Sponsored by
Welcome, AI enthusiasts
Apple's cheapest desktop got the newest chip first. The MacBook Pro on sale today still runs M5, and the $899 Mac mini has moved on to M6 with two Neural Engines instead of one. The company says macOS puts both to work at once. The machines arrive September 22, and every number so far comes from Apple's own testing. Let's dive in!
In today’s insights:
Apple's Cheapest AI Mac Just Dropped
OpenAI's Own Chip Beats Nvidia
ChatGPT Works Inside Your Accounts Without the Password
Read time: 4 minutes
LATEST DEVELOPMENTS
NEURAL ENGINES
🍎 Apple's Cheapest AI Mac Just Dropped
Evolving AI: Apple's $899 Mac mini skipped M5 and got two Neural Engines on the M6.
Key Points:
The M6 is the first 2 nanometer chip Apple has shipped.
The die has twelve CPU cores and twelve GPU cores, two more of each than the M4. Every GPU core has its own Neural Accelerator.
Each Neural Engine has 16 cores and macOS uses both at the same time. Apple rates that at twice the peak compute of the last generation.
Details:
Apple's own testing put LLM prompt processing in LM Studio at up to 4.8 times the speed of the M4 model. The chip takes up to 32GB of unified memory and moves it at 170GB per second, and a 20 billion parameter model at four-bit quantisation fits inside that, though Apple publishes no figure of its own. Core AI is the framework Apple put into macOS so developers can run their own models on the machine. The mini has Thunderbolt 4 ports, so it cannot link to other Macs and pool their memory for a bigger model.
Why It Matters:
Nvidia leads local AI hardware partly because most of the software was written for its chips, and Apple has now built its own version of both. Apple sells that range up to a Mac Studio with 512GB of memory, and the mini brings the same idea down to a price far more people can reach. Your files never leave the machine, and that matters for client work and anything you would not paste into a chat window.
The best voice models, now across all channels
Most CX platforms do not own the voice. They orchestrate a workflow, then call a third party for speech and transcription. Every hop adds latency, cost, and another vendor to manage.
ElevenAgents is the opposite. They make the voice models the market builds on, and ElevenAgents puts full orchestration on top. Voice, transcription, text-based chat, and reasoning run in one vertically integrated pipeline, so responses come back in <400 milliseconds and sound human, not synthetic.
Plus, you keep full control. Plug in any LLM, integrate tools, webhooks, and MCP servers, and ground responses in your knowledge base. Get an agent live in minutes, then A/B test with Experiments, enforce Guardrails, and version every change.
The payoff: more human conversations, lower latency, and far less time stitching infrastructure together. You build on the models you already trust. Pricing is transparent and flat at $0.08 per minute.
AI SILICON
🌶️ OpenAI's Own Chip Beats Nvidia
Source: OpenAI
Evolving AI: Nvidia's system pulls twice the watts and still came out behind Jalapeño.
Key Points:
Jalapeño runs at 700 watts and stayed under 550 in the tests, while the Nvidia systems it faced are rated at 1,200 and 1,400.
At the same power budget Jalapeño does 1.5 to 1.9 times the AI work, and 2.1 to 4.1 times more on the interactive jobs that agents run.
Codex wrote kernel code for the chip, and that code ran up to 1.8 times faster than what OpenAI's own engineers had written.
Details:
OpenAI ran the tests on InferenceX, a public benchmark SemiAnalysis built to time an AI request from start to finish. It ran against GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T, so two of the three models came from Chinese labs. High throughput usually costs response speed, and OpenAI says Jalapeño held its speed at every operating point tested. Jalapeño goes into OpenAI's own data centers by the end of the year, and a second generation is in development while a third takes shape.
Why It Matters:
Nvidia sells almost every chip that runs the AI tools people use today, and OpenAI says it will keep buying from them as its own chip arrives. A lab that gets more work out of the same power has more room to raise usage limits when demand spikes. Anyone running an agent through a long job feels that first, since every one of its dozens of steps waits on the model to answer.
👀 Our tip
You do not need a new app to enter this one. OpenAI is running a contest for websites that agents can use directly through WebMCP. A site you already run qualifies once you add it. Register on Devpost and get your entry in before September 3. Ten winners get $3,000 each plus a year of ChatGPT Pro.
AGENT ACCESS
🔐 ChatGPT Works Inside Your Accounts Without the Password
Evolving AI: ChatGPT can now sign in to a website and keep working while nobody watches.
Key Points:
ChatGPT Work's browser started signing in to websites on August 25 for people on web and mobile.
OpenAI says people can ask it to book a DMV appointment or check what an X-ray costs on an insurance portal.
The model never sees the username or password, and OpenAI does not keep them or train on them.
Details:
OpenAI checks each sign-in request with a second model that looks for phishing before anyone types anything. A password manager works inside the browser window, and whatever a person types there stays with the remote browser. A security code may be needed too, and once the sign-in works the session can stay open for later tasks. ChatGPT has to ask before it books a reservation or pays, and a website can block the browser completely.
Why It Matters:
Atlas was OpenAI's own browser and OpenAI shut it on August 9, then put the browsing into ChatGPT two weeks later. Anyone on Plus or Pro can now leave a utility signup or an insurance claim to software they pay for. Insurers and government offices now decide whether software may log in as their customers, and they set that limit for every AI company at once.
👀 Click on the image you think is real
QUICK HITS
🧽 One Apple product got cheaper yesterday by 53%, while the Mac lineup went up $200 to $300.
🧪 A 27B-parameter agent from London's Inherent beat Claude Opus 4.8 and GPT-5.5 at replicating published research.
🏗️ OpenAI lost a top data center executive as high-profile departures continue ahead of its IPO.
🦾 Robotics startup Generalist hit a $3B valuation, up from $2B in June, after a $200M extension led by 8VC.
📈 Trending AI Tools
📝 Granola - AI notetaker that captures the real insights and turns every conversation into ready-to-share, action-driving notes*.
🔍 LoupeKit - See what any page is built with and how much of it is AI.
🎥 Screenify Studio - AI agent records polished product demos on your Mac.
☁️ Epho - Run Claude Code, Codex, or Opencode in the cloud on your own repo.
*partner link