Fable 5 Doubled in Price, Then Vanished in Three Days. Don't Hardwire Your Agents to One Model.
If you’ve been following AI news this week, you saw both headlines. First: Anthropic shipped Fable 5 and Mythos 5, the most capable, and most expensive, models ever released. Then, three days later: a U.S. government export control order suspended access to both, globally, overnight.
Anthropic says it’s working to restore access, and we believe them. Frontier models aren’t going away. But this week did something useful: in the span of a few days, the price doubled and then the model vanished entirely, on a decision nobody using it could control.
That’s the real lesson, and it has nothing to do with Fable 5 specifically. Models change. Prices change. Access itself can disappear without warning. If your agents are hardwired to one model, that’s your single point of failure.
So the question worth asking isn’t “should I use Fable 5”. It’s “what happens to my agents when the model underneath them moves”. That’s where routing stops being a cost trick and becomes infrastructure.
What actually happened
Three facts worth keeping (everything is sourced at the bottom):
1. The price. Fable 5 is billed at $10 / $50 per million tokens (input / output). For comparison: Opus 4.8 is at $5 / $25, Sonnet 4.6 at $3 / $15. On output, Fable 5 costs more than 3x Sonnet. The most expensive model Anthropic has ever made generally available.
2. The disappearance. Released June 9, included in Pro, Max, Team and Enterprise plans through June 22. Then suspended June 12: an export control order, after someone demonstrated a method to bypass its safeguards. Anthropic plays it down (“a small number of previously known, minor vulnerabilities… widely available from other models”) and says it’s working to restore access, with no date. The result: the best model on the market returns errors, and nobody knows for how long.
3. The detail everyone missed. Fable 5 ships with safeguards: on certain topics, the request is served by Opus 4.8 instead, in less than 5% of sessions. In other words, Anthropic shipped its frontier model with a router built in. When a request is not meant for Fable 5, it goes somewhere else.
Keep those last two points in mind. We will come back to them.
The real problem: your agents don’t read pricing announcements
If you fire up Fable 5 in Claude Code or in a conversation, you know what you are doing. You see the answer, you feel the quality, you make the call.
Your agents don’t.
An agent configured to use “the best available model” will send Fable 5 the same things as before: the heartbeats, the log summaries, the ticket rewrites, the small JSON extractions. Requests that a model at $15 per million output tokens (or a local model at $0) handles exactly as well.
The difference doesn’t show in the answers. It shows on the bill, at the end of the month, multiplied by the number of agents running.
And this week added a second way to get burned: that same agent, hardwired to “the best model”, started pointing overnight at a model that returns errors. Not because you misconfigured anything, just because the model disappeared out from under you.
Anthropic routes. You should too.
The most interesting fact in this release is not a benchmark. It is the architecture.
Anthropic does not trust Fable 5 with everything: less than 5% of sessions get redirected to Opus 4.8, because that is the right model for those cases. Routing is not a cost-optimization hack anymore. It has become the way frontier models themselves are shipped.
The only difference is the criterion. Anthropic routes for safety. You need to route for relevance: does this request actually need a model at $50 per million tokens?
For the vast majority of an agent’s traffic, the answer is no. For the requests that deserve it (long reasoning, complex refactoring, high-stakes decisions), the answer is yes, and that is exactly where a frontier model is worth its price.
Your options
1. Stay on Opus 4.8. Right now it isn’t even a choice: it’s what’s available. Half the price, still state of the art on long agentic tasks, and it is the model Anthropic itself redirects to when Fable 5 doesn’t serve a request. If your agents run fine, you are not missing anything urgent.
2. Bet on Fable 5’s return. Anthropic will likely restore access. But you just watched it double in price and then disappear on an external decision, in the same week. Building your stack around a single model that does that is the risk, not the fix.
3. Route. Send a frontier model only the requests that justify its price, and everything else to cheaper models or your existing subscriptions. When one model disappears, the rest keep running. That’s our option, obviously. Here is how it works.
How we handle it with Manifest
Manifest is an open source LLM Router: your agents talk to a single endpoint, and each request goes to the model that matches its actual complexity. And when a model moves, they never notice.
Concretely:
Scoring decides who deserves the top shelf. Every request is analyzed and classified into four tiers (simple, standard, complex, reasoning). Only the top of your traffic justifies a frontier model. The rest goes to cheaper models, or to your local ones.
The fallback chain absorbs the disappearances. Quota exhausted, rate limit, credits dry, or a model suspended overnight: the request moves to the next provider instead of failing. That’s exactly this week’s scenario, except your agents don’t stop.
Your subscriptions come first. Already paying for a Claude plan? Manifest plugs it in as a provider and consumes it before touching your API keys. When Fable 5 returns to the plans, you’ll make the most of it without wasting it on trivial requests that drain the quota twice as fast.
Cost limits absorb the frontier rate. When Fable 5 comes back (at $50 per million, or as out-of-plan credits), a spending cap protects you from the rude awakening. The agent that loops at 3am stops at the cap, not at the end of your budget.
And none of it is a black box: every response exposes the chosen model, the detected tier and the routing reason in the headers (X-Manifest-Model, X-Manifest-Tier, X-Manifest-Reason).
The checklist
- List your agents that default to “the best model”. Those are the ones that broke this week.
- Look at how much of their traffic is actually complex. (Spoiler: measure before guessing.)
- Put a fallback chain behind them, so an unavailable model doesn’t take your agents down.
- Set a cost cap before the frontier model comes back at its rate, not after the first bill.
Fable 5 is a great model, and it will come back. But the lesson of the week is neither its price nor its suspension. It’s that depending on any single model is the real fragility. The right answer is neither “everything on the best one” nor “don’t touch it”. It is: the right request on the right model, and a way out when it moves.
That is exactly what a router is for.
Try Manifest, it’s open source.
Sources: Anthropic announcement, access suspension, GitHub Changelog, Finout pricing comparison, usage credits breakdown.
The Manifest team
