write-up of how the agent loop works and design decisions left in issues and pull requests. (Source: subject to vendor confirmation)
• OpenAI is testing a new image model (GPT Image 2.5). Multiple users on LMArena spotted a model called luna-lisa-alpha — believed to be an early GPT Image 2.5 — improving person consistency, realism and text rendering; the GPT Image 2 API preview also added transparent-background generation. OpenAI has not confirmed the model. (Source: subject to vendor confirmation)

2. Model Release Roundup

• Google's Gemma open models pass 1 billion downloads. Google DeepMind said its Gemma family of open models has surpassed one billion downloads, with more than 100,000 community variants and fine-tunes since launch about two years ago; an "Awesome Gemma" GitHub repo organizes notable community projects. (Source: Google DeepMind)
• Liquid AI ships LFM2.5-DSpark for up to 3.18x faster inference. Released Aug 20, three speculative-decoding draft models (1.2B-Instruct, 2.6B, 8B-A1B) add a path that lifts throughput up to 3.18x on an H100 and 2.87x on an M4 Max with identical outputs; function-calling latency drops 57% on average. Weights are on Hugging Face with day-one llama.cpp and SGLang support. (Source: Liquid AI / Hugging Face)
• Mistral launches Agentic Search. Mistral introduced Agentic Search, a multi-step retrieval system for complex enterprise documents (contracts, manuals, cross-file reports) that assembles a usable answer instead of a list of links. (Source: Mistral)
• Meta ships a Mac app powered by Muse Spark. Meta released a dedicated Meta AI Mac app built on its Muse Spark model, with system-wide dictation that transcribes voice across any macOS app and multimodal screen analysis of open windows. (Source: subject to vendor confirmation)
• xAI rolls Grok Build out to everyone. xAI moved Grok Build from beta to general availability across web, iOS and Android for all tiers; users describe an idea and Grok builds a runnable web app, game or dashboard, publishable to a grok.me link. (Source: xAI)

3. Industry & Capital

• Anthropic's run rate hits $65B. In an investor update, Anthropic said its annualized revenue run rate reached roughly $65B — up from about $9B at end-2025 — with Q2 revenue above $11.5B and positive adjusted operating income; a fall IPO is targeted at a valuation up to $2T, ahead of OpenAI's $40B run rate. (Source: Bloomberg / CNBC)
• Binance launches Agent OS. Binance rolled out Agent OS, a platform letting autonomous AI agents analyze market data and execute trades for its 300 million users, with MCP support and user-managed sub-accounts for risk isolation; agents can monitor live feeds and run real funds under user-set limits. (Source: Binance)
• Model routing becomes a new infrastructure layer. Ramp released Router, a self-built AI model-routing service across OpenAI, Anthropic and DeepSeek with cost/latency/fallback dashboards (free this year); NVIDIA open-sourced Switchyard, an LLM traffic-routing framework. The "token transaction layer" is consolidating. (Source: TechCrunch / NVIDIA)
• Micron commits $10B to a U.S. R&D lab. Micron said it will spend $10B over ten years on a research lab in Boise (construction from 2027), separate from its existing $250B commitment, focused on memory, compute architecture, packaging and manufacturing. (Source: subject to vendor confirmation)

4. China Power

• Zhipu GLM-5.3 tops cybersecurity benchmarks. GLM-5.3 scored 84.5% on CyberGym for vulnerability discovery — slightly ahead of Anthropic Mythos 5 and OpenAI GPT-5.6 Sol — and found 2,436 vulnerabilities across 269 real projects (1,097 medium-to-high); weights are due to open in about two weeks. (Source: subject to vendor confirmation)
• GLM-5.3 and DeepSeek-V4-Pro land on Alibaba's Qwen platform. Alibaba Cloud's MaaS expanded its model matrix, adding Zhipu GLM-5.3 and the stable DeepSeek-V4-Pro to API access, covering coding and agent scenarios alongside Qwen3.8-Max and Kimi K3. (Source: subject to vendor confirmation)
• SenseNova U1.5 Lite goes open. SenseTime open-sourced SenseNova U1.5-Lite-Preview, an 8B MoT model on a native unified multimodal architecture (understanding, reasoning and generation in one model), trained to 4K resolution; it rose from 47.14 to 55.20 on Qwen-Image-Bench. (Source: subject to vendor confirmation)
• DeepSeek V4-Pro time-of-day pricing takes effect. From Aug 17, DeepSeek V4-Pro output rose to 27 yuan per million tokens at peak (from 6 yuan, about +350%) and 13.5 yuan off-peak; V4-Flash output rose to 9 yuan at peak (from 2 yuan). The company cited capacity allocation; the DeepSeek Harness open-source release cleared 100,000 GitHub stars in roughly two days. (Source: Caixin / subject to vendor confirmation)
• Tencent and ByteDance move. Tencent's Hunyuan Hy3 offers an 8x usage discount and is open under Apache 2.0 on OpenRouter; ByteDance's Doubao added "skills · connectors · work partners" to its task tool, and Meituan extended LongCat's free trial by two weeks. (Source: subject to vendor confirmation)

5. Today's Observation

The throughline of Aug 22 is the rise of the "token transaction layer." Stripe buying OpenRouter for $7.5B, Ramp's Router and NVIDIA's Switchyard all point to the same thing: as models commoditize, the value is shifting to routing, cost control and reliability across providers. At the same time, agents are crossing from assisting into executing with real money (Binance Agent OS) and real security mandates (ClaudeSecurityScans on Claude Mythos 5), while DeepSeek's price hike signals that even the cheapest lab is capacity-constrained. The story is no longer "who has the biggest model" but "who controls the plumbing, who is safest, and who counts best."

Editor's Take

The Stripe–OpenRouter deal is the headline, but the detail I keep returning to is Anthropic routing Claude Mythos 5 into Claude Security. That is the first time a frontier lab has pointed its strongest reasoning model at its own security scanners — a quiet admission that autonomous security review is now a model-quality problem, not just a tooling one. Meanwhile, OpenAI open-sourcing the Codex Harness tells me the moat was never the code; it was the accumulated engineering judgment, now given away to pull developers onto the platform. I'd watch the routing layer more than any single new model this quarter.

Date: 2026-08-22. Sources gathered via WebSearch and vendor news pages; no local scripts or user API keys were used.