OpenAI unveiled a new API service tier called Ultrafast that runs its GPT-5.6 Sol model up to 14 times faster than standard, delivering up to 750 output tokens per second. The speed comes from specialized chips made by Cerebras, a chipmaker that competes with Nvidia. In plain terms: AI responses that used to take half a minute now arrive almost instantly, which matters most for real-time uses like live customer service agents, voice assistants, and interactive tools. The move signals the next competitive front in AI is not just how smart models are, but how fast they respond.
Google cut the price of Gemini 3.7 Flash roughly in half, and it is not alone: OpenAI has slashed pricing on GPT-5.6 Luna, and Anthropic is positioning Claude Opus 5 at about half the price of its flagship Fable 5. The driver is competitive pressure from lower-cost Chinese models, including DeepSeek and Moonshot AI, which are winning users on price. Bloomberg reports that Beijing-based Z.ai is upgrading its GLM-5.3 model specifically to close the coding gap with Anthropic and OpenAI. For communicators, the story is a shift in the industry narrative: frontier AI is rapidly becoming a commodity, and the US-China rivalry is now being fought on price as much as capability.
Anthropic introduced invisible watermarking for Claude-generated text, a hidden signature woven into the output that can later verify whether a passage was written by the AI. The move is widely read as compliance with the EU AI Act’s transparency requirements, which came into force this month. Newsletter writers are calling it “secret tattoos” for AI text. For a communications team, this is worth watching closely: watermarked AI text means detectable AI text, which has real implications for how organizations draft press materials, executive communications, and social content with AI assistance.
AI regulation moved from theory to enforcement this month. The EU AI Act’s core obligations for high-risk systems and deployer transparency took legal effect on August 2, ending the grace period. In China, regulators issued the first fines under the new companion-AI rules: 12 companies were penalized a combined 4.2 million RMB in the first week of enforcement. Meanwhile in Washington, the “Great American AI Act” has stalled in the House over its state preemption clause, leaving California and Colorado to enforce their own rules. The upshot: three major markets, three very different regulatory realities, and multinational companies must now navigate all of them at once.
Tencent announced broader international access to Hy3 (formerly Hunyuan), following its official release on July 6. Users and enterprises worldwide can access Hy3 free of charge through WorkBuddy until August 31, as well as through Tencent Design Miora and Tencent Cloud TokenHub. Call volume in Hy3’s first week grew more than 68 times over the previous generation, and the larger-parameter Hy4 is planned for release soon. Full details in the Company Watch section below.
OpenAI introduces Computer History for ChatGPTA new feature that lets ChatGPT remember and reference what you have done on your computer, extending its agentic capabilities across sessions.
AI Valley · Aug 14
04
OpenAI splits its frontier AI into ‘Doug’ and ‘Astra’Reports say OpenAI’s next generation is being developed as two tracks — internal codenames Doug and Astra — with the company sharing preliminary cybersecurity evaluations for Astra amid concerns it may be too capable to ship as-is.
OpenAI appoints Dali Rajic as Chief Revenue OfficerThe new CRO will lead OpenAI’s global revenue organization — the latest in a string of enterprise-focused executive hires, following reports of the COO’s departure.
Cursor is now part of SpaceXAI Valley reports the popular AI coding tool has been folded into the SpaceX orbit, with newsletters speculating the Cursor brand may become Grok-branded.
Amazon trains AI on Twitch streams — opt-out onlyTwitch added a setting to opt out of having channel content train Amazon’s generative AI models; its product chief admitted “if it was opt in, nobody would opt in.”
Pivot to AI · Aug 13
13
People are grieving their AIThe Signal examines the growing phenomenon of users mourning discontinued AI models and companions — an emerging reputational issue for every AI company that retires a model.
EU AI Act core obligations now legally enforcedAs of August 2, high-risk system rules and deployer transparency duties carry legal force — the “theoretical era” of EU AI compliance is over.
Cubbbix · Aug 2026
16
China issues first fines under new companion-AI rulesTwelve companies received fines totaling 4.2 million RMB in the first week of enforcement — a signal Beijing intends to police emotional-AI products aggressively.
AIFOD · Aug 2026
17
US federal AI bill stalls over state preemptionThe “Great American AI Act” is stuck in the House over whether federal rules should override state laws; California and Colorado continue enforcing their own.
The Substack connector was not available in this session, so today’s newsletter coverage was pulled from your Inoreader newsletter feeds, which capture the same Substack (and Beehiiv) editions.
Explores the wave of genuine grief users experience when AI models and companions are retired or changed — treating it as a real emotional and cultural phenomenon, not a curiosity.
Raises a comms-relevant point: model deprecations are becoming reputation events, with organized user communities pushing back publicly.
Earlier this week the same outlet covered Google’s infrastructure strategy and ByteDance’s bigger AI picture — worth a scan for the China angle.
Grok 4.6 undercuts Claude on price by a claimed 85 percent while matching it on several benchmarks.
Also covers the departure of OpenAI’s COO, Meta doubling down on “open superintelligence,” and Claude making notable progress on a Riemann-related mathematics problem.
Offers a mental model for delegating work to AI: most people automate one small step and leave the surrounding workflow manual, capturing only a fraction of the benefit.
Argues you should map the whole job end to end, then hand entire segments to AI — directly applicable to comms workflows like coverage monitoring and reporting.
Google, OpenAI, and DeepSeek all dropped models or price cuts on the same day; the Neuron reads it as a coordinated race to own the “workhorse” tier.
Claims Anthropic’s Fable 5 has “flopped with businesses despite the hype” on pricing grounds — a contrarian take worth noting alongside Anthropic’s half-price Opus 5 positioning.
Documents AI’s local trust problem: over 500 US communities have moved to block data center construction.
Also notes Grok Bot’s cross-software capabilities and argues the next important AI hire at most companies may not be technical.
From Your AI Feeds
Inoreader AI Folder.
12 articles landed in your “AI” folder in the last 24 hours. The day was dominated by the speed-and-price war (OpenAI’s Ultrafast, Gemini’s price cut, Grok 4.6) plus a strong undercurrent of AI-and-society stories.
Turkey’s Ministry of Justice announced the “CBS Organization Prediction Project,” which lets prosecutors and judges feed a case file to an AI system that predicts which terrorist organisation the defendant belongs to. Gerard’s characteristically sharp piece walks through why using a statistical classifier for criminal attribution is alarming — and it lands the same month the EU begins enforcing rules on exactly this category of high-risk AI.
OpenAI’s own announcement of the Cerebras-powered Ultrafast service tier, delivering up to 750 output tokens per second for real-time agent and voice applications.
Coverage of xAI’s Grok Bot, which provisions each AI agent with its own cloud computer so it can run apps and complete multi-step work like a human colleague.
A bipartisan, grassroots backlash against AI infrastructure is now measurable at scale — an issue with direct implications for any company announcing new compute buildouts.
Users form real attachments to specific AI models — and retiring them now triggers organized grief and backlash. A preview of a new class of reputational risk for AI product owners.
A practical mental model for delegating whole workflows to AI rather than isolated steps.
See also · Workflows & Tool Watch
Discoveries
Workflows & Tool Watch.
Turn your Obsidian vault into an active AI workspace with MCP
A comprehensive new guide walks through connecting Obsidian to Claude via MCP (the connector standard that lets Claude read and write your notes directly). The result: your vault stops being passive storage and becomes something Claude can search, summarize, and organize on your behalf — ask it to synthesize every note you have on a topic, or file meeting notes automatically. With Obsidian now past 1.5 million users, the plugin ecosystem around this is maturing fast. You already run Obsidian MCP servers for both vaults, so the guide’s workflow patterns (vault-wide synthesis, automatic filing, daily-note automation) are immediately usable.
Today’s WhatPlugin.ai issue makes a point worth internalizing: most people automate a single step (say, transcription with MacWhisper) and leave everything around it manual — capturing maybe 10 percent of the possible time savings. The suggested exercise: write out one recurring job end to end (for example, podcast episode: record, transcribe, edit, show notes, publish, promote), then identify which whole segments — not steps — could be handed to AI. A good candidate for the Red Book pipeline or your daily media monitoring.
An active thread on the DEVONtechnologies community forum collects real-world setups for DEVONthink’s built-in AI features — connecting ChatGPT, Claude, or local models via Ollama to summarize documents, auto-tag, and run natural-language searches across databases. Useful peer examples if you want to push your DEVONthink research archive further, especially the workflows combining smart rules with AI actions.
Claude Code + Figma MCP: design concepts with full rationale
A detailed writeup shows how connecting Claude Code to Figma through MCP can generate multiple design concepts complete with every state and written design rationale. For a comms team producing social and video assets at volume, this pattern — brief in, several fully-documented design directions out — is worth showing your creative leads.
Framer’s August 11 launch lets you drop AI agents directly onto the canvas to design, write, analyze, and organize websites collaboratively. Also charting on Product Hunt this week: AMP by CanyonTechs (task-executing AI agents) and Murmel, a shared cloud canvas where human teams and AI agents work side by side — a glimpse of what “AI coworker” interfaces are converging toward.
Gemini as a fact-checker for state media narratives
The GDELT Project published a first look at using Google’s Gemini to fact-check global media narratives and state propaganda, testing it against a day of Iranian state TV (IRINN) coverage. For anyone doing media monitoring across state-adjacent outlets, the methodology — AI-assisted claim extraction and verification at broadcast scale — is directly relevant to issues-monitoring work.
The company’s flagship announcement continues to draw coverage: Hy3 is accessible worldwide free of charge through WorkBuddy until August 31 (Pacific Time), plus Tencent Design Miora and Tencent Cloud TokenHub. Call volume grew more than 68 times over the previous generation in the first week, and the larger-parameter Hy4 is planned for release soon. The PR Newswire syndication is also circulating widely.
Earnings coverage highlights revenue of RMB 204.8 billion (up 11 percent) and frames AI spending as beginning to pay off, with WorkBuddy and CodeBuddy user growth called out and WeChat’s AI assistant “Xiaowei” expanding gray-scale testing. IG’s earnings preview framed the quarter around AI capex against a gaming slowdown — expect analysts and reporters to keep probing that tension.
InfoWorld covers a new Tencent model led by a former OpenAI research scientist — a talent-flow story likely to resurface in US-China AI narratives. Worth having a line ready on Tencent’s research leadership and international talent.
Bloomberg’s Z.ai story and the broader coverage of DeepSeek and Moonshot AI pressuring OpenAI and Anthropic on price puts Chinese AI competitiveness front and center. If asked, the relevant Tencent line is the Hy3 global rollout and its free-access window — Tencent’s own move in the same accessibility race.