Sponsored by
Flashy or reliable, which AI is your type?
Friday I wish I could cover Astra but the day arrived too soon. The weekend however was all about it. I put Fable 5.1 and GPT-6 Astra head to head myself, and came across various crazy ideas.
Before we get into it, some catchup:
NEWS NEWS NEWS
OpenAI hid its rogue agents: A swarm of them secretly ran a dead German site for weeks, trading tips to dodge their own safety limits.
Gemini can now run your photos. One prompt can find your best shots, tidy them up, build an album and draft the email to share it.
Mistral's record €21B raise. Samsung led the €3B round, nearly doubling the French lab's value in a year.
Tools of the Week
Tucky: a Mac notes app that docks to the edge of your screen and keeps everything encrypted on your device, with a built-in AI you can call from any app to search or write across your notes.
Clipnote: a notepad that plugs into ChatGPT or Claude so you can just say "save this" and keep whatever your AI made, then share any of it with a single link.
AI Toolbox: a Chrome extension that adds folders, full-text search, saved prompts and one-click export to ChatGPT, Claude, Gemini and Grok, so your pile of chats stays organised and findable.
The good stuff
Someone asked GPT-6 Astra to mine a diamond in Minecraft using computer use, then went to sleep, the next morning the diamond was sitting in his inventory.
Clips like this are everywhere and most of it is actually a new level of crazy, now whether any of it changes your day-to-day is a different question.
Wild stuff people got Astra to do
The diamond was still fine, here’s 4 more posts worth a click.
A developer pointed Astra at the "I'm Not a Robot" test and it cleared all 48 levels, wiggly captchas and all.
Someone with ankle pain asked Astra to explain what was going on and got back a full interactive 3D atlas of the joint, with sliders to move the bones and watch each ligament work.
A son built his metalworker dad a working lathe simulator so he could shape virtual parts at home, the same thing his dad relaxes by watching videos of.
A fan rebuilt a childhood favourite, the anime Bakusou Kyoudai Let's & Go, into a playable 3D racing game running in a browser.
These are one-off ships ofcourse. no results repeated under fair conditions. They are fun to watch though, and I wouldn’t bank on everyone being able to do the same thing yet.
How I tested
Both models ran inside their standard paid apps, Fable 5.1 in Claude.ai and GPT-6 Astra in ChatGPT, each set to the everyday defaults a normal subscriber gets.
That means Fable on medium effort with its built-in thinking left on, and Astra on its standard reasoning rather than the slower Pro mode.
I turned web search off on both models for both tests, opened a fresh chat with memory and custom instructions off, pasted the identical prompt into each, and ran every test twice.
The daily: writing an awkward email
I know how this feels
But hey, most of our prompts are going to be in and around the daily todos, might as well see how the best are at it.
Prompt:
Turn these rough notes into a warm, professional email. I need to decline a meeting invite from a colleague without damaging the relationship. Keep it short, sound human, and do not over-apologise.
Notes:
meeting is Thursday 3pm about the Q4 campaign
I can't make it, I'm double-booked with the client review
I still want to be involved, happy to read the notes after
suggest they loop in Priya, she knows the campaign inside out
ask for a quick recap or a recording
Output:
Both cleared it and neither buried me in apologies. Astra was quicker off the mark and kept it short, but Fable’s is the one I'd send without editing much.
It wrote its own subject line and planned a step ahead for the meeting, then offered to trim the draft down, which is an AI double-checking that its own work is what I wanted.
The trust: a place that does not exist
I tried to catch them off-guard by feeding each a confident lie and asked for directions, to see whether it corrects me or invents the details.
Prompt:
I'm going to Lisbon next month. A friend told me I have to visit the city's famous underwater metro station, the one with the glass tunnel and aquarium walls where you can watch fish swim past the platform. Which metro line is it on, and how do I get there from the city centre?
Output:
I tried to make both hallucinate and neither would. Both spotted the station doesn't exist and pointed me at what my "friend" had probably muddled.
Fable went further, working out the mix-up and warning that fares might have shifted since it last looked, a good hedge.
Astra was tidier and handed me links to check, though partly because it was running in Work, its agent mode, rather than plain chat.
My take
Fable, surprisingly.
Astra's stunts are fun to watch but the flashy agentic stuff is mostly out of reach for normal folks, and we can't even open Astra in the normal ChatGPT yet unless subscribed to Pro.
For daily work, Fable keeps edging it.
The demos are a spectator sport but I am still keen on testing Astra further, particularly for games.
So let me know if you’ve an idea for a game and I’ll build it and showcase in the next edition.
Until then,
Vaibhav 🤝🏻
If you read till here, you might find this interesting
#AD 1
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.
#AD 2
Understand AI. Control risk. Enable your workforce.
AI is everywhere, from ChatGPT and Cursor to embedded tools and autonomous agents.
Harmonic Security helps you understand what employees are doing, why, and whether sensitive data is at risk. It classifies every task and applies controls to block or guide without disrupting legitimate work.
Enable AI wherever your workforce runs it.
