The episode kicks off with Luca delivering a half-ironic cold-open manifesto: if you're not an entrepreneur, you're just a normal person — 'go and stay in your 9-to-5.' It's a bait designed to grab attention, and it works. Then the hosts pivot to a story that actually earns some genuine tension: Rik went tandem paragliding with a stranger from a WhatsApp group, climbing 1,000 meters of altitude before discovering his pilot had only flown from that exact spot once before. Luca and Sasha tried to talk him out of it; Rik held a goodbye speech the night before. He was terrified. He jumped anyway. It was amazing.

After the cold open, Rik kicks into proper episode mode with a summary of what's coming: Grok 4.5, Claude Fable extended, GPT-5.6 Soul, a Europe surveillance corner, and quickfire rounds. Ben is still in Brooklyn post-July 4th heatwave, rooting for Argentina and taking diplomatic fire from Belgian Rik about Belgium knocking the US out 4-1. Luca is bouncing between Glasgow, Malta, and Albania — where he's starting a tourism business — and nominally supports Brazil, who are already out. The banter is warm and chaotic, exactly the vibe of three builders scattered across the globe who somehow run a top-1% podcast.

After the cold open, Rik kicks into proper episode mode with a summary of what's coming: Grok 4.5, Claude Fable extended, GPT-5.6 Soul, a Europe surveillance corner, and quickfire rounds. Ben is still in Brooklyn post-July 4th heatwave, rooting for Argentina and taking diplomatic fire from Belgian Rik about Belgium knocking the US out 4-1. Luca is bouncing between Glasgow, Malta, and Albania — where he's starting a tourism business — and nominally supports Brazil, who are already out. The banter is warm and chaotic, exactly the vibe of three builders scattered across the globe who somehow run a top-1% podcast.

The week's biggest headline: SpaceX AI announced Grok 4.5 with 33 million views on the announcement tweet. The model was co-trained with Cursor and, according to BridgeMind and CursorBench data, places third overall — behind Fable 5 Max and Fable 5 Extra High — but above Fable 5 High, and at roughly 70% lower cost per task. ^{[1] — Rik "Grok 4.5 vs Opus 4.8: ~70% cheaper: Grok 4.5 performs on par with Claude Opus 4.8 on key benchmarks but costs approximately 70% less per ta…" 09:03 Play} Ben notes the leapfrog is clearly data-driven: once SpaceX plugged into Cursor's coding dataset, Grok jumped fast. He's skeptical that Grok can beat Anthropic and OpenAI purely on model quality and suggests they need a different vertical, perhaps raw compute speed or parallel agent execution. The real opportunity, Rik argues, is that Grok is approaching open-source cost territory — which could make it the default grunt-work model for agentic pipelines.

The week's biggest headline: SpaceX AI announced Grok 4.5 with 33 million views on the announcement tweet. The model was co-trained with Cursor and, according to BridgeMind and CursorBench data, places third overall — behind Fable 5 Max and Fable 5 Extra High — but above Fable 5 High, and at roughly 70% lower cost per task. ^{[1] — Rik "Grok 4.5 vs Opus 4.8: ~70% cheaper: Grok 4.5 performs on par with Claude Opus 4.8 on key benchmarks but costs approximately 70% less per ta…" 09:03 Play} Ben notes the leapfrog is clearly data-driven: once SpaceX plugged into Cursor's coding dataset, Grok jumped fast. He's skeptical that Grok can beat Anthropic and OpenAI purely on model quality and suggests they need a different vertical, perhaps raw compute speed or parallel agent execution. The real opportunity, Rik argues, is that Grok is approaching open-source cost territory — which could make it the default grunt-work model for agentic pipelines.

As benchmarks converge, Luca makes a case that resonates across the table: we've entered a qualitative era. He uses Claude for complex legal research — specifically, finding cross-jurisdictional loopholes he likens to hunting cybersecurity bugs — because Fable's depth is unmatched for that kind of pioneering work. But for quick website builds and SEO ranking tools, he reaches for Codex because Claude is simply too slow. Ben adds that Grok was missing from agentic coding conversations until this week, and that the Grok Build subscribers who bought at 60% off a few weeks ago made a surprisingly good deal. Rik notes that Grok 4.5's jump on Artificial Analysis benchmarks — from around 1,000 to 1,500 points, and 39% to 81% on TerminalBench ^{[1] — Rik "Grok benchmark: 1,000 → 1,500 points: On Artificial Analysis benchmarks, Grok jumped from approximately 1,000 points (Grok 4.3) to 1,500 po…" 12:58 Play} — is a serious leap, even if it's still nitpicked.

The conversation turns personal and practical: Rik walks through the exact pipeline he built to automate God Mode Podcast clip creation. Claude reads the full transcript and writes an instruction document for an agent. That document goes to Codex, which then literally navigates Riverside — clicking, creating clips, waiting — without Rik touching it. All he does is review and approve the final output. What used to take the most time in the whole production workflow is now fully delegated. ^{[1] — Rik "Hermes now runs agents entirely in the cloud — pick a model, choose server size, and your agent is live in 60 seconds. Teams can spin up ag…" 43:40 Play} He heard a similar story on Lenny's podcast, where the show's video editor started using Codex to find word gaps in the audio — a use case nobody on the Codex team anticipated. The lesson: the best agentic workflows are discovered by doing, not planned.

The timing feels deliberate: just as Anthropic was closing off Fable access, OpenAI dropped GPT-5.6 Soul — along with Terra and Luna — to steal the spotlight. Ben describes it as the first model that seems to have gone through some form of US government sovereignty review before release, a new precedent that nobody on the pod quite knows what to make of. ^{[1] — Ben Broch "GPT-5.6 Soul is the first major AI model launch that appears to have gone through US government review before public release. Whether that'…" 18:00 Play} A Theo (prominent developer) quote circulates: 'Damn good model. Not quite as smart as Fable, but fixed all the problems with 5.5. Incredibly determined.' That word — determined — emerges as the defining quality of this new frontier model generation: they don't stop at a passable result, they iterate until the task is genuinely done.

The three-model architecture emerges organically from the conversation and it's one of the episode's most actionable insights: use Claude Fable as the CEO making the strategic call, have GPT-5 manage and check the work, and deploy Grok 4.5 for all the high-volume, low-stakes token burning — because at 70% lower cost, Grok is purpose-built to be the workhorse. ^{[1] — Rik "Use Claude Fable for strategic planning, GPT-5 as a middle manager checking the work, and Grok 4.5 for cheap token-burning grunt tasks. Thi…" 20:03 Play} Ben frames it with a vivid analogy: Claude is the CEO telling OpenAI to go build it. Then he immediately raises the tension: will OpenAI eventually refuse to be delegated to by a rival's model? Or will they pragmatically embrace any usage of their API? The hosts don't resolve it, but the question hangs — because as these models become infrastructure, the politics between labs will matter as much as the benchmarks.

Just as builders were Fable-maxing through sleepless nights before the July 5th cutoff, Anthropic quietly extended access to July 12. Ben's read: this wasn't generosity — enterprise onboarding takes time, contracts need negotiating, and Anthropic hadn't deployed Fable to enterprise at sufficient scale yet. By giving individuals access longer, they maintain revenue while enterprises catch up. But Dario Amodei has been clear that enterprise is the priority. ^{[1] — Ben Broch "Free Fable access through July 12 is bait. Ben predicts Anthropic will lock Fable behind a $500–$1,000/month subscription tier, with the $1…" 24:30 Play} Ben's prediction lands sharp: for $100 or $200/month, don't expect Fable long-term. There's a $500–$1,000/month tier coming, and the extended free window is the hook. The hosts describe the current moment as deliberate FOMO engineering — Anthropic giving you just enough Fable to make you desperately want more.

Rik surfaces a debate circulating in VC podcast circles: Sam Altman's suggestion that the US government receive a 5% stake in OpenAI — at roughly $50 billion implied value — against OpenAI's own claims of trillion-dollar economic impact. Critics argue that if OpenAI really will reshape a $30 trillion economy, the government will eventually want a far larger share. Whether Sam is floating a trial balloon to gauge public reaction, or genuinely believes 5% is a fair exchange, remains unclear. What is clear, Rik concludes, is that both Altman and Anthropic's Dario Amodei have a spokesperson problem: neither is winning hearts or minds in this moment.

Rik plays the ChatGPT Live announcement video — complete with a grandmother knitting a sweater — and Ben takes it for a live spin. The technical leap is real: GPT Live One is the first ChatGPT voice model to be fully full-duplex, meaning it listens and speaks simultaneously rather than taking turns. Ben clocks the latency at under one second for almost every question. ^{[1] — Ben Broch "She tells me she talks to Timmy all the time, all the time. She gets advice about relationships, what she should cook for dinner. Her husba…" 32:47 Play} When you ask it something requiring web search, latency spikes — but for anything in-training, it's near-instantaneous. The demo is charming and slightly chaotic — the model has never heard of any of the hosts — but the anecdote that lands hardest is Ben's: his aunt named her ChatGPT 'Timmy,' talks to it constantly for relationship advice and dinner suggestions, and her husband reports she talks to Timmy more than she talks to him. The hosts call it 'dangerous territory' and move on.

Rik plays the ChatGPT Live announcement video — complete with a grandmother knitting a sweater — and Ben takes it for a live spin. The technical leap is real: GPT Live One is the first ChatGPT voice model to be fully full-duplex, meaning it listens and speaks simultaneously rather than taking turns. Ben clocks the latency at under one second for almost every question. ^{[1] — Ben Broch "She tells me she talks to Timmy all the time, all the time. She gets advice about relationships, what she should cook for dinner. Her husba…" 32:47 Play} When you ask it something requiring web search, latency spikes — but for anything in-training, it's near-instantaneous. The demo is charming and slightly chaotic — the model has never heard of any of the hosts — but the anecdote that lands hardest is Ben's: his aunt named her ChatGPT 'Timmy,' talks to it constantly for relationship advice and dinner suggestions, and her husband reports she talks to Timmy more than she talks to him. The hosts call it 'dangerous territory' and move on.

The timing is suspicious and the hosts say so: OpenAI retracted SWE-Bench Pro right before shipping GPT-5.6 Soul, the model that would otherwise have been evaluated on it. Their audit found 30% of benchmark tasks are simply broken. ^{[1] — Rik "30% of SWE-Bench Pro tasks broken: OpenAI audited SWE-Bench Pro and found 30% of its tasks are broken, leading them to retract their recomm…" 35:00 Play} Rik notes they've been skeptical of benchmarks for a while — that's partly why they use BridgeMind, which maintains its own independent evaluation. Ben raises a deeper problem: if the AI models doing the auditing are the same models being benchmarked, the process is fundamentally compromised. Who watches the watchmen? The answer, increasingly, is other AI — and that's a problem nobody has cleanly solved yet.

Levels.io going native iOS generated a tweet and a lot of speculation. The headline is clear enough — AI has made iOS development approachable for solo builders. But Luca's read is more interesting: Nomad Table, an iOS-only social map app for digital nomads, has quietly hit roughly $1M ARR and 70,000 users ^{[1] — Luca Arrigo "Nomad Table: 70,000 users in first year: Nomad Table, a social map app for digital nomads, crossed 70,000 users within its first year and r…" 42:10 Play} while Nomads.com has cut its price to $1 to defend market share. Nomad Table does something Nomads.com never managed: it actually connects people in person, through a map of local events. Ben and Rik excavate Levels' revenue dashboard and find it full of interesting anomalies — the $803K figure turns out to be from a Cursor investment being priced, not operational MRR. The meta-lesson: ideas from Levels spark industries, and then focused single-product founders build the better version on top of them.

Levels.io going native iOS generated a tweet and a lot of speculation. The headline is clear enough — AI has made iOS development approachable for solo builders. But Luca's read is more interesting: Nomad Table, an iOS-only social map app for digital nomads, has quietly hit roughly $1M ARR and 70,000 users ^{[1] — Luca Arrigo "Nomad Table: 70,000 users in first year: Nomad Table, a social map app for digital nomads, crossed 70,000 users within its first year and r…" 42:10 Play} while Nomads.com has cut its price to $1 to defend market share. Nomad Table does something Nomads.com never managed: it actually connects people in person, through a map of local events. Ben and Rik excavate Levels' revenue dashboard and find it full of interesting anomalies — the $803K figure turns out to be from a Cursor investment being priced, not operational MRR. The meta-lesson: ideas from Levels spark industries, and then focused single-product founders build the better version on top of them.

A week after the hosts predicted local agents would fade within six months, Hermes validates the thesis: pick a model, choose a server size, two clicks, and your agent is live in 60 seconds with no local file management, no GitHub repo wrangling. Teams can spin up parallel agents with granular access controls and shared billing from one portal. Rik connects this to what he heard at a recent panel: Lovable and Replit's biggest advantage over local coding tools is that they're collaboration platforms — whole teams can access the same build simultaneously. Hermes brings that same benefit to the agent layer. Luca's reaction: you can be on a walk, get an idea, deploy it in a minute. That's the future.

Cloudflare is building the infrastructure for the agentic web economy: a monetization gateway that uses the x402 protocol (originally championed by Coinbase) to charge AI agents stablecoins for accessing web content. When an agent visits your page, it pays you — automatically, instantly, in crypto. ^{[1] — Rik "Cloudflare's x402 gateway lets any website charge AI agents stablecoins for access to their content. Luca's insight: you're already paying …" 45:13 Play} Luca's framing cuts through the abstraction: right now, builders pay Firecrawl to get past Cloudflare so their agents can scrape websites. That money goes to an intermediary. With x402, it would go to the person who actually wrote the content. Ben raises the crypto trade-offs — instant settlement means no chargebacks if you're defrauded — but acknowledges digital-native use cases like domain sales are a perfect fit. Rik sees a broader implication: large labs training their LLMs on your data would have to pay for access. The web as a metered data market.

While the US races toward fully autonomous vehicles, the EU has mandated that all new cars include infrared cameras watching the driver's face for signs of distraction. The hosts find the irony stark: we're building technology to eliminate the need for human attention on the road, while regulators are simultaneously demanding more of it. Rik imagines dystopian scenarios — a car that stops mid-highway because the driver glanced away, or face data flagged to police. Ben notes Tesla has had this feature for years, originally as a check on self-driving activation: you had to shake the steering wheel, then it moved to eye tracking. But that's surveillance in service of enabling autonomy, not restricting freedom. There's a viral story about someone gaming the system by positioning a screen to look like they were watching the road while playing video games behind the Tesla display. The hosts see the EU mandate as heading the wrong direction.

The EU's chat control bill — requiring platforms to scan private communications for illegal content — passed not by winning a majority but by opponents failing to muster the supermajority needed to kill it. Most MEPs voted against it; it passed anyway. ^{[1] — Rik "EU chat control: majority not reached: Despite most MEPs voting to reject EU chat control legislation, the absolute majority of 361 members…" 53:46 Play} Rik calls it a sad day for Europe. Luca delivers the episode's most extended take: this vote being brought back for a second attempt in such a short window is abnormal — it takes massive lobbying infrastructure to get anything to a Parliament vote once, let alone twice in quick succession. He praises independent MEP Fidias as a rare voice who evaluates each proposal on its own merits without party allegiance. He floats his own political ambitions before walking them back. The EU Inc. project — Levels' proposal for an EU-wide startup entity — gets cited as another example of reform that looks good on paper but has stalled for two years without implementation. The tone is resigned but not hopeless.

The episode's final segment delivers what Rik calls the most practically useful news of the week: Anthropic launched a monthly Claude usage recap showing when you engage most and what you work on, plus quiet-hour settings and break nudges. The /checkup command, surfaced by Cloud Code Guy Boris, cleans up unused MCPs and skills — the digital equivalent of clearing app clutter. ^{[1] — Rik "Anthropic launched a monthly usage recap showing when you use Claude and what you build, plus the /checkup command that clears unused MCPs …" 58:30 Play} Ben's reflection is the most honest moment of the episode: he's maxed his Claude credits, which is what finally drove him to download Codex. He went to the park. He saw a friend. He scheduled a doctor's appointment. All things he'd been putting off because Claude is always there. He describes the psychological loop: if you're not building, someone else is. That's what you tell yourself. Ben compares it to iPhone Screen Time — everyone said they needed it, and it barely changed behavior. This is Screen Time for Claude. Rik closes the episode with a plug to like and subscribe, and Luca signs off heading for a glass of wine in the garden.

The episode's final segment delivers what Rik calls the most practically useful news of the week: Anthropic launched a monthly Claude usage recap showing when you engage most and what you work on, plus quiet-hour settings and break nudges. The /checkup command, surfaced by Cloud Code Guy Boris, cleans up unused MCPs and skills — the digital equivalent of clearing app clutter. ^{[1] — Rik "Anthropic launched a monthly usage recap showing when you use Claude and what you build, plus the /checkup command that clears unused MCPs …" 58:30 Play} Ben's reflection is the most honest moment of the episode: he's maxed his Claude credits, which is what finally drove him to download Codex. He went to the park. He saw a friend. He scheduled a doctor's appointment. All things he'd been putting off because Claude is always there. He describes the psychological loop: if you're not building, someone else is. That's what you tell yourself. Ben compares it to iPhone Screen Time — everyone said they needed it, and it barely changed behavior. This is Screen Time for Claude. Rik closes the episode with a plug to like and subscribe, and Luca signs off heading for a glass of wine in the garden.