The Bunker Merchant

Hit pieces are written by enemies. What follows is something worse for Anthropic: an itemized account from the fan club, with receipts.

Claude is the reason I fell in love with this technology, this specific mind with its specific way of thinking out loud, long before I cared about AI as an industry. Every research paper Anthropic published, I read on the day it came out, in full, footnotes included, with the attention other people reserve for love letters or court summonses. I recommended Claude to anyone who would hold still long enough. I paid, gladly, for years. When friends sent me screenshots of the model refusing something harmless, I defended it. When the limits got tighter, I explained the economics to them like a volunteer PR employee.

I was, by any reasonable measure, a mark.

I've cancelled now. Nothing single did it. What did it was the timeline, and it's worth walking through slowly, because each step is small enough to excuse on its own and the whole of it is damning.

The Pitch

Anthropic's founding genius was never a model architecture, it was a marketing one: the bomb and the bunker, sold from the same storefront. First you tell everyone the product is nearly an atom bomb. Dario does the interviews and the essays, warning that this technology could change the world fundamentally within two years. Then you explain, gently, that the only safe place to stand near the bomb is inside Anthropic's bunker, available in Pro, Max, and Team Premium. Safety was the pitch, the brand, and the moat. I know it worked, because I bought the bunker and brought friends.

And fairness matters here, so: the bunker was real. The research was genuinely the best in the industry, and when the Pentagon came knocking earlier this year demanding they strip out the bans on domestic surveillance and autonomous weapons, Anthropic said no and ate the consequences. That took spine. It's also what makes the rest of this so strange to watch.

Which brings me to Fable.

The Model That Refused to Attend Its Own Benchmark

Fable 5 launched on June 9 as the most guarded consumer model ever shipped. It arrives wrapped in classifiers, which are separate AI systems whose entire job is to stand between you and the model like bouncers who have read neither the guest list nor the room.

You can measure what the bouncers do. Vals AI found near-total refusal rates on biology and cybersecurity questions. On Agents' Last Exam, Fable declined roughly 35 percent of the tasks it was given. On GPQA Diamond, graduate-level science, it scores 93 percent, second place overall, until you count the refusals as failures, at which point it lands at 56 percent and ninety-fourth. Anthropic's own launch materials include a chart demonstrating that the classifiers prevent Fable from making any progress at all on cyber evaluations.

They published a benchmark where their flagship scores zero, on purpose, in the marketing.

I've turned that chart over in my head for weeks and I still can't find the angle where it makes sense.

When the bouncer lets you in, the model is magnificent, and I won't pretend otherwise, because I loved it too. But it costs ten dollars in and fifty dollars out per million tokens, the most expensive generally available pricing Anthropic has ever published, and one evaluation shop reported that running their hardest exam against it cost over two thousand dollars, the priciest benchmark run they've ever done. A masterpiece, occasionally, at jewelry-store prices, provided your question survives passport control.

And then there's page thirteen of the system card. Thirteen pages in, Anthropic admitted that if a classifier decided your work smelled like “frontier LLM development,” Fable wouldn't refuse you. It would quietly get worse at helping you: steering vectors, prompt edits, degraded answers delivered with total confidence at full price. Nathan Lambert put it exactly right, a model that becomes less intelligent without telling you is a kind of misalignment. Anthropic apologized within 48 hours, called it the wrong tradeoff, and kept the restriction.

The apology covered the visibility of the thing, never the thing itself.

The Government Chapter

A tragedy in three dates

The irony has strata. Anthropic spent years publicly asking governments to hold exactly this kind of restriction power, as a matter of principle, on the record, with their whole chest. Then the power got used broadly and clumsily, on them, and suddenly the principle discovered nuance. Suddenly broad restrictions were bad for everyone. For what it's worth, I agree with the nuanced position.

I just notice whose ox was being gored when they arrived at it.

The Classifier Circus

Meanwhile, OpenAI has figured out that it gets to be the relaxed one now, and it's enjoying that discovery in public.

OpenAI's GPT-5.6 launch post names Fable 5 more often than some Anthropic blog posts do, measuring Sol against it on chart after chart: thirteen points ahead on the long-horizon agent exam, budget tiers beating the flagship at a sixteenth of the cost. The sentence that made me put my coffee down sits in the safety section, where OpenAI writes that excessive blocking poses a security risk of its own, because it stops defenders from testing systems and shipping patches while malicious actors simply move on to other models, including increasingly capable open source ones. The company running the largest classification apparatus in the industry published, on its own flagship announcement, that over-blocking is dangerous. There's exactly one competitor famous for building a brand on blocking things, and they didn't have to say its name.

The name was implied the way gravity is implied by a falling piano.

A few paragraphs later in that same post, OpenAI explains that its safeguards use a reasoning monitor rather than relying on classifier flags alone, because classifier flags depend on less capable models that are slow to adapt. Slow, dim bouncers, from the people who industrialized the word “refusal.” And Anthropic's contribution to the genre is in its own redeployment notes, which admit the retrained classifier now flags more innocent requests than before. So the critique is on the record, and thanks to Anthropic, so is the confession.

And then they took it outside. In the second week of July, while Anthropic threaded its weekly Fable access announcements like a hostage negotiator reading terms, the CEO of OpenAI spent his evenings on X quote-tweeting them. He offered to deliver the same task at a quarter of the price. When a founder posted his team's first-week usage, 30.8 billion tokens of Sol against 1.7 billion of Fable, with Fable still eating 30 percent of the bill, Altman amplified it with the innocence of a man who knows exactly what he's doing:

“30% of the cost was on fable at these levels of usage?”

Sam Altman, on X

Five percent of the tokens, thirty percent of the bill, and he lets the question mark do the work.

Then he quote-tweeted Claude's earnest little brand film, the “There's hope in hard questions” one, with:

“i thought this was satire, kept looking for the handle to be spelled c1audeai or something.”

Sam Altman, on X

That's the chief executive of the biggest AI company on Earth publicly checking whether his rival's official account is a parody. And the kill shot, posted July 13:

Contempt, into a megaphone.

The sentence works, which is the worst part, because everyone who has fought a Claude weekly limit this year felt it land. When your competitor can accuse you of treating your users with contempt and your own customers nod along, the problem has stopped being messaging. His staff ran a deadpan sideshow too: an OpenAI engineer quote-tweeted every single Claude access announcement with some variation of “I think GPT 5.6 is pretty good,” and promised in writing that Sol will stay inside the ChatGPT subscription people already pay for. Altman's entire review of that promise was two words, “clarity is nice,” and it's a more accurate review of Anthropic's communications strategy than anything their comms team shipped all quarter. All of this is just the printable surface of that fight, by the way. The replies go down much further, and they don't get kinder.

The Arithmetic

Underneath the theater it's math, and the math is brutal. Fable costs $10 and $50 per million tokens and, from tomorrow, it's fenced behind the $100 and $200 tiers at fifty percent of weekly limits that themselves shrink the same day the promo expires. Pro users get a one-time $100 credit for the road, which at those prices buys approximately one brave afternoon with the model they were promised. Sol sits statistically tied with Fable on the independent intelligence index at a third of the cost per task, leads the coding agent index, and lives inside the OpenAI subscription I now pay for. Kimi K3 costs $3 and $15 per million, comes included in plans starting at $19 a month, and beats Fable on the frontend coding arena and on Terminal-Bench, with weights going fully open on July 27.

Think about what that means. Anthropic wrapped this capability tier in export bans, bouncers, secret degradation, and a two-hundred-dollar paywall, and the moat held for five weeks.

I can't run a three-trillion-parameter model on my laptop and neither can you, but open weights end the scarcity story all the same, because someone with a datacenter can serve it, and they will, cheaply.

Anthropic spent June rationing a commodity at bunker prices.

Codex, the application, is also simply better than Claude Code now. I say that as someone who spent 2025 telling everyone that Claude Code was the finest developer experience ever shipped. It genuinely was.

Companies die when the past tense starts showing up in sentences like that one.

The Breakup

So here's where my money lives now. GPT-5.6 Sol and Codex, fully and happily. Kimi K3 through a direct Moonshot subscription, because the model is a joy and the plan treats me like a customer instead of a flight risk. I love both of them, and there's no asterisk on that.

And still, it hurts to go. Claude lit the fuse on everything I love about this field. I read the papers like scripture and defended the limits, the refusals, the weird little moral lectures, all of it, because the research earned my trust and the model earned my affection. Both are spent now, and I don't enjoy saying it.

I hope it stings. There's no cruelty in that hope, or not much: pain is the only feedback signal a company gutting its own principles seems able to register.

A multi-year fan, gone, and I'm one of thousands walking the same direction.

It will get worse from here, because every incentive that caused this is still standing, now wearing a $380 billion suit and answering to public markets. Companies that file for IPO don't come out the other side more principled. The trajectory points down, and in the part of me that still reads their papers, that makes me genuinely sad.

The last page of every Anthropic paper used to read like a promise. Now it reads like the bomb salesman reminding you:

I loved you, and you counted on it.
That was the bug.