Anthropic Just Split the Frontier in Two
Fable 5 is the model we get. Mythos 5 is the model we don’t. The gap between them is the product strategy.
Jul 1, 2026, critical update: The U.S. government lifted the export controls on Fable 5 and Mythos 5. Anthropic restored Fable 5 globally across Claude.ai, Claude Code, Cowork and the API, while Mythos 5 returned only for a limited group of approved U.S. organizations.
Jun 12, 2026, critical update: The U.S. government, citing national security authorities, issued an export control directive suspending access to Fable 5 and Mythos 5 for foreign nationals, regardless of whether they are inside or outside the United States, including foreign-national Anthropic employees. The net effect is that Anthropic suspended customer access to Fable 5 and Mythos 5 altogether to ensure compliance.
My original analysis below.
TL;DR: Anthropic shipped Claude Fable 5 on June 9, 2026, the same underlying model as Claude Mythos 5, with a classifier layer that re-routes 5% of high-risk queries to Opus 4.8. It's $10 input / $50 output per million tokens, exactly 2x Opus 4.8. The buried trap: every "show your reasoning" prompt from the GPT-4 era silently downgrades Fable 5 to Opus 4.8 via the reasoning_extraction classifier. You don't pay Fable prices for the downgrade, but you do pay in stamina, context, and the half-hour you spent assuming a long-running agent was running on Fable. Nothing in the API warns you before it happens. Here's the routing playbook and the math on when 2x is worth it.
When a company announces two products from one model, the model is no longer the only thing being productized.
Today, Anthropic launched Claude Fable 5 and Claude Mythos 5.
Same underlying model. Different doors. Different rules.
This must be an interesting time to be a PM at Anthropic.
The model creates the capability, but the product is shaped in the access layer: segments, permissions, routing, eligibility, pricing, governance, and trust.
For builders, that access layer is part of the product we are building on top of.
What’s Inside
What Fable 5 Is
Fable 5 is the version you and I get: Anthropic’s most capable generally available model, built for software engineering, knowledge work, vision, scientific research, and long-running agentic tasks.
In API terms, the model ID is claude-fable-5. It supports a 1M token context window by default and up to 128k output tokens per request.
And Fable 5’s advantage seems to grow with task length.
Short tasks test fluency. Long tasks test endurance: whether the model can hold context, recover from ambiguity, follow constraints, inspect its own work, and keep going after the easy part is over.
Software migrations. Financial analysis. Chart and table reasoning. Vision-heavy coding.
Long tasks are also where enterprises burn money: handoffs, review loops, retries, context rebuilding, and humans babysitting work that was supposed to be automated. Less supervision when work gets long enough to become expensive is where the real enterprise value begins.
Hey, I’m Karo Zieminski 🤗.
I’ve been smuggling ethics into tech since 2019 and spent the last several years working at the edge of AI product development, ethical AI, and AI-assisted building.
I also build with AI myself and write Product with Attitude, an AI newsletter for tens of thousands of readers across 146 countries, helping them develop critical AI literacy the only way it sticks: through practice.
If you’re new here, welcome! Here’s what you might have missed:
What Mythos 5 Is
Mythos 5 is the same model with safeguards lifted in selected areas. You can’t buy it.
This is frontier AI with a guest list. Access is limited because the capability can do real damage in the wrong context. The same skills that help defenders can also help attackers.
Anthropic first introduced Mythos as a limited preview in April. At the time, only a small number of partners could access it through Project Glasswing, including the US government and trusted cyberdefense partners.
Last week, Anthropic widened access to hundreds of organizations in 15 countries.
But this is still not open access.
My PM read: gated access is a trust-as-a-feature play. Anthropic created a restricted tier that money cannot unlock. Only vetting can.
I personally think limiting access to capabilities considered dangerous can be a good idea. But I would not be surprised if this creates backlash.
A private company is now deciding who gets access to certain frontier AI capabilities, using criteria the public cannot fully inspect.
Who counts as “trusted”?
What does “trusted” mean?
These are difficult ethical choices. Open access can create real harm. Restricted access can create private gatekeeping.
Again: an interesting time to be on Anthropic’s product team.
The New AI Safety Pattern: Route, Don’t Refuse
Fable 5 comes with new safeguards for cybersecurity, biology, chemistry, and model distillation.
When a request triggers those safeguards, Fable 5 does not simply refuse. It routes the response through Claude Opus 4.8 instead.
Luckily, we won't be charged Fable prices for rerouted requests.
My PM read: routing, not refusing, is graceful degradation applied to safety. Refusal is a dead end in the user journey. Fallback keeps the session alive.
I wouldn’t be surprised if other labs copy this quickly.
But graceful degradation is still degradation. If an agent starts on Fable 5 and finishes on Opus 4.8, the product may keep moving while you lose the capability you selected the model for.
Build Like Task Length Is the New Benchmark
For builders, the interesting part is Fable 5’s stamina.
We can use it to audit old codebases, which I’m planning to do for all my 2025 builds.
We can use it to migrate systems.
We can use it to reason through charts, tables, screenshots, and messy product flows without resetting the context every five minutes.
And vision-heavy coding gets fun here too. If Fable 5 can reliably build from screenshots, tools like Magic Patterns will need a new moat.
My AI builder read: If task stamina is the new axis of competition, our evaluation criteria need to change with it.
For builders, the winning model is not always the one with the best answer on the first turn, but the one that can preserve context, follow constraints, recover from ambiguity, and keep going.
My critical AI literacy read: longer autonomous runs mean longer stretches where no human is checking the work.
So design checkpoints. Know where the human review must happen, and don’t automate that moment away.
The “Show Your Reasoning” Trap
I found this in Anthropic’s prompting guide, and I’m not sure how I feel about it.
A common way to get better answers from AI models is to ask them to show their reasoning.
Things like:
Show me your reasoning.Walk me through your thinking.Think step by step and show your work.But Anthropic’s documentation says that if a prompt asks Fable 5 to echo, transcribe, or explain its internal reasoning in the final answer, it can trigger a safety category called reasoning_extraction, causing what the docs call "elevated fallbacks to Claude Opus 4.8."
In plain English: the model may treat the prompt as asking for something it shouldn’t reveal, then route the response to Claude Opus 4.8.
I still want to understand how the models work, so instead of asking for reasoning, I’ll be asking for sources, assumptions, and checks. Let’s see if that works.
My critical AI literacy read: prompt patterns are not portable forever.
The same instruction that improves one model can create routing, reliability, and cost problems on another.
I explained this more in my prompt engineering article:
