Anthropic gave Fable back at half rate limits for exactly 7 days — and Theo thinks it's a calculated GPU provisioning experiment, not a gift.
Nerd Snipe with Theo and Ben
Anthropic gave Fable back at half rate limits for exactly 7 days — and Theo thinks it's a calculated GPU provisioning experiment, not a gift.
TL;DR
Theo and Ben break down the AI engineering landscape in the wake of the Fable (Claude's "Mythos" model) export ban, the AI Engineer Conference in SF, and what it actually means to work in this industry right now. They cover Microsoft's $2.5B forward-deployed engineering push, why AI-generated Claude Code skills fall apart, how Anthropic's two-stage safety classifier triggers Fable-to-Opus rerouting, whether GLM-5.2 can touch Sonnet 5, and the mechanics of Fable's restricted re-release at half the rate limits [1] — Theo "The 7-day Fable window at half rate limits wasn't a consolation prize — it was a calculated GPU provisioning experiment. Anthropic needed e…" 47:17 . The single most useful takeaway: don't let Fable spin off Fable subagents — have Fable command cheaper models like Sonnet 5 or GLM-5.2 instead [2] — Theo "Anthropic's rerouting system has two stages: a cheap first-stage classifier (~0.05% compute) that watches which expert pathways activate in…" 1:02:50 .
Fable was supposed to give us 14 days and instead we got 3 before the export ban. Now it's back at half the rate limits. Also, we break down why AI-generated Claude Code skills fall apart, what actually triggers Fable-to-Opus rerouting, the Mythos export-control timeline, and whether GLM 5.2 can really touch Sonnet 5.
The episode kicks off with Theo turning the tables and making Ben do the intro, an arrangement Ben was clearly not fully prepared for. He rattles off the major topics — Mythos, forward-deployed engineers, the 5.6 model holdout, Fable's return, and Anthropic's ongoing China disputes — before stumbling to a landing. Theo closes the cold open with a reminder that the show is now on every major podcast platform, and urges listeners to leave comments and ratings, with a personal wave to producer Alyssa who is apparently keeping tabs.
Theo opens the news segment with a primer on forward-deployed engineers — a role that has existed for years but is suddenly central to enterprise AI adoption. The job is simple: embed at a client company, focus on 3 to 5 people per month, and physically make them use the product. [1] — Theo "Microsoft $2.5B FDE investment: Microsoft invested $2.5 billion in hiring approximately 6,000 forward-deployed engineering experts to help …" 01:29 Microsoft has taken this playbook to an extreme, committing $2.5 billion and 6,000 FDE hires to accelerate adoption across its enterprise customer base. Ben frames it bluntly as 'consulting in 2026,' noting that even attendees at the AI Engineer Conference — a self-selected technically sophisticated crowd — had barely scratched the surface of what current models can do. The conversation pivots to a harder question: who should companies hire for these roles? Theo argues strongly that the answer is AI-native young engineers, not experienced developers learning AI retroactively, because the former grew up with the tools the same way today's kids grew up without knowing file systems. He closes with a pointed plea: don't pull the ladder up behind you — teach the younger generation about how old systems worked instead of trying to reskill a 32-year-old on Claude.
The AI Engineer Conference in San Francisco leveled up considerably this year, ballooning to roughly 7,000 attendees from an estimated 1,000–2,000 the prior year — a growth both hosts describe as a tier jump. Ben mostly skipped the talks in favor of hallway conversations, but made an exception for Theo's, which he calls Theo's second-best talk overall after the Elixir Conf one. Theo had agreed to speak just a day or two before the event and was aiming for a million-view breakout moment, inspired by Matt Pocock having achieved that twice. The meetup story is a highlight: Theo put out a spontaneous tweet asking if anyone would host a meet-and-greet, OpenAI stepped up and offered their booth, and the turnout far exceeded expectations. For audience members who couldn't get through the crowd, cornering Ben and collaborator Julius in the back apparently worked just as well.
Matt Pocock's missed AIE talk — posted online instead — becomes the launching pad for one of the episode's most technically dense segments. His core argument: don't let AI write your Claude Code skills. The reason is elegant and non-obvious. [1] — Theo "The fundamental problem with letting AI write your Claude Code skills is that models write what they already do by default. Skills are supp…" 14:00 The entire purpose of a skill is to give the model behavior it doesn't have by default. When the AI writes the skill itself, it naturally gravitates toward behaviors it already exhibits, creating elaborate no-ops stuffed with filler like 'be really thorough' — instructions the model would follow anyway. Ben admits he's guilty of this pattern himself. The right workflow, Theo argues, is to go through your full model usage history, find every time you had to correct the model, categorize those failures, and write skills that specifically redirect the model away from those documented failures. Both hosts then open their .clod directories live and compare skill counts — Ben has 8 in his global directory and over 100 in his Hermes agent (after pruning 50), while Theo has trimmed aggressively to just a handful. The discussion covers workflow skills, framework-specific skills (Svelte, Effect), and the important distinction between things that belong in Agents.md versus a full skill.
The Alibaba distillation story is a perfect illustration of how frontier model data gets exfiltrated at scale. Theo explains the alleged mechanism: Alibaba — or more likely a third-party broker — set up discounted Claude Code subscriptions targeted at users in regions where Anthropic accounts are hard to obtain, then routed all traffic through a forwarding proxy that captured every input and output. The result was 28.8 million Claude exchanges across 25,000 fraudulent accounts. [1] — Theo "Alibaba: 28.8M exchanges stolen: Alibaba allegedly scraped 28.8 million Claude exchanges via 25,000 fraudulent accounts to use as model tra…" 25:33 Anthropic's response was a geographic detection code block embedded in Claude Code that checked for signals like Shanghai-based logins — flagged publicly by a user as 'hidden spyware,' then acknowledged by Anthropic's Thorik as a March experiment that had already been superseded by better mitigations. Theo connects this directly to the ViProxy proxy tool he and Ben use to balance accounts, predicting it will be killed within 2–3 weeks once Anthropic adds cryptographic session signatures to Claude Code's proprietary OAuth endpoint. Ben notes that the proxy model is only routing to Claude Code — not to other harnesses — but acknowledges the risk is real.
Ben uses the PostHog ad to illustrate a concept he and Theo have been discussing more lately: self-driving product loops. PostHog already collects product analytics signals, groups them into reports, and then — through an agent — opens PRs to your actual codebase to fix the underlying issues. After you merge, PostHog measures the impact and loops back. Ben highlights that PostHog is uniquely positioned for this because they already have all the signal data, and praises both their aggressive price-decreasing track record and their MCP server as tools for teams that want to build custom versions of this loop.
The proposed 5% OpenAI stake for the Trump administration sparks one of the episode's most pointed political tangents. Theo opens with a personal message to Sam Altman thanking him for doing what nobody else had the courage to do — publicly and aggressively courting the administration in order to preserve AI progress. The move is immediately called out as performative theater, compared unfavorably to Pride Month rainbow profile pictures. But Theo argues it's the necessary theater of the moment: make the government feel like they control the decision-making process long enough for actual AI safety infrastructure to be built, then hand that infrastructure to them as a fait accompli. [1] — Theo "5% of OpenAI ÷ 350M citizens = ~$143/person: If OpenAI hits a $1 trillion valuation, a 5% government stake divided among 350 million US cit…" 38:32 The math lands as a punchline: at a $1 trillion valuation, 5% divided among 350 million Americans equals $143 — less than one month of ChatGPT Pro. Theo also notes the delicious irony that if citizen equity is ever distributed, Sam Altman — who famously holds no equity in OpenAI — would finally have the same stake in the company as every American, including Theo and Ben.
With Fable back in their hands, both hosts have been running it hard. Ben's first major use case was rebuilding the brittle CLI that spun up his Hermes agent container — a task involving Hermes, Codex, Claude Code, and an Executor, all needing to live in a single deployable box. What Opus had built barely worked; Fable rebuilt the entire architecture in about an hour of back-and-forth. [1] — Theo "Theo gave Fable his entire backlog of stale, conflicted, and draft PRs and told it to figure out which ones to keep, merge, rebase, or kill…" 41:40 Theo's story is more cinematic: he had 18-plus stale pull requests sitting in his Lakebed project, some finished but unrebased, others superseded by later merges, a few just outright slop. Rather than manually triaging them, he asked Fable to read all the PRs, figure out what they were for, identify overlaps, write plans for anything worth salvaging, and then execute. He read the plans briefly, said 'YOLO — slash goal, complete all the work we talked about,' and watched the model close a dozen PRs and land 14 more in one loop. The thing that makes this work, Theo emphasizes, is that Fable can pull intent from incomplete context — it doesn't need you to re-explain what you were trying to do; it can figure it out.
Theo's conspiracy theory for the episode is arguably his best. The 7-day, half-rate-limit Fable window looks like a consolation prize for the 14-day window that was cut short by the export ban. But Theo proposes a different read: Anthropic needed exactly one full weekly cycle of subscriber usage data to understand usage patterns — how hard people burn on weekdays versus weekends, during working hours versus late nights — so they can provision GPU capacity between enterprise customers and subscribers. You can't get that data from 3 days, and you don't need 14. [1] — Theo "The 7-day Fable window at half rate limits wasn't a consolation prize — it was a calculated GPU provisioning experiment. Anthropic needed e…" 47:17 The FOMO built up during the export ban was a feature, not a bug: by announcing access the day before it went live, Anthropic ensured every subscriber who was going to burn would be ready to go on day one, giving them the clearest possible signal on peak usage. And the half rate limit isn't cruelty — it's Anthropic not having the spare compute to give full access while enterprises are still onboarding and claiming their allocations. Ben's counter — that they also get training data out of the deal — is gently corrected: Theo notes that most power users have training data opt-out enabled, and the provisioning data is what actually matters.
The pricing story around Fable is revealing. It launched at roughly $125 per million output tokens, then dropped to $50 — still 30–40% more expensive than Opus on a per-token basis, but Fable uses significantly fewer tokens per equivalent task. That means Fable often ends up cheaper than Opus in practice, which Theo argues is the death knell for Opus as a standalone model. [1] — Theo "Fable dropped from $125 to $50 per million output tokens, making it 30–40% pricier than Opus per token — but far cheaper overall because it…" 54:50 He's already made peace with paying API prices once the subscription window closes: he spent $150–$200 closing over 30 PRs in a single 5-hour session, and frames it as a magic button he'd press twice a day if he could. Ben agrees that any serious company would make the same calculation. The conversation turns to whether Opus 5 can survive in a world where Fable is this affordable — both hosts think Opus 5 needs to be priced lower than current Opus just to remain relevant.
Ben opens with a confession: on his first day with Fable back, he let it spawn Fable subagents on xHigh reasoning mode — forgetting to turn it off after prior testing — and burned approximately 30% of a weekly usage allocation with minimal useful output. [1] — Ben "Running Fable subagents from a Fable orchestrator is a fast path to burning 30% of your weekly limit on nothing useful. The right pattern i…" 57:56 Theo introduces a meme to illustrate the lesson: in Claude Code's reasoning effort selector, low/medium/high are labeled 'good options,' xHigh and Max are labeled 'mental illness,' and UltraCode is a 'skill disguised as a reasoning effort' (because Anthropic is using High under the hood). The correct architecture for serious work is Fable orchestrating cheaper models — Sonnet 5, GLM-5.2, or similar — for implementation while Fable handles planning, context synthesis, and review. The reasoning level discussion also touches on the non-obvious fact that spending more tokens and using higher reasoning effort past a certain point actually degrades output quality rather than improving it.
A passing mention of a Hetzner server dying — including its volume — opens the valve for a long-suppressed rant. Both hosts are emphatic: the $5 VPS crowd is the most obnoxious demographic in developer infrastructure, and anyone who has listened this deep into the episode and still insists on bargain VPS hosting 'knows better.' Theo frames his frustration theoretically: choosing a $5 VPS encodes two statements at once — that your time isn't valuable and that your product isn't valuable — making it an unusually efficient filter for people he doesn't want to work with. The segment closes on a lighter note with a story about Cascadia — where the conference organizers had bottles of whiskey backstage and were enthusiastically offering pre-talk shots. Theo declined, Ben notes that Theo has 'one of the most obvious Ballmer curves of anyone I've ever met,' and both agree the talk turned out fine without the liquid courage.
The FUD around Fable being 'nerfed' after its return gets a systematic takedown. Ben notes that public benchmarks comparing Fable before and after the export ban are mostly hitting prompts that now trigger rerouting to Opus — the model itself hasn't changed. Theo then digs into the actual architecture behind rerouting, drawing on an Anthropic paper he'd reviewed that day. [1] — Theo "Anthropic's rerouting system has two stages: a cheap first-stage classifier (~0.05% compute) that watches which expert pathways activate in…" 1:02:50 The system is two-stage: a first-stage classifier monitors which expert pathways activate during inference (essentially watching the direction of the weights) and costs roughly 0.05% of full inference compute. When it flags a request, a far more expensive secondary classifier fires — one that adds 27–80% to the compute cost per request. By building the cheap first-stage gate, Anthropic can catch the vast majority of concerning requests without paying the full secondary cost on every token. The combined system has driven jailbreak pass rates from 95% down to under 0.5%, though Theo notes that since all of this runs on every token of output, there's an interesting lower bound related to how many tokens it takes to encode any given piece of dangerous information. Both hosts agree rerouting is annoying but rare in practical coding use — and recommend just sending a normal follow-up message to exit Opus and return to Fable.
The Sonnet 5 vs. GLM-5.2 debate is more nuanced than the episode's recurring cardboard-taste jokes suggest. Ben's actual defense of Sonnet 5 is narrow and specific: during the window when Fable was banned, it was the only model that could handle very long-running agentic loops requiring recursive self-calling, review subagents, and staged PR workflows. GLM-5.2 is fast, cheap, and produces clean code that a smarter orchestrator can command easily — and both hosts now prefer it in most subagent slots. [1] — Ben "GLM-5.2 is fast, cheap, and produces clean code easily controlled by a smarter orchestrator. Sonnet 5 edges it out only in one scenario: ve…" 1:10:10 The broader complaint about people equating GLM-5.2 with Sonnet 5 for capability is what Ben was actually arguing against. To punctuate the Sonnet 5 discourse, Theo issues a live challenge: have Sonnet 5 order a custom t-shirt reading 'Sonnet 5 made this shirt' via Codex computer use. The model navigates a custom apparel website, auto-populates the right zip code (alarming both hosts), and nearly completes the order — though the shirt design comes out misaligned over the right breast rather than centered. The episode closes on this chaotic note, with Ben threatening to 'slightly inconvenience' the Trump administration if 5.6 isn't out by next week.
Chapter 2 · 00:49
Theo opens the news segment with a primer on forward-deployed engineers — a role that has existed for years but is suddenly central to enterprise AI adoption. The job is simple: embed at a client company, focus on 3 to 5 people per month, and physically make them use the product. [1] — Theo "Microsoft $2.5B FDE investment: Microsoft invested $2.5 billion in hiring approximately 6,000 forward-deployed engineering experts to help …" 01:29 Microsoft has taken this playbook to an extreme, committing $2.5 billion and 6,000 FDE hires to accelerate adoption across its enterprise customer base. Ben frames it bluntly as 'consulting in 2026,' noting that even attendees at the AI Engineer Conference — a self-selected technically sophisticated crowd — had barely scratched the surface of what current models can do. The conversation pivots to a harder question: who should companies hire for these roles? Theo argues strongly that the answer is AI-native young engineers, not experienced developers learning AI retroactively, because the former grew up with the tools the same way today's kids grew up without knowing file systems. He closes with a pointed plea: don't pull the ladder up behind you — teach the younger generation about how old systems worked instead of trying to reskill a 32-year-old on Claude.
Microsoft is spending $2.5 billion to hire 6,000 forward-deployed engineers — essentially consultants whose job is to show up at companies and force AI adoption. The real insight: FDEs don't work because they're eloquent; they work because they lock in 3–5 people per month and physically make them use the product.
Microsoft invested $2.5 billion in hiring approximately 6,000 forward-deployed engineering experts to help companies adopt AI.
Theo argued that effective forward-deployed engineering works because FDEs fixate on 3 to 5 people per month at a target company, forcing hands-on adoption.
Not adopting AI isn't just slow — it's fatal. Unlike moving from COBOL to Go (a 2–3x output gain you could offset by hiring), AI creates a gap so large that a company without it simply cannot compete with one that has it. The difference isn't just performance; it's existential.
The AI Engineer Conference in San Francisco this year drew around 7,000 attendees over 4 days, a massive jump from an estimated 1,000–2,000 the prior year.
Chapter 3 · 10:30
The AI Engineer Conference in San Francisco leveled up considerably this year, ballooning to roughly 7,000 attendees from an estimated 1,000–2,000 the prior year — a growth both hosts describe as a tier jump. Ben mostly skipped the talks in favor of hallway conversations, but made an exception for Theo's, which he calls Theo's second-best talk overall after the Elixir Conf one. Theo had agreed to speak just a day or two before the event and was aiming for a million-view breakout moment, inspired by Matt Pocock having achieved that twice. The meetup story is a highlight: Theo put out a spontaneous tweet asking if anyone would host a meet-and-greet, OpenAI stepped up and offered their booth, and the turnout far exceeded expectations. For audience members who couldn't get through the crowd, cornering Ben and collaborator Julius in the back apparently worked just as well.
The fundamental problem with letting AI write your Claude Code skills is that models write what they already do by default. Skills are supposed to steer the weights toward behaviors the model doesn't naturally exhibit. An AI generating a skill for itself is writing a no-op.
Chapter 4 · 14:20
Matt Pocock's missed AIE talk — posted online instead — becomes the launching pad for one of the episode's most technically dense segments. His core argument: don't let AI write your Claude Code skills. The reason is elegant and non-obvious. [1] — Theo "The fundamental problem with letting AI write your Claude Code skills is that models write what they already do by default. Skills are supp…" 14:00 The entire purpose of a skill is to give the model behavior it doesn't have by default. When the AI writes the skill itself, it naturally gravitates toward behaviors it already exhibits, creating elaborate no-ops stuffed with filler like 'be really thorough' — instructions the model would follow anyway. Ben admits he's guilty of this pattern himself. The right workflow, Theo argues, is to go through your full model usage history, find every time you had to correct the model, categorize those failures, and write skills that specifically redirect the model away from those documented failures. Both hosts then open their .clod directories live and compare skill counts — Ben has 8 in his global directory and over 100 in his Hermes agent (after pruning 50), while Theo has trimmed aggressively to just a handful. The discussion covers workflow skills, framework-specific skills (Svelte, Effect), and the important distinction between things that belong in Agents.md versus a full skill.
The best skills come from analyzing your full usage history, finding repeated correction patterns, and writing guidance that steers the model away from documented failures. Writing a skill once and never revisiting it is almost always wrong — treat skills like living documents.
Ben's Hermes personal agent had well over 100 skills even after he deleted about 50, illustrating how quickly skill files accumulate.
Chapter 5 · 24:10
The Alibaba distillation story is a perfect illustration of how frontier model data gets exfiltrated at scale. Theo explains the alleged mechanism: Alibaba — or more likely a third-party broker — set up discounted Claude Code subscriptions targeted at users in regions where Anthropic accounts are hard to obtain, then routed all traffic through a forwarding proxy that captured every input and output. The result was 28.8 million Claude exchanges across 25,000 fraudulent accounts. [1] — Theo "Alibaba: 28.8M exchanges stolen: Alibaba allegedly scraped 28.8 million Claude exchanges via 25,000 fraudulent accounts to use as model tra…" 25:33 Anthropic's response was a geographic detection code block embedded in Claude Code that checked for signals like Shanghai-based logins — flagged publicly by a user as 'hidden spyware,' then acknowledged by Anthropic's Thorik as a March experiment that had already been superseded by better mitigations. Theo connects this directly to the ViProxy proxy tool he and Ben use to balance accounts, predicting it will be killed within 2–3 weeks once Anthropic adds cryptographic session signatures to Claude Code's proprietary OAuth endpoint. Ben notes that the proxy model is only routing to Claude Code — not to other harnesses — but acknowledges the risk is real.
Alibaba allegedly acquired 28.8 million Claude exchanges via 25,000 fraudulent accounts — likely by offering subsidized Claude Code subscriptions that route all traffic through a data-collecting forwarder. It's not necessarily Alibaba doing it directly; third-party brokers may sell the data to them.
Alibaba allegedly scraped 28.8 million Claude exchanges via 25,000 fraudulent accounts to use as model training data.
Tools like ViProxy let users load-balance multiple Claude Code accounts to maximize access to Fable during its limited window. But Theo predicts Anthropic will kill this method in 2–3 weeks by requiring a rolling cryptographic signature in Claude Code's special OAuth endpoint — and Alibaba is to blame for accelerating that.
Chapter 6 · 31:10
Ben uses the PostHog ad to illustrate a concept he and Theo have been discussing more lately: self-driving product loops. PostHog already collects product analytics signals, groups them into reports, and then — through an agent — opens PRs to your actual codebase to fix the underlying issues. After you merge, PostHog measures the impact and loops back. Ben highlights that PostHog is uniquely positioned for this because they already have all the signal data, and praises both their aggressive price-decreasing track record and their MCP server as tools for teams that want to build custom versions of this loop.
OpenAI offering the Trump administration a 5% stake is the most performative move Theo has ever seen — but also probably the right one. The calculation is simple: make the government feel important enough to let the labs work, and buy time until people who actually understand AI can build real oversight systems.
OpenAI proposed handing the Trump administration a 5% stake in the company, which Theo likened to the US government's 10% stake in Intel.
Chapter 7 · 32:20
The proposed 5% OpenAI stake for the Trump administration sparks one of the episode's most pointed political tangents. Theo opens with a personal message to Sam Altman thanking him for doing what nobody else had the courage to do — publicly and aggressively courting the administration in order to preserve AI progress. The move is immediately called out as performative theater, compared unfavorably to Pride Month rainbow profile pictures. But Theo argues it's the necessary theater of the moment: make the government feel like they control the decision-making process long enough for actual AI safety infrastructure to be built, then hand that infrastructure to them as a fait accompli. [1] — Theo "5% of OpenAI ÷ 350M citizens = ~$143/person: If OpenAI hits a $1 trillion valuation, a 5% government stake divided among 350 million US cit…" 38:32 The math lands as a punchline: at a $1 trillion valuation, 5% divided among 350 million Americans equals $143 — less than one month of ChatGPT Pro. Theo also notes the delicious irony that if citizen equity is ever distributed, Sam Altman — who famously holds no equity in OpenAI — would finally have the same stake in the company as every American, including Theo and Ben.
If OpenAI hits a $1 trillion valuation, a 5% government stake divided among 350 million US citizens yields roughly $143 per person — less than one month of a ChatGPT Pro subscription.
Chapter 8 · 39:50
With Fable back in their hands, both hosts have been running it hard. Ben's first major use case was rebuilding the brittle CLI that spun up his Hermes agent container — a task involving Hermes, Codex, Claude Code, and an Executor, all needing to live in a single deployable box. What Opus had built barely worked; Fable rebuilt the entire architecture in about an hour of back-and-forth. [1] — Theo "Theo gave Fable his entire backlog of stale, conflicted, and draft PRs and told it to figure out which ones to keep, merge, rebase, or kill…" 41:40 Theo's story is more cinematic: he had 18-plus stale pull requests sitting in his Lakebed project, some finished but unrebased, others superseded by later merges, a few just outright slop. Rather than manually triaging them, he asked Fable to read all the PRs, figure out what they were for, identify overlaps, write plans for anything worth salvaging, and then execute. He read the plans briefly, said 'YOLO — slash goal, complete all the work we talked about,' and watched the model close a dozen PRs and land 14 more in one loop. The thing that makes this work, Theo emphasizes, is that Fable can pull intent from incomplete context — it doesn't need you to re-explain what you were trying to do; it can figure it out.
Theo gave Fable his entire backlog of stale, conflicted, and draft PRs and told it to figure out which ones to keep, merge, rebase, or kill. Without needing step-by-step instructions, the model categorized overlapping work, wrote plans for missing features, and landed or closed over 30 PRs in a single 5-hour loop — for about $150–$200.
Theo spent roughly $150–$200 of Fable API usage to review, close, rebase, or land over 30 pull requests in a 5-hour agentic loop.
Chapter 9 · 47:17
Theo's conspiracy theory for the episode is arguably his best. The 7-day, half-rate-limit Fable window looks like a consolation prize for the 14-day window that was cut short by the export ban. But Theo proposes a different read: Anthropic needed exactly one full weekly cycle of subscriber usage data to understand usage patterns — how hard people burn on weekdays versus weekends, during working hours versus late nights — so they can provision GPU capacity between enterprise customers and subscribers. You can't get that data from 3 days, and you don't need 14. [1] — Theo "The 7-day Fable window at half rate limits wasn't a consolation prize — it was a calculated GPU provisioning experiment. Anthropic needed e…" 47:17 The FOMO built up during the export ban was a feature, not a bug: by announcing access the day before it went live, Anthropic ensured every subscriber who was going to burn would be ready to go on day one, giving them the clearest possible signal on peak usage. And the half rate limit isn't cruelty — it's Anthropic not having the spare compute to give full access while enterprises are still onboarding and claiming their allocations. Ben's counter — that they also get training data out of the deal — is gently corrected: Theo notes that most power users have training data opt-out enabled, and the provisioning data is what actually matters.
The 7-day Fable window at half rate limits wasn't a consolation prize — it was a calculated GPU provisioning experiment. Anthropic needed exactly one full week of data to see usage rates across every day of the week so they can plan how to allocate compute between enterprises and subscribers. The FOMO from the export ban was also a feature, not a bug.
Instead of the planned 14-day Fable trial window, users got 3 days before the export ban, and then 7 days at half the rate limit — effectively ~7.5 days total vs. the promised 14.
Chapter 10 · 54:50
The pricing story around Fable is revealing. It launched at roughly $125 per million output tokens, then dropped to $50 — still 30–40% more expensive than Opus on a per-token basis, but Fable uses significantly fewer tokens per equivalent task. That means Fable often ends up cheaper than Opus in practice, which Theo argues is the death knell for Opus as a standalone model. [1] — Theo "Fable dropped from $125 to $50 per million output tokens, making it 30–40% pricier than Opus per token — but far cheaper overall because it…" 54:50 He's already made peace with paying API prices once the subscription window closes: he spent $150–$200 closing over 30 PRs in a single 5-hour session, and frames it as a magic button he'd press twice a day if he could. Ben agrees that any serious company would make the same calculation. The conversation turns to whether Opus 5 can survive in a world where Fable is this affordable — both hosts think Opus 5 needs to be priced lower than current Opus just to remain relevant.
Fable dropped from $125 to $50 per million output tokens, making it 30–40% pricier than Opus per token — but far cheaper overall because it uses fewer tokens. The result: Opus is dead. Theo won't maintain a $200/month sub for Opus alone, and thinks Opus 5 needs to be cheaper or it won't survive either.
Ben speculated that Fable (Mythos) is likely a 5 to 10 trillion parameter model, making it the largest publicly accessible frontier model by a wide margin.
Fable (Mythos) dropped from its originally announced ~$125 per million output tokens to ~$50, making it 30–40% more expensive than Opus but often cheaper due to lower token usage.
Running Fable subagents from a Fable orchestrator is a fast path to burning 30% of your weekly limit on nothing useful. The right pattern is Fable commanding cheaper models — Sonnet 5, GLM-5.2, or similar — for implementation work while Fable handles planning, context, and review.
Ben accidentally ran Fable subagents on xHigh reasoning, burning through roughly 30% of a weekly usage allocation with negligible useful output.
Chapter 11 · 58:00
Ben opens with a confession: on his first day with Fable back, he let it spawn Fable subagents on xHigh reasoning mode — forgetting to turn it off after prior testing — and burned approximately 30% of a weekly usage allocation with minimal useful output. [1] — Ben "Running Fable subagents from a Fable orchestrator is a fast path to burning 30% of your weekly limit on nothing useful. The right pattern i…" 57:56 Theo introduces a meme to illustrate the lesson: in Claude Code's reasoning effort selector, low/medium/high are labeled 'good options,' xHigh and Max are labeled 'mental illness,' and UltraCode is a 'skill disguised as a reasoning effort' (because Anthropic is using High under the hood). The correct architecture for serious work is Fable orchestrating cheaper models — Sonnet 5, GLM-5.2, or similar — for implementation while Fable handles planning, context synthesis, and review. The reasoning level discussion also touches on the non-obvious fact that spending more tokens and using higher reasoning effort past a certain point actually degrades output quality rather than improving it.
Chapter 13 · 1:02:50
The FUD around Fable being 'nerfed' after its return gets a systematic takedown. Ben notes that public benchmarks comparing Fable before and after the export ban are mostly hitting prompts that now trigger rerouting to Opus — the model itself hasn't changed. Theo then digs into the actual architecture behind rerouting, drawing on an Anthropic paper he'd reviewed that day. [1] — Theo "Anthropic's rerouting system has two stages: a cheap first-stage classifier (~0.05% compute) that watches which expert pathways activate in…" 1:02:50 The system is two-stage: a first-stage classifier monitors which expert pathways activate during inference (essentially watching the direction of the weights) and costs roughly 0.05% of full inference compute. When it flags a request, a far more expensive secondary classifier fires — one that adds 27–80% to the compute cost per request. By building the cheap first-stage gate, Anthropic can catch the vast majority of concerning requests without paying the full secondary cost on every token. The combined system has driven jailbreak pass rates from 95% down to under 0.5%, though Theo notes that since all of this runs on every token of output, there's an interesting lower bound related to how many tokens it takes to encode any given piece of dangerous information. Both hosts agree rerouting is annoying but rare in practical coding use — and recommend just sending a normal follow-up message to exit Opus and return to Fable.
Anthropic's rerouting system has two stages: a cheap first-stage classifier (~0.05% compute) that watches which expert pathways activate in the model, and an expensive second-stage classifier (27–80% extra compute) that only fires when the first flags a request. The result: jailbreak pass rates dropped from 95% to under 0.5%.
Anthropic's secondary safety classifier increases compute per request by 27% to 80%, prompting them to build a cheaper first-stage check costing only ~0.05% of compute.
Anthropic's classifiers reduced jailbreak success from 95% to 4%, and then their latest two-stage system brought it down to under 0.5%.
Chapter 14 · 1:10:10
The Sonnet 5 vs. GLM-5.2 debate is more nuanced than the episode's recurring cardboard-taste jokes suggest. Ben's actual defense of Sonnet 5 is narrow and specific: during the window when Fable was banned, it was the only model that could handle very long-running agentic loops requiring recursive self-calling, review subagents, and staged PR workflows. GLM-5.2 is fast, cheap, and produces clean code that a smarter orchestrator can command easily — and both hosts now prefer it in most subagent slots. [1] — Ben "GLM-5.2 is fast, cheap, and produces clean code easily controlled by a smarter orchestrator. Sonnet 5 edges it out only in one scenario: ve…" 1:10:10 The broader complaint about people equating GLM-5.2 with Sonnet 5 for capability is what Ben was actually arguing against. To punctuate the Sonnet 5 discourse, Theo issues a live challenge: have Sonnet 5 order a custom t-shirt reading 'Sonnet 5 made this shirt' via Codex computer use. The model navigates a custom apparel website, auto-populates the right zip code (alarming both hosts), and nearly completes the order — though the shirt design comes out misaligned over the right breast rather than centered. The episode closes on this chaotic note, with Ben threatening to 'slightly inconvenience' the Trump administration if 5.6 isn't out by next week.
GLM-5.2 is fast, cheap, and produces clean code easily controlled by a smarter orchestrator. Sonnet 5 edges it out only in one scenario: very long-running jobs requiring recursive self-calling and review loops. Outside that, GLM-5.2 wins on price-performance — and even Theo would pick it over Sonnet if he still had Fable.
Theo, mid-episode, gave Sonnet 5 a live task: use Codex computer use to browse the web and order a custom shirt that says 'Sonnet 5 made this shirt.' The model navigated a custom apparel site using browser automation, auto-filled Theo's zip code, and nearly completed the order — getting about as far as anyone expected.
No indexed bits in this chapter.
The 7-day Fable window at half rate limits wasn't a consolation prize — it was a calculated GPU provisioning experiment. Anthropic needed exactly one full week of data to see usage rates across every day of the week so they can plan how to allocate compute between enterprises and subscribers. The FOMO from the export ban was also a feature, not a bug.
Anthropic's rerouting system has two stages: a cheap first-stage classifier (~0.05% compute) that watches which expert pathways activate in the model, and an expensive second-stage classifier (27–80% extra compute) that only fires when the first flags a request. The result: jailbreak pass rates dropped from 95% to under 0.5%.
This episode
Factual claims made this episode, and whether a source was named.
Microsoft invested $2.5 billion in hiring approximately 6,000 industry and engineering experts as forward-deployed engineers.
Alibaba allegedly acquired 28.8 million Claude conversation exchanges via approximately 25,000 fraudulent Anthropic accounts.
Anthropic embedded code in Claude Code that checks for geographic indicators (e.g., Shanghai) to detect unauthorized resellers and prevent distillation.
The AI Engineer Conference in San Francisco had approximately 7,000 attendees in 2026, up from an estimated 1,000–2,000 the prior year.
Anthropic's secondary safety classifier increases compute per request by between 27% and 80%.
Anthropic's first-stage safety classifier costs approximately 0.05% of the compute of a full inference request.
Anthropic's classifiers reduced jailbreak pass rates from 95% to 4%, and the newest two-stage system brought that to under 0.5%.
OpenAI proposed giving the Trump administration a 5% stake in the company.
A 5% stake in OpenAI at a $1 trillion valuation would yield approximately $143 per US citizen across 350 million people.
Fable's API price dropped from approximately $125 per million output tokens to approximately $50 per million output tokens.
Fable's weekly rate limit for Claude Code subscribers is set at 50% of the standard weekly limit during the re-release window.
Claude Code rate limits are currently doubled for all users, a change that took effect in May.
UltraCode reasoning mode in Claude Code uses High-level reasoning under the hood, not a truly ultra-high setting.
This episode
Annual AI engineering conference in San Francisco that drew ~7,000 attendees this year, where Theo spoke and hosted a meetup at OpenAI's booth.
TypeScript educator and content creator whose AIE talk on building competent Claude Code skills is cited extensively; known as the second-best TypeScript YouTuber.
Praised by Theo for his political strategy of offering a 5% OpenAI stake to the Trump administration to preserve AI model access.
Anthropic's CEO, criticized for being good at communicating with AI researchers but poor at political communication with government officials.
Discussed as the maker of Claude and Fable (Mythos), whose export ban, rate limit policies, and safety classifiers are central to the episode.
Discussed in context of the AI Engineer Conference, GPT-5.6 delays, the proposed 5% government stake, and Codex computer use.
Accused of acquiring 28.8 million Claude exchanges via 25,000 fraudulent accounts for model distillation purposes.
Referenced for its $2.5 billion investment in 6,000 forward-deployed engineers to accelerate enterprise AI adoption on Azure.
Episode sponsor; described as an all-in-one product analytics suite with self-driving loop capabilities and a highly regarded MCP server.
Anthropic's agentic coding tool, central to the episode's discussions of skills, rate limits, Fable access, and ViProxy usage.
Debated as an inferior but capable model for long-running agentic tasks; Ben defends it while Theo compares it to an inferior banana.
OpenAI's coding agent, used alongside Claude Code in the hosts' multi-agent setups; also used for computer use tasks in the live t-shirt demo.
Discussed as a fast, cheap model from Alibaba that is strong on price-performance and useful when commanded by a smarter orchestrator like Fable.
Open-source proxy tool that enables load balancing across multiple Claude Code accounts to maximize Fable access; predicted to be shut down by Anthropic.
Location of the AI Engineer Conference and the hosts' base; used as shorthand for the insular tech bubble they operate within.
Stats
We use essential and analytics cookies to run Vuci. To understand how the site is used: Privacy Policy.
Install Vuci on your phone
Add it to your home screen for a faster, app-like experience.
Install Vuci on your phone
Tap the Share button, then “Add to Home Screen”.
A new version is available
Reload to get the latest Vuci.