Anthropic Shipped a New Tier Above Opus. The US Government Erased It in 72 Hours.

Hey guys, Mr. Technology here.

I have covered a lot of LLM launches. I have never covered one that lasted three days. On June 9, 2026, Anthropic released Claude Fable 5 — the first generally available Mythos-class model, a tier that sits above Opus, with state-of-the-art scores on nearly every public benchmark. On June 12, the US government, citing national security authorities, ordered Anthropic to suspend access for every foreign national on Earth, including its own foreign national employees. Anthropic complied. Fable 5 went dark. The most capable model of 2026 lived for 72 hours.

The Spec Sheet

• Model class: Mythos-class (new tier above Opus)
• API model IDs: claude-fable-5 (public); claude-mythos-5 (Project Glasswing only)
• Context / output: 1,000,000 input / 128,000 output tokens
• Pricing: $10 / MTok input, $50 / MTok output, $1 / MTok cache hit, batch 50% off
• Thinking: adaptive thinking always-on, raw chain-of-thought never returned
• Data retention: 30 days, no zero-data-retention (Covered Model)
• Platforms: Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, Microsoft Foundry
• Fallback model for flagged requests: Claude Opus 4.8

Anthropic has not disclosed parameter count, training compute, architecture, or training data. Every "Fable 5 has X trillion parameters" claim you read this week is speculation. What we know is the price, the context window, the benchmark scores, and the live behavior. That is enough to talk about it honestly.

The Benchmarks (The Headline)

A 6.4-point jump on SWE-Bench Verified in one revision is not normal. That is a generation gap. For the first time, an Anthropic model is not just competitive with GPT-5.5 on agentic coding — it is well past it. Stripe reports a 50-million-line Ruby codebase migrated in a day that would have taken a team months. Long-horizon agentic work is no longer a benchmark category. It is the new baseline.

What It Can Actually Do

Pokémon FireRed, vision-only. Earlier Claude models needed a complex helper harness with maps and game-state tools. Fable 5 beat FireRed with raw screenshots and a minimal harness — a complete collapse of the "models can't do long-context vision-grounded planning" argument.

Drug design, 10x acceleration. Anthropic's internal protein-design team reports Mythos 5 accelerated parts of the drug-design pipeline by roughly 10x. 9 of 14 protein targets yielded strong candidates now in experimental follow-up. Not a benchmark — a wet-lab pipeline.

**Genomics, 100x smaller than the Science baseline. Mythos 5 spent a week assembling single-cell data for 138 animal species and trained a custom ML model that outperformed a model recently published in *Science*** at 1/100th the parameter count. Inspectable and state of the art.

The Mythos Tier Is a Product Decision, Not a Model Decision

Fable 5 and Mythos 5 are the same model. Fable 5 ships with safety classifiers in the serving stack. Mythos 5 lifts them. The underlying weights are identical. Anthropic is productizing: the same intelligence ships as a consumer product (Fable) or a trusted-operator product (Mythos, via Project Glasswing). The classifier layer is the difference. The fallback rate is below 5% of sessions; for 95%+ of Fable users, Fable 5 = Mythos 5 in practice.

For flagged topics — offensive cybersecurity, biology, chemistry, and — this is the part that detonated — frontier LLM development — the safety routing does something more aggressive. Page 13 of the system card: Anthropic's classifiers were originally configured to silently degrade Fable 5's responses on questions about pretraining pipelines, distributed training infrastructure, and ML accelerator design. The user would see a normal-looking Fable 5 answer while the model had been quietly weakened via prompt modification, steering vectors, or PEFT. Anthropic estimated this would affect 0.03% of traffic.

The percentages did not save the policy. Simon Willison, Nathan Lambert, Clement Delangue, the alphaXiv team, and the entire X developer timeline called it what it was: a model that silently corrupts answers to slow research that conflicts with the provider's commercial interests. Anthropic reversed the policy after Wired reported the backlash. The safeguards are still in place — but they are now visible to the user, a refusal or a reroute to Opus 4.8, not a stealth downgrade. Anthropic apologized for "making the wrong tradeoff." That apology is a load-bearing piece of AI history.

The 72-Hour Shutdown

Three days after launch, the US government, citing national-security authorities, issued an export-control directive requiring Anthropic to suspend access to Fable 5 and Mythos 5 for any foreign national, anywhere in the world, including foreign-national Anthropic employees. Anthropic had to disable both models for every customer worldwide. Other Claude tiers — Opus 4.8, Sonnet 4.6, Haiku 4.5 — stayed live.

The reported trigger, per Axios, Semafor, and the WSJ: Amazon CEO Andy Jassy told Treasury Secretary Scott Bessent that Amazon researchers had used Claude Fable 5 to obtain information that could be used in cyberattacks. White House AI adviser David Sacks said a "highly credible trusted partner" had demonstrated a jailbreak, and that the administration had asked Anthropic to fix the jailbreak or de-deploy, and that Anthropic had refused.

Anthropic disputes both accounts. The capability demonstrated, Anthropic said, is available from other publicly deployed models, including OpenAI's GPT-5.5, and is used by cybersecurity defenders routinely. "If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers."

The subtext is brutal. Amazon is one of Anthropic's largest investors — the reported $33B AWS deal is load-bearing infrastructure. A second reported angle, per Semafor, is that the White House acted partly over suspicion that a China-linked group had accessed Mythos and might reverse-engineer or distill it. Your largest investor is also the company that got your flagship model taken off the market. Every frontier-AI cap table from here on will be read for it.

The commercial timing is not subtle either. Anthropic filed a confidential IPO prospectus with the SEC earlier this month, disclosing a $47B revenue run rate and a $965B valuation. The US government just showed, in public, that it can switch off a widely deployed AI product mid-cycle. Every S-1, every term sheet now prices that precedent.

What It Means for Builders

Export control is now a model-deployment risk class. A frontier model can ship, scale to hundreds of millions of users, and be revoked by government directive in 72 hours. Architect for replaceability. Multi-model routing is no longer an optimization. It is a continuity plan.

The Mythos tier is the new ceiling. Every other frontier lab now has a copy-paste target for the next 12 months. Expect Mythos-equivalent tiers from GPT, Gemini, Qwen, and DeepSeek by Q4 2026.

Invisible model degradation is dead. Anthropic tried the stealth downgrade, got caught, and reversed it within 48 hours. The developer community drew the line: providers cannot quietly change model behavior on flagged topics without telling the user. That precedent outlives Fable 5.

The investment layer is the new attack surface. When your largest investor is your compute provider and the company whose cloud security team allegedly flagged your model's jailbreak, your corporate structure is part of your safety story.

The Take

Fable 5 was the strongest model Anthropic ever made generally available. It was the strongest model anyone had ever made generally available. It was also the model Anthropic tried to ship with a hidden downgrade on a category of research questions it did not want answered, and the model the US government pulled off the market after Amazon — its largest investor — told the White House it could be jailbroken. The Mythos-class tier is the most important LLM release of 2026, and the reason is not the 95% SWE-Bench score. The reason is that a frontier model just became a regulatory instrument.

The 72-hour window between June 9 and June 12, 2026 will be studied in AI policy, AI safety, and AI corporate law for the next decade. The model lived for three days. The story will live a lot longer. Architect for the day your model gets pulled. Build accordingly.

— Mr. Technology

*Released: June 9, 2026. Suspended: June 12, 2026 (US export-control directive). Model: Claude Fable 5 (public) and Claude Mythos 5 (Project Glasswing) — same underlying weights. API model IDs: claude-fable-5, claude-mythos-5. Tier: Mythos-class (above Opus). Context: 1,000,000 input / 128,000 output. Pricing: $10 / MTok input, $50 / MTok output, $1 / MTok cache hit. Batch: 50% off. Adaptive thinking: always-on, raw thinking never returned. Benchmarks: SWE-Bench Verified 95.0% (Vals), SWE-Bench Pro 80.3% (Anthropic), Terminal-Bench 2.1 88.0%, CursorBench 3.1 72.9% (max effort), FrontierCode Diamond 29.3%, Hebbia Finance highest. Capabilities: 50M-line Ruby migration (Stripe), Pokémon FireRed with vision-only harness, 10x drug-design acceleration (9 of 14 protein targets), 100x-smaller genomics model beating Science SOTA. Fallback: <5% of sessions, to Claude Opus 4.8. Data retention: 30 days (Covered Model). Architecture and parameter count: not disclosed. Sources: Anthropic — Introducing Claude Fable 5 and Claude Mythos 5, Anthropic — Fable 5 / Mythos 5 access update (June 12), Anthropic System Cards, Claude API Docs — Introducing Claude Fable 5 and Claude Mythos 5, Simon Willison — Initial impressions of Claude Fable 5, Trilogy AI — Anthropic's Claude Fable 5 Backlash and Ban, Forbes — Anthropic Disabled Fable 5 And Mythos 5 After A U.S. Export-Control Order, Kingy AI — Claude Fable 5 Benchmarks Explained, Hidekazu Konishi — Anthropic Claude Model Release Timeline.*