OpenAI announced a preview of Ultrafast, a new service tier that runs its flagship GPT-5.6 Sol model up to 14 times faster than the standard offering, delivering up to 750 output tokens per second. The speed comes from a partnership with Cerebras, a chipmaker that builds unusually large processors optimized for AI. In plain terms: tasks that used to feel like waiting for a page to load will now feel close to instant, which matters most for AI agents that chain many steps together. OpenAI paired the launch with a new “builder’s guide to GPT-5.6” aimed at startups designing faster, cheaper agents. Expect competitors — including inference-speed specialists like Groq and Cerebras’s other partners — to respond quickly.
Starting today, August 14, Anthropic flips Claude Code’s Auto Mode on by default for Pro, Max, and Team users. Instead of asking permission before every action, Claude now proceeds automatically unless a built-in safety check judges an action to be irreversible, destructive, or outside your environment. Anthropic says that in testing with 1,053 paid users, the automated check caught 89% of harmful actions versus just 13.6% caught by humans clicking through approval prompts — an argument that people rubber-stamp prompts anyway. The change generated a lively Hacker News discussion and was the lead item in several of your newsletters today. If you use Claude Code or Cowork, it is worth knowing where the setting lives in case you prefer manual control for sensitive work.
Anthropic is reportedly negotiating to buy Decart AI, an Israeli startup specializing in real-time generative video, “world models” that simulate interactive environments, and GPU optimization technology, in a deal valued around $6 billion. If completed, it would be Anthropic’s largest acquisition and a signal that the company wants to compete beyond text — in the video and simulation territory where OpenAI’s Sora and Google’s Veo have dominated headlines. For communicators, real-time generative video moving into mainstream enterprise hands raises both creative opportunities and deepfake-era verification headaches.
Elon Musk’s xAI released Grok 4.6, with early benchmark chatter suggesting it competes at the level of OpenAI’s GPT-5.6, and Musk claiming Grok 4.7 is only weeks behind. The release lands alongside “Grok Bot,” xAI’s new agent product that gives AI coworkers their own cloud computers to run applications independently. The pace here is the story: frontier labs are now shipping meaningful model updates on a cycle measured in weeks, not quarters, which keeps the competitive narrative — and the media cycle — in constant motion.
Anthropic began embedding invisible watermarks in text generated by Claude — a statistical “fingerprint” woven into word choices that lets detection tools identify AI-written content without changing how the text reads. Several of your newsletters framed it memorably (“Claude’s generated texts are getting secret tattoos”). This is a big deal for communications teams: it points toward a future where AI-drafted press materials, statements, and articles can be flagged as machine-generated by platforms, journalists, or verification tools. Worth watching how detection tools roll out, whether other labs follow, and what it means for disclosure norms in PR.
OpenAI publishes “The builder’s guide to GPT-5.6”A playbook showing startups how to build faster, more cost-efficient AI agents with smarter model selection and new Responses API capabilities — companion piece to the Ultrafast launch.
OpenAI · Aug 13
02
“Grok Bot” gives AI coworkers their own cloud computersxAI’s new agent product provisions each AI worker with its own virtual machine to run apps end-to-end — the latest step in the industry’s shift from chatbots to autonomous “AI employees.”
OpenAI appoints Dali Rajic as Chief Revenue OfficerRajic, a veteran enterprise sales leader, will run OpenAI’s global revenue organization — a sign the company is professionalizing its enterprise go-to-market as competition for corporate AI budgets intensifies.
OpenAI · Aug 13
06
Lovable raises $400M Series C at a $13.3B valuationThe Swedish “vibe-coding” platform — which lets non-programmers build software by describing it — was led by Menlo Ventures and the Scaleup Europe Fund. AI code-review firm CodeRabbit also raised $143M at ~$1.5B.
California holds “suspense” votes on roughly 30 AI billsBoth the Assembly and Senate held decisive appropriations votes August 13 on a stack of AI bills covering transparency, automated decision-making, and training data — decisions that often set the template for other states.
Transparency Coalition
09
EU AI Act high-risk obligations are now in forceAs of August 2, 2026, obligations for many high-risk AI systems — including credit scoring and insurance pricing — moved from roadmap to reality, with companies operating in Europe facing new transparency requirements.
European Commission
10
85 new AI laws passed across 27 US states so far in 2026California, Colorado, and New York lead a patchwork of state AI legislation covering automated decision-making and training data transparency — a compliance maze for any company operating across the US.
Gunderson Dettmer
Media & Communications
11
GDELT tests Gemini as a fact-checker for state media narrativesThe GDELT Project used Google’s Gemini to analyze a day of coverage from Iran’s IRINN state broadcaster, separating accurate, misleading, and false claims — an early glimpse of AI-assisted propaganda monitoring at scale.
GDELT · Aug 13
12
An automated news site beat legacy media to a breaking storyThe Automated reports on an AI-run news site that published a breaking story before traditional outlets — a marker for newsrooms (and comms teams timing announcements) of how fast machine-driven publishing is getting.
The Automated · Aug 13
From Your Inbox
Substack Highlights.
The dedicated Substack tools were unavailable in this run, so today’s newsletter coverage comes from the Substack editions that flow into your Inoreader feeds. Summaries reflect each edition’s lead items.
Ben digs into xAI’s Grok Bot launch and argues the framing of “AI coworker with its own computer” hides what the product actually is — and where it falls short of the marketing.
Useful counterweight to the day’s breathless Grok coverage before you cite it in any internal discussion of agent products.
A hands-on guide to OpenCode, the open-source alternative to Claude Code that runs in your terminal and can be pointed at different AI models.
Relevant if you want a fallback coding agent that is not tied to a single AI provider — or want to understand what your technical colleagues are experimenting with.
Lead story: Claude Code’s Auto Mode going default — what changes for everyday users and why Anthropic believes automated safety checks beat human approval fatigue.
Also covers Anthropic’s invisible watermarking of Claude-generated text (“Claude now watermarks everything it writes”).
Reports that researchers cracked open the reasoning processes of OpenAI, Claude, and Gemini models — with Claude able to identify its own generated text.
Also covers Grok Bot and the LTX 2.5 open-source video model release.
From Your AI Feeds
Inoreader AI Folder.
12 articles landed in your “AI” folder in the last 24 hours — 9 unique stories after removing duplicates. OpenAI’s speed push dominated, with GDELT’s media-monitoring experiments the most interesting outlier for your line of work.
OpenAI’s companion playbook showing how startups use GPT-5.6 to build faster, more cost-efficient agents with smarter model selection and new Responses API capabilities.
Kalev Leetaru runs Gemini across a full day of Iran’s IRINN state broadcaster coverage to separate accurate, misleading, contested, and false claims. For a comms team monitoring global narratives about your company, this is a working preview of AI-scale media monitoring — the same techniques could track how any story travels through state and state-adjacent media.
Contractor Darana Hybrid claims it was stiffed for $137 million on the Colossus I and II data centres in Memphis and $431 million on facilities in nearby Southaven, filing liens against two SpaceX subsidiaries. Gerard’s skeptical take is a useful pressure test on the AI infrastructure boom narrative.
See also · News Roundup
Discoveries
Workflows & Tool Watch.
Check your Claude Code / Cowork permission settings today
With Auto Mode becoming the default today, Claude will act without asking for step-by-step approval unless it judges an action risky. For most daily work this removes real friction — but for anything touching sensitive files, email sending, or publishing (like this briefing’s own workflow), it is worth a two-minute review of your settings to decide where you still want a human in the loop. The Hacker News thread has practical takes from heavy users on when to trust it.
DEVONthink MCP Server — let AI agents work inside your database
A community-built MCP server for DEVONthink is being actively discussed on the DEVONtechnologies forum. It lets an AI assistant like Claude search your DEVONthink databases, read and file documents, and feed results from other tools (web research, transcripts) directly back into DEVONthink. For your setup — where DEVONthink is the archive of record — this is the bridge that turns it from a filing cabinet into something Claude can actively research against.
Obsidian + Claude: updated 2026 setup guides for vault-connected AI
Two freshly updated guides walk through connecting an Obsidian vault to Claude via MCP servers — covering search, note creation, frontmatter and tag management, and even a lighter-weight approach using a plain filesystem connection with no plugins at all. If you have been meaning to let Claude search and update your Obsidian notes rather than copy-pasting between apps, these are the current best starting points.
Bagel Bots: “The Prompt That Fixes Your Messy Workflows”
A practical newsletter edition built around a single reusable prompt: describe a repetitive task or bottleneck in your work, and the prompt walks the AI through mapping the workflow, spotting the busywork, and proposing an automated version. Simple, but a good template for auditing recurring comms-team processes — media list updates, coverage reports, weekly summaries — before deciding what to hand to Claude, Zapier, or Keyboard Maestro.
GDELT’s playbook: vibe-coding internal tools with Gemini
Beyond the fact-checking experiment, the GDELT blog this week showed how it used Gemini 3.6 Flash to “vibe code” a mobile-friendly thumbnail browser for its TV news archive — a worked example of a non-developer-friendly path from idea to working internal tool. A good pattern for small comms-team utilities (clip browsers, coverage dashboards) that would never justify an engineering request.
No major new Tencent AI stories in the past 24 hours
Today’s web sweep and your AI feeds surfaced no fresh Tencent-specific AI coverage. The active storylines reporters are most likely to reference are below.
The global rollout of Tencent Hy3 (formerly Hunyuan), released July 6 under an Apache 2.0 licence, remains the company’s flagship AI narrative: 295B total parameters (21B active), agent capabilities integrated into Yuanbao, and API input pricing cut to 1 yuan per million tokens. The free worldwide access window through WorkBuddy closes August 31 (Pacific Time) — expect usage-milestone and “what happens next” questions as that date approaches.
The global launch of the Hunyuan 3D engine — generating 3D assets from text, images, or sketches — continues to draw creator-tools coverage and positions Tencent in the AI-for-creativity conversation alongside its language models.
InfoWorld’s piece on a former OpenAI research scientist launching a new AI model for Tencent keeps circulating — a useful proof point in the “Western AI talent joining Chinese labs” narrative that reporters periodically revisit.