# EP25: Anthropic vs OpenAI, SpaceX chaos, 25 lessons in 25 episodes > Source: > Published: 2026-08-08 21:49:04+00:00 # EP25: Anthropic vs OpenAI, SpaceX chaos, 25 lessons in 25 episodes A developer squeezed $50,000 in AI coding value from just $800 in subscriptions — and the hosts think the subsidy window is closing fast. God Mode Podcast # EP25: Anthropic vs OpenAI, SpaceX chaos, 25 lessons in 25 episodes A developer squeezed $50,000 in AI coding value from just $800 in subscriptions — and the hosts think the subsidy window is closing fast. TL;DR Rik and Ben mark 25 episodes of God Mode Pod by grading their biggest AI calls against what actually happened. SpaceX crashed 35% from IPO exactly as Rik predicted in episode 15 [1] — Ben Broch "Ben pulled live Google Trends data mid-episode: Claude searches dropped 30–40% over the summer while ChatGPT fell only 10%. The Codex vs Cl…" 22:40 , Claude's Google Trends share dropped 30–40% while ChatGPT rebounded [2] — Ben Broch "Claude Google Trends down 30–40%: Ben observed that Claude's Google Trends score dropped roughly 30–40% over the summer while ChatGPT fell …" 22:45 , and Rik cancelled his $100 Claude Max plan in favour of Cursor and Codex [3] — Ben Broch "Opus 5 benchmarks well but consistently underdelivers; Fable 5 just gets it done. Ben has learned to use Fable when he has credits and acce…" 32:55 . The episode's sharpest takeaway: average intelligence will be free long-term, so the real investment opportunity is in inference providers routing cheap models, not in frontier labs subsidising tokens with VC money [4] — Rik "Average AI intelligence will eventually be free, and frontier models will command a premium. The real investment opportunity isn't in the l…" 20:35 . SpaceX was down 35% since IPO and the unlock ladder is landing exactly as Rik called it on episode 15. Rik and Ben grade 25 episodes of AI calls: Anthropic vs OpenAI, Claude vs Codex vs Cursor, and why DeepSeek's pricing doesn't even show up on the chart. - The episode opens mid-argument with four rapid-fire clips that double as the episode's thesis statements — Ben's token supply warning, Rik's SpaceX call, and the proclamation that coding is the new podcasting. Rik then formally opens episode 25, noting the milestone with a characteristic flourish: 'a quarter of 100.' He explains that he fed all 25 episodes to Claude and asked it to distil the dominant themes, which it returned as three wars: OpenAI vs Anthropic, open vs closed source, and the Elon Corner. The episode's unusual retrospective format is set up here — a structured grading of past calls against what actually happened, bookended by the week's live news. - The episode opens mid-argument with four rapid-fire clips that double as the episode's thesis statements — Ben's token supply warning, Rik's SpaceX call, and the proclamation that coding is the new podcasting. Rik then formally opens episode 25, noting the milestone with a characteristic flourish: 'a quarter of 100.' He explains that he fed all 25 episodes to Claude and asked it to distil the dominant themes, which it returned as three wars: OpenAI vs Anthropic, open vs closed source, and the Elon Corner. The episode's unusual retrospective format is set up here — a structured grading of past calls against what actually happened, bookended by the week's live news. - The hosts open with a brief life update: Rik is turning 30 tomorrow and plans a pizza-and-cocktails night, while Ben is settling into Brooklyn and leaning into a new content experiment called Indi2Millie. Ben explains that friction in his previous content workflow — the editing grind of clipping, graphics, and posting — has been largely offloaded to Claude, freeing him to focus on recording. The catch, as Rik immediately spots, is that Ben now spends 80% of his time on content and only 20% on the three SaaS businesses he was previously racing to build. Ben is relaxed about this: the SaaS products are 'running on their own' and he wants to spend one to two months getting genuinely comfortable on camera. Rik frames it as the cringe mountain — the social discomfort of filming in public — and both hosts agree it's a skill worth building. - The hosts open with a brief life update: Rik is turning 30 tomorrow and plans a pizza-and-cocktails night, while Ben is settling into Brooklyn and leaning into a new content experiment called Indi2Millie. Ben explains that friction in his previous content workflow — the editing grind of clipping, graphics, and posting — has been largely offloaded to Claude, freeing him to focus on recording. The catch, as Rik immediately spots, is that Ben now spends 80% of his time on content and only 20% on the three SaaS businesses he was previously racing to build. Ben is relaxed about this: the SaaS products are 'running on their own' and he wants to spend one to two months getting genuinely comfortable on camera. Rik frames it as the cringe mountain — the social discomfort of filming in public — and both hosts agree it's a skill worth building. - Rik reads Finn's tweet in full: waking up, walking with AirPods, brain-dumping priorities to ChatGPT Voice, watching them execute in parallel, spending only 15 minutes at a desk. It sounds like liberation. Ben is unconvinced. His own experience, he says, is that babysitting AI requires a screen — you need to see the output to catch errors, and voice alone doesn't give you that fidelity. Where Finn claims only 5% of his day is review, Ben estimates 40% of his is. Rik then delivers the kicker: in January, Finn's big claim was that his agent worked while he slept; now he has to go for a walk with it. That's not six months of progress. Ben clinches it by pointing out Finn said the exact same desk-liberation things about his Telegram bot. He's a hype machine, and the timeline loves him for it. - Rik reads Finn's tweet in full: waking up, walking with AirPods, brain-dumping priorities to ChatGPT Voice, watching them execute in parallel, spending only 15 minutes at a desk. It sounds like liberation. Ben is unconvinced. His own experience, he says, is that babysitting AI requires a screen — you need to see the output to catch errors, and voice alone doesn't give you that fidelity. Where Finn claims only 5% of his day is review, Ben estimates 40% of his is. Rik then delivers the kicker: in January, Finn's big claim was that his agent worked while he slept; now he has to go for a walk with it. That's not six months of progress. Ben clinches it by pointing out Finn said the exact same desk-liberation things about his Telegram bot. He's a hype machine, and the timeline loves him for it. - After dismissing ChatGPT Goals as little more than a branded name for the prompt loop every coding agent already runs, Rik shares the week's most-talked-about developer thread. Xianjin Zhu ran both the $400 Claude Max and $400 Codex plans for 30 days, hitting usage caps daily, and found the Claude plan delivered $32,000 in API equivalent value — an 81x return — while Codex delivered $17,000. Total: $50,000 in subsidised compute for $800 paid. Rik traces this back to episode 1, when the hosts first started gaming the subscription math. Ben notes something curious: the report now shows Claude as the better deal per dollar, whereas previously Codex was the winner — suggesting the relative token economics are shifting. The uncomfortable question they both land on: this subsidy is being funded by VC money, and as SpaceX's post-IPO slide puts pressure on startup valuations, the window could close faster than builders expect. - The conversation sharpens into token efficiency. The viral report measures token value per dollar of subscription, but Rik flips the frame: what matters for anyone not on a capped plan is how many tokens each model burns per task. Opus 5 can consume close to 100 million tokens where Terra completes the same work in around 1.5 million — a difference approaching 65 to 100 times. Ben contextualises how enterprises are justifying this by comparing it to headcount savings; a $20,000-per-employee AI bill can beat a 20% staffing increase. But both hosts agree: they would never want to pay Anthropic's API rates for Opus at scale. The subsidy is the only reason the current arrangement works for small builders. - Rik pulls up the token-cost chart that had been circulating on his timeline: Anthropic, OpenAI, Moonshot, Alibaba Qwen all show visible bars. DeepSeek V4 Flash is barely a sliver. The gap between proprietary subscriptions and DeepSeek's open-source API is so small that the subscriptions actually win on pure token value. But then the news drops mid-recording: DeepSeek has announced major API price increases. Ben unpacks the nuance — DeepSeek is open-source, so users can theoretically pull the model and run it elsewhere, meaning the official API hike is really about DeepSeek trying to make money rather than a fundamental model cost increase. Still, the direction of travel is clear: even the cheapest corner of the AI market is now moving toward monetisation, and the era of austerity pricing may be giving way to something more expensive across the board. - Ben pulls in the macro framing from a recent Gavin Baker podcast appearance. Baker's argument is simple and stark: demand for AI tokens is growing roughly 10x over the next year, but data center commitments and supply-side build-out will only deliver around 3x more capacity. Something has to give — either labs cut pre-training compute to free up inference supply, or prices go up. Baker's view is that labs will not cut pre-training, because their investor stories are all built on reaching AGI, and you only reach AGI by continuing to pour resources into the next frontier model. Ben notes that both Fable and Claude are now at their sixth generation of models, converging in parallel. With only about 500,000 people globally using agentic coding workflows today against 8 billion potential users, the demand curve is almost incomprehensibly steep. The hosts' conclusion: the inference layer — companies like Cerebras, and the OpenRouter routing layer recently bid on by Stripe — is where the investment value will accumulate. - OpenAI's decision to cut SOL 5.6 Terra/Luna pricing by 80% provides the jumping-off point. Rik has been using Luna heavily precisely because 80% of his tasks don't need the best model, and the cost savings are dramatic. He ties this to a quote from the Palo Alto Networks CEO: average intelligence will be free in the long run, and as the frontier advances, today's exceptional intelligence becomes tomorrow's average. For consumers, this is almost purely good news — cheap, capable AI for most tasks, with frontier models reserved for work that genuinely demands them. The routing layer — knowing which model to deploy when — becomes the chokepoint, which is why OpenRouter's reported acquisition by Stripe is so strategically interesting. The implication for compute investment is that inference spend will bifurcate: commodity inference grows cheaply, but frontier inference becomes the premium product with pricing power. - Ben opens Google Trends live on air and shares what he sees: Claude peaked in February, roughly when both hosts deleted their ChatGPT subscriptions and went all-in on Anthropic. Since then, Claude has slid from a relative score of around 33 down to 21 — a drop of 30 to 40% — while ChatGPT's summer dip has been a comparatively mild 10%. OpenAI's CFO had boasted in July that the company added more net ARR that month than in all of Q2, and the Trends data corroborates it. The Codex vs Claude Code comparison tells a similar story: Claude Code dominated Q2, but by June Codex had pulled level. Rik connects this to what they observed in practice — Claude Code was everywhere on Instagram in the spring, but OpenAI was aggressively pushing Codex migrations on Twitter and it worked. - Ten weeks ago on episode 15, Rik made a structured bear case for SpaceX in the near term, anchored on the $101 billion unlock ladder. Today, recording on 6 August 2026, that ladder is landing. The stock peaked at $229 on IPO day, faded to $108, briefly spiked to $125 on Elon's Nvidia news, and is now sitting at $106 — well below the IPO price. More than double the public float is becoming tradeable today as insiders and team members who bought in far earlier become eligible to sell. A clip from the episode 15 discussion is played back: Rik forecasted a pop followed by a significant dump once the unlock arrived, and that is precisely what happened. The lesson he draws is dual: the thesis was right, but the more important lesson for investors is that entry price and patience matter as much as narrative conviction. At current levels, he says, he is starting to get interested in SpaceX for the long term. - With ChatGPT prompting the structure, Rik runs through each lesson as a headline and Ben riffs on how it aged. The series opens with the moment Dario Amodei and Sam Altman refused to hold hands at a group photo at a conference — a petty but revealing signal of how far relations between the two labs had deteriorated. Lesson two spotlights OpenAI's decision to focus on coding agents above all else, which Ben calls pivotal: the coding-agent flywheel (build better agents → agents improve themselves → everything else improves) turned out to be the year's biggest strategic insight. Lesson three surfaces the Mythos speculation, with Ben now claiming Anthropic trained Mythos in February and almost certainly has a Mythos 2 scored in the 80s or 90s on intelligence benchmarks. Lesson four addresses the Opus nerf — Ben was among the first to notice labs quietly throttling model effort during supply crunches, a practice he traces to around February. The GPT-5.5 86% hallucination benchmark from episode 10 gets a failing grade: fifteen weeks on, OpenAI still hasn't meaningfully moved the needle. The section closes with the Codex vs Claude Code call — once a clear Claude Code victory, now a dead heat — as OpenAI's aggressive developer migration campaign on Twitter bore fruit. - This section is the episode's most personal — a live confession of subscription fatigue. Rik traces the decision back to May 16th when he first cancelled Claude Max, and says the feeling now is permanent. Cursor running on auto feels genuinely unlimited; GPT-5.6 SOL and Luna do everything he needs; and the Anthropic models simply don't feel like upgrades. The specific irritant: Sonnet, Claude's third-best model, consuming 10% of his monthly token allowance in a single prep session for this podcast. For a plan that supposedly delivers $32,000 in API value, this is a jarring user experience. Ben adds his own data point: at roughly 65–70% through the week, he's already at 95% of his Claude credits and 100% of his Codex plan. He briefly tried downgrading from the $200 to the $100 plan two months ago; he upgraded back the next day. The $200 plan is the minimum viable subscription for anyone doing serious agentic work. - The open vs closed source war has been the most economically consequential theme of the archive. Lesson nine recasts the Bloomberg terminal moment from episode 2 — building a $30,000 professional tool for free with vibe coding — as the moment the SaaS disruption thesis went from speculative to obvious. The 850x token cost crash since 2020 is then immediately complicated by everything the hosts have been discussing today: supply constraints, DeepSeek price hikes, Gavin Baker's 10x demand warning. The lesson isn't that costs will keep falling; it's that Jevons paradox is real and demand will absorb any efficiency gains. Airtable's acquisition brings the SaaS apocalypse into sharp relief: $1.2 billion for a company worth $11 billion four years ago. Ben's analytical frame is clean — products whose value lived in their user interface are being destroyed, while infrastructure plays like Vercel and Supabase (the vibe-coding era's equivalent of Airtable for the no-code era) are thriving. ElevenLabs closes the section as the episode's positive case study: an application-layer AI company that has survived commoditisation by staying specialised, customisable, and execution-focused. - The Elon Corner is the episode's most speculative and entertaining section, combining confirmed calls with forward-looking chess moves. The $380 billion compute contract between Anthropic and SpaceX gets recapped first — a deal that now looks like both parties building leverage over each other. The dark fiber analogy resurfaces: unlike the 2000s telecom bust, every GPU in production today is being utilized, which is why the bubble thesis doesn't quite fit. SpaceX's acquisition of Cursor is flagged as a sleeper story: Rik has been getting near-unlimited auto-mode use from Cursor since the deal, routing through both the Composer model and Grok 4.5. The most intriguing speculation is Ben's 4D chess framing around Grok: if Elon open-sources Grok, it instantly makes every proprietary frontier model look overpriced by comparison. Grok is already one-tenth the cost of Claude and ChatGPT. Combined with Elon feeding all of SpaceX's engineering history into the next model, the hosts suggest Grok could become the default go-to model for anyone unwilling to pay frontier prices. - Ben lands the episode's emotional close with characteristic self-awareness. He walks back to the Manus moment — dozens of phones and computers running AI in tandem, everyone on the timeline declaring AGI had arrived. A year on, the models are significantly better, but they still require constant human direction. He frames this not as disappointment but as opportunity: builders who can steer AI well, catch its mistakes, and direct its output are still essential, and that skill is durable in a way pure hype cycles are not. Rik adds his own lesson 25 from the investing angle: average intelligence will be free, and the capital will flow through the inference providers routing those cheap models. The routing layer — OpenRouter, Stripe's apparent interest in it, Elon Web Services — is where the structural opportunity lives. Both hosts close bullish on the next 25 episodes, and Rik puts out a direct call for sponsors, offering a referral fee to any listener who connects them with a brand deal. - AGI (Artificial General Intelligence) - AI that matches or exceeds human-level intelligence across all cognitive tasks; debated as a milestone the hosts argue has not yet been reached. - Agentic coding - Using an AI model to autonomously execute multi-step coding tasks in loops — planning, executing, reviewing — rather than responding to single prompts. - Token - The unit AI models use to process text; roughly 0.75 words. Pricing and usage caps are measured in tokens consumed per request. - Inference - Running a trained AI model to generate outputs for users, as opposed to training the model. Inference spend is the cost of serving AI at scale. - Distillation attack - The practice of using outputs from a proprietary AI model to train a competing or open-source model, effectively copying its capabilities without access to training data. - Nerf - In gaming slang, to weaken something; here used to describe AI labs secretly reducing model effort or quality during periods of compute scarcity. - Vibe coding - Writing software by describing intent to an AI coding tool in natural language, without writing traditional code manually; associated with tools like Cursor and Codex. - Jevons paradox - The economic observation that as a resource becomes cheaper and more efficient, total consumption rises rather than falls; applied here to AI tokens. - SWE-Bench - A benchmark measuring an AI model's ability to solve real-world software engineering tasks; cited in the episode as an example of a benchmark OpenAI retracted. - Pre-training - The initial large-scale training phase of an AI model on vast datasets; requires enormous compute and is distinct from fine-tuning or inference. - Float (public float) - The proportion of a company's shares available for public trading. SpaceX's float more than doubled when the $101B insider unlock occurred. - ARR (Annual Recurring Revenue) - A SaaS metric measuring the annualised value of subscription revenue; cited in reference to OpenAI's CFO claiming record ARR additions in July. - 4D chess - A metaphor for an extremely complex, multi-layered strategic move whose full implications are non-obvious to most observers. - Austerity (age of austerity) - Period of cost-cutting and reduced spending; used by Ben to describe a shift even among Chinese open-source AI providers toward monetisation over growth-at-any-cost. - Commensurate - Proportionate or corresponding in size or degree; Ben used it to say Opus 5's confidence level is not commensurate with its actual task performance. - Dark fiber - Unused fibre-optic cable built out speculatively during the 1990s telecom boom; invoked as an analogy for potential AI infrastructure overinvestment. - Neo Clouds - A new generation of cloud compute providers (e.g. CoreWeave, Elon Web Services) focused on GPU-intensive AI workloads, distinct from traditional hyperscalers. - Computer use - An AI capability where a model can autonomously control a computer — clicking, browsing, and interacting with applications — rather than just generating text. - Harness - In AI engineering, a framework or scaffolding that wraps an AI model and feeds it prompts in loops to accomplish a complex goal. - Capitulation - In investing, the point where the last sellers give up and sell, often marking a market bottom; Rik suggested Airtable's sale might signal SaaS capitulation. Chapter 1 · 00:00 ## Cold open, the SpaceX call The episode opens mid-argument with four rapid-fire clips that double as the episode's thesis statements — Ben's token supply warning, Rik's SpaceX call, and the proclamation that coding is the new podcasting. Rik then formally opens episode 25, noting the milestone with a characteristic flourish: 'a quarter of 100.' He explains that he fed all 25 episodes to Claude and asked it to distil the dominant themes, which it returned as three wars: OpenAI vs Anthropic, open vs closed source, and the Elon Corner. The episode's unusual retrospective format is set up here — a structured grading of past calls against what actually happened, bookended by the week's live news. Chapter 2 · 00:31 ## Hello everyone, 25 episodes and three wars The episode opens mid-argument with four rapid-fire clips that double as the episode's thesis statements — Ben's token supply warning, Rik's SpaceX call, and the proclamation that coding is the new podcasting. Rik then formally opens episode 25, noting the milestone with a characteristic flourish: 'a quarter of 100.' He explains that he fed all 25 episodes to Claude and asked it to distil the dominant themes, which it returned as three wars: OpenAI vs Anthropic, open vs closed source, and the Elon Corner. The episode's unusual retrospective format is set up here — a structured grading of past calls against what actually happened, bookended by the week's live news. Chapter 6 · 06:46 ## Alex Finn's ChatGPT Voice hype, graded Rik reads Finn's tweet in full: waking up, walking with AirPods, brain-dumping priorities to ChatGPT Voice, watching them execute in parallel, spending only 15 minutes at a desk. It sounds like liberation. Ben is unconvinced. His own experience, he says, is that babysitting AI requires a screen — you need to see the output to catch errors, and voice alone doesn't give you that fidelity. Where Finn claims only 5% of his day is review, Ben estimates 40% of his is. Rik then delivers the kicker: in January, Finn's big claim was that his agent worked while he slept; now he has to go for a walk with it. That's not six months of progress. Ben clinches it by pointing out Finn said the exact same desk-liberation things about his Telegram bot. He's a hype machine, and the timeline loves him for it. Chapter 7 · 10:22 ## The goals myth, then $50K for $800 in subscriptions After dismissing ChatGPT Goals as little more than a branded name for the prompt loop every coding agent already runs, Rik shares the week's most-talked-about developer thread. Xianjin Zhu ran both the $400 Claude Max and $400 Codex plans for 30 days, hitting usage caps daily, and found the Claude plan delivered $32,000 in API equivalent value — an 81x return — while Codex delivered $17,000. Total: $50,000 in subsidised compute for $800 paid. Rik traces this back to episode 1, when the hosts first started gaming the subscription math. Ben notes something curious: the report now shows Claude as the better deal per dollar, whereas previously Codex was the winner — suggesting the relative token economics are shifting. The uncomfortable question they both land on: this subsidy is being funded by VC money, and as SpaceX's post-IPO slide puts pressure on startup valuations, the window could close faster than builders expect. [$50K in AI Value for $800 — The Subsidy Window Is Open](/bit/podbit/11137/) A developer extracted $50,000 in API value from just $800 in Claude and Codex subscriptions over 30 days. The hosts warn this subsidy era is finite — VC money is propping it up and the window will close. [$50K coding value for $800 subscriptions](/bit/snapshot/12183/) A developer tested Claude Code and Codex subscriptions for 30 days and unlocked $50,000 in API value for only $800 in subscription costs. [81x return on $400 Claude plan](/bit/snapshot/12184/) The $400 Claude Max plan delivered $32,000 in API equivalent value, an 81x return over the subscription price. Chapter 8 · 13:16 ## Opus burns 100x more tokens than Terra The conversation sharpens into token efficiency. The viral report measures token value per dollar of subscription, but Rik flips the frame: what matters for anyone not on a capped plan is how many tokens each model burns per task. Opus 5 can consume close to 100 million tokens where Terra completes the same work in around 1.5 million — a difference approaching 65 to 100 times. Ben contextualises how enterprises are justifying this by comparing it to headcount savings; a $20,000-per-employee AI bill can beat a 20% staffing increase. But both hosts agree: they would never want to pay Anthropic's API rates for Opus at scale. The subsidy is the only reason the current arrangement works for small builders. [Opus burns 100x more tokens than Terra](/bit/snapshot/12185/) Opus 5 can spend nearly 100 million tokens on a task where a cheaper model like Terra uses only 1.5 million, a ~65x difference in token consumption. [DeepSeek: Too Cheap to See on the Chart — Then a Price Hike](/bit/podbit/11147/) DeepSeek's API was so cheap it literally didn't register on a token-cost comparison chart. Then, mid-episode, news broke of significant DeepSeek price hikes. Even the cheapest open-source alternative is now joining the move toward monetisation. Chapter 9 · 15:02 ## DeepSeek, too cheap to see on the chart Rik pulls up the token-cost chart that had been circulating on his timeline: Anthropic, OpenAI, Moonshot, Alibaba Qwen all show visible bars. DeepSeek V4 Flash is barely a sliver. The gap between proprietary subscriptions and DeepSeek's open-source API is so small that the subscriptions actually win on pure token value. But then the news drops mid-recording: DeepSeek has announced major API price increases. Ben unpacks the nuance — DeepSeek is open-source, so users can theoretically pull the model and run it elsewhere, meaning the official API hike is really about DeepSeek trying to make money rather than a fundamental model cost increase. Still, the direction of travel is clear: even the cheapest corner of the AI market is now moving toward monetisation, and the era of austerity pricing may be giving way to something more expensive across the board. [DeepSeek price hike mid-episode](/bit/snapshot/12191/) DeepSeek, previously so cheap its pricing barely registered on token-cost charts, announced significant API price increases mid-episode. Chapter 10 · 17:40 ## Gavin Baker, 10x demand, 3x supply Ben pulls in the macro framing from a recent Gavin Baker podcast appearance. Baker's argument is simple and stark: demand for AI tokens is growing roughly 10x over the next year, but data center commitments and supply-side build-out will only deliver around 3x more capacity. Something has to give — either labs cut pre-training compute to free up inference supply, or prices go up. Baker's view is that labs will not cut pre-training, because their investor stories are all built on reaching AGI, and you only reach AGI by continuing to pour resources into the next frontier model. Ben notes that both Fable and Claude are now at their sixth generation of models, converging in parallel. With only about 500,000 people globally using agentic coding workflows today against 8 billion potential users, the demand curve is almost incomprehensibly steep. The hosts' conclusion: the inference layer — companies like Cerebras, and the OpenRouter routing layer recently bid on by Stripe — is where the investment value will accumulate. [Gavin Baker's Warning: Tokens Are About to Get More Expensive](/bit/podbit/11138/) Investor Gavin Baker laid out a stark supply-demand mismatch for AI tokens: demand is growing 10x next year while supply grows only 3x. Frontier labs won't cut pre-training to free up capacity, so consumer prices are heading up. [Token demand 10x-ing, supply only 3x-ing](/bit/snapshot/12186/) Gavin Baker argued on Invest Like the Best that AI token demand is growing 10x next year while supply is only growing 3x, implying prices must rise. [Only 500K people using agentic coding](/bit/snapshot/12187/) Ben estimated that only around 500,000 people globally currently use any form of agentic coding workflow, out of 8 billion people. [Average Intelligence Will Be Free — The Inference Investing Thesis](/bit/podbit/11136/) Average AI intelligence will eventually be free, and frontier models will command a premium. The real investment opportunity isn't in the labs subsidising tokens with VC money — it's in the inference providers routing cheap models at scale. Chapter 12 · 22:32 ## Google Trends, Claude versus ChatGPT, Codex versus Claude Code Ben opens Google Trends live on air and shares what he sees: Claude peaked in February, roughly when both hosts deleted their ChatGPT subscriptions and went all-in on Anthropic. Since then, Claude has slid from a relative score of around 33 down to 21 — a drop of 30 to 40% — while ChatGPT's summer dip has been a comparatively mild 10%. OpenAI's CFO had boasted in July that the company added more net ARR that month than in all of Q2, and the Trends data corroborates it. The Codex vs Claude Code comparison tells a similar story: Claude Code dominated Q2, but by June Codex had pulled level. Rik connects this to what they observed in practice — Claude Code was everywhere on Instagram in the spring, but OpenAI was aggressively pushing Codex migrations on Twitter and it worked. [Claude's Google Trends Collapse vs ChatGPT's Resilience](/bit/podbit/11141/) Ben pulled live Google Trends data mid-episode: Claude searches dropped 30–40% over the summer while ChatGPT fell only 10%. The Codex vs Claude Code race is now neck-and-neck after Claude Code led all of Q2. [Claude Google Trends down 30–40%](/bit/snapshot/12188/) Ben observed that Claude's Google Trends score dropped roughly 30–40% over the summer while ChatGPT fell only about 10%, signalling a sentiment shift back to OpenAI. Chapter 13 · 25:02 ## The SpaceX bear case, graded Ten weeks ago on episode 15, Rik made a structured bear case for SpaceX in the near term, anchored on the $101 billion unlock ladder. Today, recording on 6 August 2026, that ladder is landing. The stock peaked at $229 on IPO day, faded to $108, briefly spiked to $125 on Elon's Nvidia news, and is now sitting at $106 — well below the IPO price. More than double the public float is becoming tradeable today as insiders and team members who bought in far earlier become eligible to sell. A clip from the episode 15 discussion is played back: Rik forecasted a pop followed by a significant dump once the unlock arrived, and that is precisely what happened. The lesson he draws is dual: the thesis was right, but the more important lesson for investors is that entry price and patience matter as much as narrative conviction. At current levels, he says, he is starting to get interested in SpaceX for the long term. [SpaceX Bear Case Confirmed: $101B Unlock Lands Exactly as Called](/bit/podbit/11140/) Ten episodes ago Rik laid out a SpaceX bear case centred on a $101 billion insider unlock. Today it's playing out live: the stock is 35% below IPO, more than doubling the public float. The lesson isn't just about SpaceX — it's about entry price mattering as much as thesis quality. [SpaceX down 35% from IPO](/bit/snapshot/12189/) SpaceX dropped from an IPO peak of $229 to around $106, roughly 35% below IPO price, as $101 billion in insider share unlocks hit the market. [$101B in SpaceX share unlocks](/bit/snapshot/12190/) SpaceX was facing $101 billion in share unlocks, more than doubling the public float from roughly 4–5% to about 11% of total shares. Chapter 14 · 28:03 ## 25 lessons, war one, OpenAI versus Anthropic With ChatGPT prompting the structure, Rik runs through each lesson as a headline and Ben riffs on how it aged. The series opens with the moment Dario Amodei and Sam Altman refused to hold hands at a group photo at a conference — a petty but revealing signal of how far relations between the two labs had deteriorated. Lesson two spotlights OpenAI's decision to focus on coding agents above all else, which Ben calls pivotal: the coding-agent flywheel (build better agents → agents improve themselves → everything else improves) turned out to be the year's biggest strategic insight. Lesson three surfaces the Mythos speculation, with Ben now claiming Anthropic trained Mythos in February and almost certainly has a Mythos 2 scored in the 80s or 90s on intelligence benchmarks. Lesson four addresses the Opus nerf — Ben was among the first to notice labs quietly throttling model effort during supply crunches, a practice he traces to around February. The GPT-5.5 86% hallucination benchmark from episode 10 gets a failing grade: fifteen weeks on, OpenAI still hasn't meaningfully moved the needle. The section closes with the Codex vs Claude Code call — once a clear Claude Code victory, now a dead heat — as OpenAI's aggressive developer migration campaign on Twitter bore fruit. [25 Episodes, Three Wars: OpenAI vs Anthropic, Open vs Closed, the Elon Corner](/bit/podbit/11148/) Rik used Claude to analyse all 25 episodes and distil them into three recurring wars: OpenAI vs Anthropic, open vs closed source, and the Elon Corner. The 25-lesson retrospective grades every major call against what actually happened. [Mythos 2 could score in the 90s](/bit/snapshot/12195/) Ben speculated that Anthropic has already trained a Mythos 2 model scoring in the 80s or even 90s on intelligence benchmarks, but has not released it. [Labs Secretly Nerf Models When Supply Gets Tight](/bit/podbit/11143/) When AI labs are supply-constrained, they quietly reduce model effort levels in the backend. Users experience a noticeably worse product with no announcement. Ben says this has been happening since at least Opus 4.6 in early 2026. [GPT-5.5 hallucinates 86% of the time](/bit/snapshot/12194/) According to a benchmark cited in episode 10, GPT-5.5 had an 86% hallucination rate compared to Grok's 17% at the time. [Fable vs Opus: Why Ben Always Picks the OpenAI Model](/bit/podbit/11142/) Opus 5 benchmarks well but consistently underdelivers; Fable 5 just gets it done. Ben has learned to use Fable when he has credits and accepts a worse product the rest of the time — a revealing admission about the gap between benchmark and lived experience. Chapter 15 · 34:36 ## Rik cancels the $100 Claude Max plan again This section is the episode's most personal — a live confession of subscription fatigue. Rik traces the decision back to May 16th when he first cancelled Claude Max, and says the feeling now is permanent. Cursor running on auto feels genuinely unlimited; GPT-5.6 SOL and Luna do everything he needs; and the Anthropic models simply don't feel like upgrades. The specific irritant: Sonnet, Claude's third-best model, consuming 10% of his monthly token allowance in a single prep session for this podcast. For a plan that supposedly delivers $32,000 in API value, this is a jarring user experience. Ben adds his own data point: at roughly 65–70% through the week, he's already at 95% of his Claude credits and 100% of his Codex plan. He briefly tried downgrading from the $200 to the $100 plan two months ago; he upgraded back the next day. The $200 plan is the minimum viable subscription for anyone doing serious agentic work. [Rik Cancels Claude Max — Again](/bit/podbit/11139/) Rik cancelled his $100 Claude Max subscription, this time for good, because Cursor and Codex now handle everything he needs. The irony: Anthropic's plans feel stingy in practice even while delivering record token subsidies on paper. Chapter 16 · 37:59 ## 25 lessons, war two, closed versus open source The open vs closed source war has been the most economically consequential theme of the archive. Lesson nine recasts the Bloomberg terminal moment from episode 2 — building a $30,000 professional tool for free with vibe coding — as the moment the SaaS disruption thesis went from speculative to obvious. The 850x token cost crash since 2020 is then immediately complicated by everything the hosts have been discussing today: supply constraints, DeepSeek price hikes, Gavin Baker's 10x demand warning. The lesson isn't that costs will keep falling; it's that Jevons paradox is real and demand will absorb any efficiency gains. Airtable's acquisition brings the SaaS apocalypse into sharp relief: $1.2 billion for a company worth $11 billion four years ago. Ben's analytical frame is clean — products whose value lived in their user interface are being destroyed, while infrastructure plays like Vercel and Supabase (the vibe-coding era's equivalent of Airtable for the no-code era) are thriving. ElevenLabs closes the section as the episode's positive case study: an application-layer AI company that has survived commoditisation by staying specialised, customisable, and execution-focused. [Token cost crashed 850x since 2020](/bit/snapshot/12196/) The hosts noted that token costs had crashed 850x since 2020, though the trend appeared to be reversing with supply constraints pushing prices back up. [The SaaS Apocalypse Gets a Price Tag: Airtable's $1.2B Exit](/bit/podbit/11144/) Airtable was sold for $1.2 billion against a 2021 peak of $11 billion. The pattern is clear: products whose value lived in their user interface are being destroyed by AI, while infrastructure plays like Vercel and Supabase are thriving. [Airtable sold for $1.2B vs $11B peak](/bit/snapshot/12192/) Airtable was acquired by European private equity for approximately $1.2 billion, compared to its peak valuation of $11 billion in 2021. [Inference spend doubling every 45 days](/bit/snapshot/12193/) Chamath's firm 8090 reported that inference spend is doubling every 45 days, driven by continuously expanding AI use cases. Chapter 17 · 45:52 ## 25 lessons, war three, the Elon corner deep dive The Elon Corner is the episode's most speculative and entertaining section, combining confirmed calls with forward-looking chess moves. The $380 billion compute contract between Anthropic and SpaceX gets recapped first — a deal that now looks like both parties building leverage over each other. The dark fiber analogy resurfaces: unlike the 2000s telecom bust, every GPU in production today is being utilized, which is why the bubble thesis doesn't quite fit. SpaceX's acquisition of Cursor is flagged as a sleeper story: Rik has been getting near-unlimited auto-mode use from Cursor since the deal, routing through both the Composer model and Grok 4.5. The most intriguing speculation is Ben's 4D chess framing around Grok: if Elon open-sources Grok, it instantly makes every proprietary frontier model look overpriced by comparison. Grok is already one-tenth the cost of Claude and ChatGPT. Combined with Elon feeding all of SpaceX's engineering history into the next model, the hosts suggest Grok could become the default go-to model for anyone unwilling to pay frontier prices. [10% improvement costs 100% more tokens](/bit/snapshot/12197/) Ben observed that to make an AI product 10% better than before now requires spending 100% more tokens, meaning optimisation gains are shrinking. [Grok Open Source: Elon's Potential 4D Chess Move](/bit/podbit/11146/) If Elon makes Grok open source, it could be a decisive competitive strike: Grok is already one-tenth the price of Claude and ChatGPT, and open-sourcing it would let anyone run it on commodity compute. Ben calls it a 'badass 4D chess move'. Chapter 18 · 51:26 ## Lesson 25, we haven't hit AGI Ben lands the episode's emotional close with characteristic self-awareness. He walks back to the Manus moment — dozens of phones and computers running AI in tandem, everyone on the timeline declaring AGI had arrived. A year on, the models are significantly better, but they still require constant human direction. He frames this not as disappointment but as opportunity: builders who can steer AI well, catch its mistakes, and direct its output are still essential, and that skill is durable in a way pure hype cycles are not. Rik adds his own lesson 25 from the investing angle: average intelligence will be free, and the capital will flow through the inference providers routing those cheap models. The routing layer — OpenRouter, Stripe's apparent interest in it, Elon Web Services — is where the structural opportunity lives. Both hosts close bullish on the next 25 episodes, and Rik puts out a direct call for sponsors, offering a referral fee to any listener who connects them with a brand deal. [We Haven't Hit AGI — And That's Bullish](/bit/podbit/11145/) Despite the Manus demo that looked like AGI a year ago, and despite Fable 5's capabilities, AGI hasn't arrived. Ben is more bullish because of it: models are great but still need human direction, and that's exactly where skilled builders create value. No indexed bits in this chapter. ## Show stoppers ## Snapshots () ## Key Quotes () Sign in to keep viewing Create a free account to keep exploring this episode's insights, snapshots, quotes and claims.