AI Briefing: May 26–27, 2026

🚨 Breaking (last 24h)

OpenClaw Ships v2026.5.26-beta.1 — Major Performance & Voice Release

OpenClaw released a substantial beta (33,702-character changelog) on May 26 with sweeping improvements across performance, voice, channels, and observability:

Faster replies and startup: Visible reply delivery now separates user-facing sends from slower follow-up work; command/model/plugin metadata is cached on hot paths; Gateway startup avoids repeated plugin, channel, session, and filesystem scans.
Better voice and Talk: Realtime Talk runs can be inspected, steered, cancelled, or followed up from Web UI and Discord voice; wake-name handling is more tolerant without letting ambient speech trigger agents.
More production-ready channels: Telegram keeps typing/progress context and forum topics; iMessage handles attachment roots and duplicate local Messages sources; WhatsApp restores group/media behavior; Discord improves voice playback and model picking; Signal and iMessage get reaction approvals.
Safer agents: Codex app-server auth improvements, sandbox path handling, usage-limit recovery, and OpenAI-compatible provider fixes for empty-tool and malformed payload failures.
New observability: Activity tab, gateway secret-prep traces, tool/model stream progress, OpenTelemetry LLM spans, and richer missing telemetry signals.
Media backend overhaul: Replaced Sharp with Rastermill for metadata, resizing, EXIF orientation, and PNG alpha-preserving optimization.
Platform support: Alpine Linux CLI installs, Windows native improvements, iOS Talk mode enhancements, Android pair-new-gateway action.
Source: GitHub Release

OpenClaw v2026.5.25-beta.1 — Late Fixes Beta

Released earlier on May 26 with targeted fixes:

iMessage attachment roots threaded into the image tool; deduped watcher startup for duplicate accounts
Codex sandbox bootstrap path style preservation
Alpine Linux apk Node.js/npm/Git support
Windows cross-platform launcher for Node package scripts
Source: GitHub Release

OpenRouter Raises $113M Series B at $1.3B Valuation

Popular AI model gateway OpenRouter, founded in 2023, raised a $113 million Series B led by CapitalG (Google's growth venture fund). The round values the company at approximately $1.3 billion post-money — more than double the ~$547 million valuation from its Series A a year ago. OpenRouter provides access to 400+ models from Anthropic, Google, OpenAI, xAI, and DeepSeek, serving 8 million global users and processing 100 trillion tokens per month (a 5x increase from six months ago).

Source: TechCrunch | NYT

Greg Brockman: "GPT-5.5 is a uniquely good coding model"

OpenAI co-founder Greg Brockman continued his heavy Codex and GPT-5.5 promotion on May 26, tweeting that GPT-5.5 is "a uniquely good coding model" (589 likes) and describing a "magical experience with codex on iPad" (653 likes). This follows a week of high-engagement posts including "under appreciated that codex is open source" (5,100 likes), "self improvement prompt for codex" (3,938 likes), and the strategic declaration "the model alone is no longer the product" (7,524 likes).

Source: @gdb on X | Thread

📊 Market Moves (last 48h)

Anthropic Appoints KiYoung Choi as Korea Representative Director

Anthropic appointed KiYoung Choi as Representative Director of Korea on May 26, ahead of opening a Seoul office. Korea is one of Claude.ai's most active markets — Koreans use Claude at more than 3.5x the rate expected for the population size, with usage skewing heavily toward technical and creative work. KiYoung joins from Snowflake where he served as General Manager for Korea.

Source: Anthropic Blog

Anthropic Co-founder Chris Olah Speaks at Vatican on AI Ethics

On May 25, Anthropic co-founder Chris Olah spoke at the presentation of Pope Leo XIV's encyclical "Magnifica Humanitas" in Vatican City. Olah participated as part of Anthropic's initiative to "widen the conversation on the important questions raised by AI." The encyclical addresses safeguarding the human person in the time of artificial intelligence.

Source: Anthropic Blog

DuckDuckGo Installs Surge 30% as Users Reject Google's AI Search

Privacy-focused search engine DuckDuckGo saw U.S. app installs rise 18.1% week-over-week during May 20–25, peaking at 30.5% on May 25, following Google's massive Search overhaul announcement. iOS install growth hit a 33% average, peaking at 69.9%. Visits to DuckDuckGo's AI-free search page (noai.duckduckgo.com) averaged 22.7% WoW growth. CEO Gabriel Weinberg stated: "Google is force-feeding AI with no way to opt out."

Source: TechCrunch

Uber President Says AI Spending Getting "Harder to Justify"

Uber's president said AI spending is getting "harder to justify" in a May 26 report, noting there's no clear connection between AI usage and productivity. The comments reflect growing enterprise skepticism about AI ROI despite massive investment.

Source: The Verge

🔬 Research (last 7 days)

arXiv May 25 Batch — 15 Papers

The most recent arXiv batch (May 25, 2026) includes notable papers:

MobileGym (2605.26114): A verifiable and highly parallel simulation platform for mobile GUI agent research — browser-hosted environment for everyday mobile use.
From Model Scaling to System Scaling (2605.26112): Argues the next major bottleneck in agentic AI is system scaling — auditable, persistent, modular architectures around foundation models.
Language Models Need Sleep (2605.26099): Proposes a sleep-like consolidation mechanism for LLMs to handle long-horizon tasks without attention scaling poorly.
Claw-Anything (2605.26086): Benchmarking always-on personal assistants with broader access to users' digital worlds.
VeriTrace (2605.26081): Evolving mental models for deep research agents facing vast, interdependent, uncertain information.
Automated Benchmark Auditing (2605.26079): Framework for verifying AI benchmarks at scale, addressing implicit assumptions and incomplete specifications.
Source: arXiv API

Apple macOS Kernel Vulnerability Found by Claude

CVE-2026-28952, an authorization issue in macOS Tahoe 26.5 that could allow an app to gain root privileges, was discovered by Calif.io "in collaboration with Claude and Anthropic Research." The vulnerability was addressed with improved state management in Apple's May 2026 security update.

Source: Apple Security | Hacker News

🛠️ Tools (last 48h)

OpenClaw v2026.5.26-beta.1 — Major beta with performance, voice, channel, and observability improvements. Release notes
OpenClaw v2026.5.25-beta.1 — Late fixes beta with iMessage and Alpine improvements. Release notes
Gemini 3.5 Flash — Released according to Twitter reports; described as "noticeably more powerful than Gemini 3.1 Pro, but token prices have tripled." Source: @0xSoulAI on X

💭 Industry Pulse (last 48h)

@gdb (Greg Brockman) continued his Codex promotion campaign with multiple high-engagement tweets about GPT-5.5's coding capabilities and Codex use cases. The sustained promotional push signals OpenAI's strategic priority on productizing Codex beyond raw model access.
@ylecun (Yann LeCun) posted a viral thread (3,050 likes) on May 25 distinguishing engineers from scientists: "an engineer, given a problem, invents and tries multiple solutions and picks the one that works." He also defended the peer review system against criticism.
Nolan Lawson published "Using AI to write better code more slowly" (topped HN May 26), arguing LLMs can be used for high-quality code written more carefully rather than as "slop cannons." He cites Claude's effectiveness at finding bugs when used deliberately.
HN May 26: Top AI-related stories included the Nolan Lawson essay, the Apple CVE found by Claude, Uber's AI spending concerns, and Spain blocking prediction markets.

📡 Sources & Data Provenance

Source	Status	URL
Twitter/X API	✅ Working	https://twitterapi.io
GitHub Releases	✅ Working	https://github.com
arXiv API	✅ Working	https://arxiv.org
Wiki Raw Archive	✅ Available	~/wiki/raw/
TechCrunch	✅ Working	https://techcrunch.com
Hacker News	✅ Working	https://news.ycombinator.com
Web Search	✅ Available	Built-in

Each story above links directly to its primary source. Unlinked claims were cross-referenced from multiple sources.