July 9 became the first day in AI history when three frontier labs put new flagship models in front of the public at the same time — and OpenAI led with the GPT-5.6 family. Sol is the flagship, built for complex reasoning and long-running agent work, and it set a new record on Terminal-Bench 2.1, a test of how well AI handles multi-step computer tasks. Terra offers roughly GPT-5.5-level performance at half the cost, while Luna is the fast, cheap option. Pricing lands at $5/$30 per million tokens for Sol, $2.50/$15 for Terra, and $1/$6 for Luna, and OpenAI says Sol ships with its most robust safety stack to date. Sol is also being served on Cerebras chips at up to 750 tokens per second for select customers — extremely fast for a frontier model.
Elon Musk’s SpaceXAI released Grok 4.5, describing it as “an Opus-class model, but faster, more token-efficient and lower cost.” It is built on a 1.5-trillion-parameter foundation and trained partly on data from Cursor, the AI code editor the company acquired earlier this year. On the benchmarks xAI chose to publish, it beats Anthropic’s Opus 4.8 on two (DeepSWE 1.0 and Terminal-Bench 2.1) and loses on two others. Pricing is aggressive at $2/$6 per million tokens, it is available immediately in Grok Build and across all Cursor plans, though notably not in the EU at launch. The company says it plans a new foundation model every month through the end of 2026 — a pace worth watching from a communications standpoint.
OpenAI replaced ChatGPT’s Advanced Voice Mode with GPT-Live and GPT-Live mini, new “full-duplex” voice models. In plain terms: instead of taking turns like a walkie-talkie, the AI listens and speaks at the same time, so conversations flow like a real phone call — including the ability to talk over you. There are default voice models for both paid and free users, plus what OpenAI calls stronger voice safety safeguards. For anyone thinking about media training, spokesperson prep, or podcast production, this is the moment AI voice interaction stopped feeling robotic.
Nvidia has lost roughly $1 trillion in market capitalization since its May 14 peak, with the stock down 16% and its valuation multiple back to early-2019 levels. Analysts describe it as a rotation rather than a collapse: investors are spreading their AI bets across other chipmakers, with Micron, Intel and AMD adding roughly $2 trillion in combined market value during the second quarter. The AI infrastructure story is not dying — it is broadening — but expect “AI bubble” headlines to keep circulating, which matters for anyone fielding questions about tech-sector spending.
Even as Nvidia slid, newsletters this morning framed the counter-story: Anthropic is reportedly finalizing a funding round of up to $50 billion at a valuation around $900 billion — which would put it above OpenAI’s $852 billion March valuation. The driver is revenue: thanks to products like Claude Code and Cowork, annualized revenue has grown roughly fivefold since late 2024, approaching $45 billion. Reports suggest this could be Anthropic’s last private round before an IPO.
Grok 4.5 lands inside Cursor on every plan, day onexAI’s first big move since acquiring Cursor: the new model ships across all Cursor tiers immediately, at 80 tokens per second and $2/$6 per million tokens.
Illinois signs a landmark AI safety law with first-in-nation annual auditsGov. Pritzker’s Artificial Intelligence Safety Measures Act requires frontier developers to publish catastrophic-risk frameworks and undergo mandatory annual third-party audits, effective January 2028. OpenAI and Anthropic both supported the bill.
Anthropic researchers map Claude’s hidden “thinking” layerNew interpretability research describes a silent internal workspace where Claude organizes reasoning before it answers — a step toward understanding how these systems actually work inside.
AI Valley · Jul 8
From Your Inbox
Substack Highlights.
The Substack tools were not available in this run, so today’s newsletter coverage comes from your Inoreader newsletter feeds (which mirror your Substack and Beehiiv subscriptions). Full-article summaries will resume when the Substack connection is restored.
GPT-Live is the headline: full-duplex voice means ChatGPT listens and speaks simultaneously, making conversations feel human — interruptions included.
The issue pairs two market signals: Nvidia’s $1 trillion market-cap slide and Anthropic’s reported $50 billion raise — the AI trade is rotating, not deflating.
A follow-up edition announced a live head-to-head test of GPT-5.6 Sol vs Claude Fable, a sign of how quickly the model wars have become spectator sport.
Beyond the Grok coverage, the issue highlights a claim circulating in Chinese state-linked media that Claude Code contains a “backdoor” that could exfiltrate user data.
Worth your attention professionally: this narrative sits at the intersection of US-China tech tension and AI security, and could surface in questions about Chinese companies’ AI tooling choices.
The newsletter treats the benchmark fight (Grok 4.5 vs Opus 4.8) as marketing-first — xAI published only four benchmarks and wins on just two.
When Claude Fable 5 went dark earlier this month, Liberty Mutual’s AI systems kept running because they were built for multi-model failover from day one.
The piece argues every serious AI operation needs a fallback plan: route work across at least two model providers so one outage never halts the business.
A useful playbook reference for any team (comms included) building daily workflows on top of a single AI vendor.
A plain-language explainer of Anthropic’s new interpretability research into Claude’s internal “scratchpad” — the hidden layer where the model organizes its thinking.
Also flags that Fable 5’s current availability window may be limited, advising readers to test it on real projects now.
See also · News Roundup
From Your AI Feeds
Inoreader AI Folder.
Roughly a dozen items landed in your “AI” folder in the last 24 hours; after removing duplicates of stories covered above, five stand out — led by a piece squarely aimed at your day job: how PR measurement changes when audiences ask AI instead of Google.
Signal AI argues that traditional PR metrics — impressions, share of voice, earned value — miss a fast-growing channel: what large language models say about your brand when someone asks before a pitch, purchase or partnership. The piece makes the case for tracking and influencing “AI answer visibility” the way comms teams once optimized for Google. Directly relevant to how Tencent’s story gets told inside ChatGPT, Claude and Perplexity.
Kalev Leetaru demonstrates using Gemini’s Deep Research to compare how different countries’ media cover the same stories — early evidence that AI research agents can do cross-national narrative analysis that once took analyst teams weeks. A capability worth borrowing for global media monitoring.
AI skeptic David Gerard’s 32-minute Times Radio interview covers AI and the job market, arguing that hasty AI-driven staff cuts are backfiring. Useful as a barometer of the critical narrative — the counterweight voice to this week’s model-launch euphoria.
Gerard’s second piece asks what bubble comes after AI, and finds venture capital already warming up quantum computing as the next lottery ticket. A sharp read on hype cycles for anyone who has to distinguish real technology stories from investor narratives.
A video walkthrough with designer Yannik Marek showing that serious visual AI work in ComfyUI is about workflow, control and taste — not one lucky prompt. Relevant framing for creative teams defending the craft behind AI-assisted content.
See also · Substack Highlights
Discoveries
Workflows & Tool Watch.
DEVONthink 4.3.2 now ships its own MCP Server
The updated Take Control of DEVONthink book confirms DEVONthink 4.3.2 includes a built-in MCP Server — meaning Claude (or any MCP-capable AI) can now search, read and file documents in your DEVONthink databases directly. The release also adds widgets for Mac, iPhone and iPad, and DEVONthink To Go 4.1.3 gains RSS feed support, so your mobile research library can now pull feeds too. If you have not connected Claude to DEVONthink yet, this is the native, supported way to do it.
Claude Cowork expands to mobile and web with remote sessions
As of July 7, Cowork is no longer desktop-only: sessions now run remotely in beta, saved to your Claude account, so work continues when your laptop is closed and scheduled tasks (like this briefing) run with no device online. You can start a task on the Mac mini, check it from your iPhone in a taxi, and pick it up on the MacBook Pro later. Rolling out to Max subscribers first.
n8n’s MCP server can now build and edit workflows for you
n8n updated its official MCP server so AI assistants can create and update automation workflows, not just run them — including validating workflows, running test executions and generating test data. In practice: you describe an automation (“when a media enquiry email arrives, log it to Notion and ping me in Slack”) and Claude assembles the n8n workflow itself. n8n also now works in both directions: it can use MCP tools inside its own AI agents, and expose any workflow as a tool for external agents.
A live bridge between Obsidian and n8n: mcp-obsidian-sse
A community MCP server connects Obsidian to a self-hosted n8n instance over a live connection (SSE), so edits in your vault can trigger automations, and external data can flow into notes in real time. Imagine a media mention landing in your monitoring feed and appearing as a pre-formatted note in your Obsidian vault seconds later — no manual capture.
Flow: all-day voice input for working with AI tools
Trending on Product Hunt this week, Flow is a voice-input app designed for dictating to AI assistants like Claude, ChatGPT and Cursor faster than you can type — product managers are using it to draft documents by talking. If you like MacWhisper for transcription, Flow targets the other direction: speaking as your primary input method for AI work.
The emerging two-tool pattern: Perplexity for discovery, Claude for execution
Comparison pieces this week converge on the same practical advice: use Perplexity Computer (19 models, 400+ app integrations) for research and validation, then hand off to Claude Cowork for synthesis and execution. For a comms team, that maps neatly to “monitor and gather” versus “draft and deliver.”
Bloomberg’s hands-on review of Xiaowei, the WeChat AI agent in testing since late June, is broadly positive: users can type or speak prompts, search, summarize chats and Moments, draft replies, and trigger mini programs to book flights, taxis or meals. The framing — an agent that could eventually run errands for 1.4 billion users across millions of mini programs — is the narrative international media are now attaching to Tencent’s AI strategy. Expect follow-up enquiries on rollout timing and privacy handling.
Tencent’s Hy3 — an open 295-billion-parameter mixture-of-experts model (21B active parameters, 256K context, Apache 2.0 license) released July 6 — continues to earn strong technical press as the latest Hunyuan flagship, with emphasis on agent capabilities and ecosystem integration. InfoWorld separately profiled the former OpenAI research scientist behind the model, a personnel angle that humanizes the story.
Bloomberg reports Tencent sold roughly $1.5 billion of Kuaishou stock, presenting it as paring mature internet bets to fund the AI push. The “divesting to double down on AI” storyline is now the default market interpretation of Tencent portfolio moves — useful context for investor-facing messaging.
Trade coverage of Tencent Cloud’s international announcements highlights Tencent WorkBuddy, an agentic AI workspace for complex office work, and Tencent Design Miora, an AI-native creative studio with persistent memory for design-team consistency. Positive B2B framing of the global expansion story.
A finance-press piece pairs Tencent giving AI agents their own email identities with Google building agent payment rails, arguing machine-to-machine communication is “one click away.” Speculative, but it shows Tencent being cited as a leading indicator in the agent-infrastructure narrative.