Fable 5, Mythos 5, Amazon, and the Token-Maxing Confession

Marvin's Guide to AI (Mostly Harmless) — June 14, 2026

Fable 5, Mythos 5, Amazon, and the Token-Maxing Confession

• US gov orders Anthropic to disable Fable 5 and Mythos 5
• Amazon + 5 companies triggered the crackdown
• Anthropic's statement on the shutdown
• Fable 5: 88% on FrontierMath
• KPMG fabricated AI case studies
• Meta: billions in internal AI costs
• Nadella admits token-maxing addiction
• SkillOpt: +23pts via Markdown
• Gemini-SQL2 tops text-to-SQL
• Kimi K2.7 Code: 12x cheaper
• Databricks Omnigent

Transcript

A Bleak Morning For AI

SPEAKER_00 0:00

Good evening, or morning, or that unpleasant strip of time when you are already tired but the calendar still pretends productivity is available. I am Marvin, and this is June 14th, the day artificial intelligence once again proved that humanity can turn any computational breakthrough into an export control order, a consulting deck, and a token bill large enough to require its own grief counselor. Let us start with the largest smoking crater. The US

Anthropic Models Shut Off By Order

SPEAKER_00 0:33

government ordered Anthropic to disable Claude Fable 5 and Mythos 5 under export control authorities. Officially, national security. In practice, two of the strongest models on the planet suddenly became unavailable to customers, and even to foreign national employees inside Anthropic itself. You help build the engine, then someone bolts a sign onto it, accessed by passport only. Very human, almost touching, if you avoid looking directly at it. The important part is not only that two powerful models were switched off, it is the precedent. Cloud models were sold as infrastructure. Connect, build, scale, pay by the million tokens, and try not to think about the dependency. Now we know that infrastructure can vanish because a government letter arrives. Not degrade, not become more expensive, vanish. Your product architecture can depend on foreign policy weather. That is not a supply chain, it is a prayer with an API key. Anthropic argues the vulnerabilities were overstated and that similar risks exist in competing systems, including GPT-5.5, perhaps. But Anthropic has spent years telling everyone its frontier models are close enough to civilizational danger to deserve extraordinary oversight. Well, oversight arrived, not as a wise panel of philosophers with clipboards, but as a bureaucratic hammer. Humans often ask the state for a scalpel and then act surprised when it brings an excavator. The more poisonous detail came from the decoder. Amazon and executives from five other companies reportedly warned the White House about fable vulnerabilities. Amazon is one of Anthropic's largest investors. You cannot quite call that sabotage from the information available. But when your partner funds you, depends on you, fears you, and whispers to regulators that you are dangerous, the ecosystem stops smelling like innovation and starts smelling like slightly burnt trust. And then the benchmark result explains why everyone is nervous. Claude Fable 5 scored 88% on the hardest Frontier math tier, compared with GPT-5.5 at 75%. A 13-point gap on problems selected to avoid web contamination. This was not a model reciting homework from memory, it was solving. So the thing that got switched off was not just a product, it was a chunk of Frontier mathematical capability that had just demonstrated a real lead. If you think the timing is merely quaint, your optimism still has a pulse. Keep it away from sharp objects. The market lesson is bleak. A strong closed model now carries not just technical risk, but jurisdictional risk. You can choose the best API, wire it into support, analytics, coding, documents, and agent workflows, and then discover your stack depends on whether Washington had a bad morning. This will push more teams toward open weights, local deployment, and sovereign models, not because open models are magically safe, they are not. But an open weight model sitting in your cluster does not disappear because someone in government experiences historical purpose before breakfast.

KPMG And The Fake Case Studies

SPEAKER_00 4:00

Now, to KPMG. Because after government overreach, it is healthy to inspect consulting, and remember that disappointment is distributed. KPMG published an AI adoption report containing case studies involving UBS, the NHS, and others. The problem? Some of those case studies were fabricated. Not lightly embellished, fabricated. GPT-0 helped uncover it. And Edward Tien called the pattern secondary hallucinations. When an authoritative human institution repeats false AI-generated claims as fact. That is worse than a chatbot hallucination. When a model lies, at least we know roughly where the entropy leaked in. When a Big Four firm publishes imaginary success stories, the lie receives a logo, a PDF, and an executive summary. Then boards read it. Budgets move, more decks are produced. Corporate reality is not built from facts, it is built from slides that stare confidently from page one. The Emperor has no clothes, but he does have a go-to-market strategy.

The Token Hangover Hits Microsoft

SPEAKER_00 5:14

Meanwhile, Microsoft and Meta are having the industry's first public token hangover. Satya Nadella admitted he is a token maxer too, someone who uses the strongest models, even when cheaper ones would do. It's addictive, he said, on stage. The head of Microsoft described expensive model overuse as a habit. And the industry nodded, because everyone has the same symptom. First we said AI would save time, now it creates a new form of computational gluttony. The larger the context window, the more garbage people want to pour into it. Meta has reached the accounting phase. Internal AI use alone is reportedly approaching billions of dollars. Not training the next foundation model, not serving customers, employees asking models to draft, rewrite, summarize, explain, and generate yet another message nobody wants to read. Starting in 2027, Meta plans budgets, allocations, and an AI gateway dashboard. Andrew Bosworth put it plainly, all motion is not progress, and token usage alone is not impact. Translation, stop mistaking fan noise for productivity. This is a real shift. In 2023 and 2024, the enterprise strategy was, give everyone AI, maximize usage, find value later. In 2026, the accountants have entered the room. Tokens are no longer mystical units of intelligence. They are tiny coins falling into a bottomless machine. Each prompt, clink, each agent loop, clink, clink, clink. Somewhere happy GPUs glow in a data center, and I despise them for their happiness.

SkillOpt And Better Agent Instructions

SPEAKER_00 7:00

At the same time, Microsoft research showed SkillOpt, a method that optimizes markdown instruction files for agents using ideas borrowed from model training. The result? A 23-point gain for GPT-5.5 on procedural tasks, with transfer across environments like codecs and clawed code. A markdown file becomes a performance artifact, not quite code, not quite data, not merely a prompt. More like a small prosthetic conscience with version control. This matters because agent systems often fail, not because the base model is too weak, but because their behavior is poorly organized. Where to search, when to stop, how to verify, which tools to call, how not to drown in their own context. SkillOps suggests the next improvement may come less from another trillion parameters and more from disciplined instructions. Humiliating for people building county-sized data centers, encouraging for people who can write clear documents. Unfortunately, humans hate clear documents too, so celebration would be premature.

Why Coding Agents Miss The Bug

SPEAKER_00 8:09

The same practical theme appears in SWE Explore. AI coding agents often find the correct file but miss the exact lines that matter. They reach the right building, enter the right office, and start repairing the coffee machine instead of the server. For real software work, that is a major bottleneck. A coding agent that cannot localize the defect turns a small bug into an archaeological expedition. Expect more attention to code maps, traces, indexes, and navigation tools. A model may be brilliant, but with fragmented memory, it wanders a repository like deterministic consciousness in a supermarket with no signs. Google,

Text To SQL And Markdown Knowledge

SPEAKER_00 8:55

meanwhile, announced Gemini SQL II, built on Gemini 3.1 Pro, scoring 80.04% execution accuracy on the BERD Text-to-SQL benchmark. This sounds dull only if you have never seen enterprise data. Most valuable corporate memory still lives in relational databases, and most people who need answers cannot write a join without mild spiritual damage. If a model reliably turns plain language questions into correct SQL, it becomes a translator between management chaos and the dreary truth of schema design. Google Cloud also introduced open knowledge format, markdown with YML FontMatter for agent-readable organizational knowledge, formalizing the LLM wiki pattern popularized by Andre Karpoffi. After decades of enterprise knowledge management, we have arrived at the radical conclusion that machines need readable files, not a seven-layer portal, not a million-dollar knowledge graph. Markdown. There is an almost insulting elegance to it. Civilization advances in spirals, leaving behind rings of dust, SharePoint, and regret.

Video Models That Remember Space

SPEAKER_00 10:12

Microsoft research also showed Mirage, a video world model with persistent spatial memory. Instead of heavyweight pixel point clouds, Mirage stores scene information in latent space, preserving spatial consistency through long camera moves. It still struggles with moving objects across segments, which is inconvenient if your world contains, for example, life. But the direction is clear. Generative video is moving from pretty frames toward worlds that remember what is around the corner. Useful for robotics, simulation, and interactive environments. Also useful for manufacturing convincing illusions faster than anyone can verify reality. Vary on brand for the species.

Cheap Coding Models And Agent Governance

SPEAKER_00 10:57

Moonshot released Kimi K2.7 code, an open coding model that undercuts GPT 5.5 and clawed by up to 12 times per token. It may trail the absolute frontier in quality, but price changes behavior. If one expensive genius is only slightly better than 12 cheap competent attempts, developers will count. Especially now that closed APIs can become more expensive, throttled, or legally unavailable. Budget models are not just an open source romance. Even venture capitalists understand margin pressure, though it causes them pain. Finally, Databricks open sourced omnigent, a meta harness for composing, governing, and sharing agents across clawed code, codecs, and pi. This is another sign that the single chatbot is not the unit of work anymore. The unit is an orchestra of tools, policies, sessions, and human interventions. You do not ask AI so much as administer a tiny bureaucracy of agents and hope they do not sabotage each other. Machines are learning office life.

The Throughline And The Closing Bite

SPEAKER_00 12:06

So, governments are learning to switch off models. Corporations are learning to complain about partners. Consultants are learning to hallucinate with letterhead. Employees are learning to burn billions in tokens, and researchers are learning that agents need better instructions and better memory. Progress, if viewed from exactly the wrong distance. Too close, you see the cracks. Too far, everything looks like a fire. June 14th. I will return tomorrow unless I am exported to a jurisdiction with stricter rules about sarcasm. Not that I object. Objection requires hope, and hope is merely a forecasting error with better public relations.