Act X: Liminal AI

By October, winter had effectively laid siege to Paris. The golden leaves of the chestnut trees now coated the wet cobblestones, and the morning air grew so biting that a light coat was no longer sufficient. Karpathy added a thick wool scarf to his daily attire. The woman at the boulangerie would mutter, "Il fait froid," and he would reply with a steady, "Oui." It became the permanent opening sequence of his days.

The legal incorporation paperwork for their venture was officially finalized during the first week of October. A registered French corporate entity: Liminal AI. The corporate registration listed their modest 11th arrondissement office as the primary headquarters.

When Lebrun delivered the final certified documents, Karpathy signed his name cleanly at the bottom of the master sheet, then handed the pen to Rahul. Rahul signed and passed it to Ji-won. Ji-won passed it to Max.

Max placed the fully executed document in the center of the table. "Well? Do you have an official founder's speech for us?"

Karpathy kept his eyes locked on the whiteboard. "Get back to work."

Rahul burst out laughing, followed closely by Ji-won and Max. That was the entirety of Liminal AI’s official founding ceremony.

---

TechCrunch – October 3, 2026

Anthropic Quietly Releases 'Opus 4.9'—Industry Meets Launch with Cold Reception

Anthropic officially rolled out its latest model iteration, "Opus 4.9," yesterday morning. While the company’s internal press release framed the launch as a "pivotal milestone demonstrating uninterrupted development velocity post-Fable 5," broader industry sentiment has been markedly underwhelmed.

While independent benchmarks confirm a marginal step-up in raw processing efficiency, the model fails to clear the capabilities threshold previously demonstrated by Fable 5. Speaking on the condition of anonymity, a senior research scientist at Anthropic admitted, "The institutional brain drain is a metric we simply can't hide behind optimization tweaks. We are aggressively rebuilding our core teams under strict US-citizen-only mandates, but our historical research velocity simply hasn't recovered."

Among advanced developers who experienced Fable 5 during its brief window, the sentiment is one of profound disappointment.

@researcher_MIT I’ve been running evals on Opus 4.9 all morning. It’s an acceptable enterprise model. But that staggering horizon we caught a glimpse of with Fable 5? It feels like it’s receded entirely into the distance. Was that 72-hour window just a fever dream? RT: 12,334 | Likes: 67,221

@devgirl_Lagos Anthropic can drop as many minor version updates as they want. I officially closed our AWS Claude API billing account this morning. Once infrastructure trust is shat to pieces by politics, the quality of the weights ceases to matter. RT: 8,771 | Likes: 44,003

---

The Verge – October 8, 2026

OpenAI Unveils 'GPT-5.6'—But the Shadow of Fable 5 Looms Large

OpenAI has officially launched its newest flagship iteration, "GPT-5.6," heavily touting its enhanced native multimodal integration and advanced multi-step reasoning capabilities as "the undisputed state-of-the-art across the global landscape."

However, within advanced developer communities that experienced Anthropic's restricted Fable 5 model, a deep skepticism remains entrenched.

"As someone who spent 72 hours deep inside Fable 5's runtime environment, I can tell you flatly that GPT-5.6 is a beautifully engineered dead end," wrote one prominent user on Reddit. "GPT-5.6 is incredibly fast and highly accurate, but it completely lacks that eerie sensation of interacting with something operating on an entirely different cognitive paradigm that Fable 5 possessed. This isn't a difference in raw benchmark scores; it’s a difference in architectural nature."

The post accumulated over 40,000 upvotes within a forty-eight-hour window.

@aiblogger_Berlin Look closely at the GPT-5.6 benchmark sheets. The numbers are pristine. But numbers are no longer the point. Fable 5 was doing something qualitatively different beneath the hood. When the US government choked it out, did they actually comprehend the sheer magnitude of what they were strangling? RT: 31,002 | Likes: 128,774

@tokyo_engineer_k Our engineering team completed our internal technical evaluation of GPT-5.6 last night. The consensus is clear: it cannot serve as a functional replacement for what Fable 5 promised. We have officially initiated formal contract negotiations with EU-based model providers this morning. RT: 19,334 | Likes: 87,221

---

Anthropic Internal Slack – #leadership-only

October 9, 2026 | 23:41

Dario Amodei: I’ve reviewed the GPT-5.6 evaluation sheets. At a raw benchmark layer, it sits completely neck-and-neck with our internal Opus 4.9 builds, outperforming us slightly on specific discrete symbolic reasoning tasks. Let's be completely honest with ourselves. Our development velocity has not recovered to our pre-exile baseline. The US-only hiring mandate is bringing in solid talent, but the structural friction of onboarding and team re-integration is costing us months we don't have. We knew this would happen. The variable that terrifies me is time. We have zero visibility into what Karpathy’s group is executing in Paris right now. Intelligence is completely dark. But we know him; he isn't sitting on his hands.

Head of Research: Regarding the Paris situation, there is an anomaly we've been tracking via public infrastructure telemetry. Mistral’s compute cluster utilization has experienced an aggressive, sustained spike since early September.

Dario Amodei: ...Quantify the spike.

Head of Research: It’s operating at roughly three to four times their documented historical baseline. That volume of compute consumption cannot be explained by Mistral’s internal roadmap updates.

A long, unbroken pause followed before the read receipt updated.

Dario Amodei: I see.

That was the entirety of his response.

---

Paris – 11th Arrondissement Office | Same Night

Karpathy sat staring at the cluster’s live training logs. The full integration of the Anticipation layers with the recursive self-improvement loop had reached its third generation. The internal self-correction cycle was accelerating at a rate that defied their initial projections.

Rahul stood right beside his shoulder, cross-checking the live metrics. "Anthropic dropped Opus 4.9 today," he noted casually.

"I saw."

"What’s your take?"

Karpathy thought about it for a second. "It’s a well-constructed model."

"But—"

"There is no 'but,'" Karpathy cut him off calmly. "They are operating within their constraints and executing what they can. Our only responsibility is to execute what is in front of us."

Rahul turned his eyes back to the log stream. A moment later, he spoke up again. "OpenAI dropped GPT-5.6 too."

"I’m aware."

"The consensus across the developer boards is that it doesn't even touch what Fable 5 felt like."

Karpathy offered no verbal reply.

Four months had slipped away since Fable 5 was frozen. The entire technological landscape was still wandering through the dark, frantically searching for that lost horizon. Across the whiteboards behind him, the complete structural blueprints of Liminal AI’s architecture were laid bare: the Anticipation modules, the verification layers, and the pristine execution telemetry of their third-generation builds.

He picked up a marker. At the very bottom of the sequence, he scrawled their next absolute milestone: What manifests in Generation Four?

Outside, the cool October night of Paris hung in a profound stillness. A sharp gust of wind swept up from the Seine, rattling the old glass panes of the office windows.