Claude Fable 5 Is Back: What Changed and How to Start Using It Again

Claude Fable 5 is back online. Anthropic restored its most capable model on July 1, 2026 across Claude.ai, the Claude Platform API, Claude Code, and Claude Cowork, ending a 19-day government-ordered outage. The U.S. Commerce Department lifted its export controls on June 30, and Anthropic’s redeployment announcement followed within a day.

If you built on Fable 5 in early June, watched it go dark on June 12, and moved your workloads elsewhere, this article is for you. We reported the shutdown as it happened in our suspension coverage; this is the follow-up. The model that came back is not identical to the one that went dark. Anthropic retrained a safety classifier, added automatic rerouting to Claude Opus 4.8 for flagged requests, and set a July 7 deadline on plan-included usage. Each of those changes affects how you should re-adopt the model. And if your first step is confirming the API behaves the way it did before the outage, a testing tool like Apidog lets you replay your old Fable 5 requests and diff the responses before anything touches production.

Here’s the full picture: what happened, what changed, and how to get back to work on every surface.

The timeline: June 12 to July 1

Fable 5 launched in early June 2026 as the first model in the Claude 5 family, and the first “Mythos-class” model Anthropic made generally available. We covered the specs in detail in our launch explainer: a 1M-token context window by default, up to 128k output tokens per request, and pricing at $10 per million input tokens and $50 per million output tokens. The launch announcement introduced Fable 5 alongside Claude Mythos 5, a restricted sibling we’ll come back to in the FAQ.

Less than two weeks in, everything stopped. On June 12, the U.S. Commerce Department placed Fable 5 and Mythos 5 under export controls after administration officials cited severe cybersecurity risks tied to a reported jailbreak technique. Anthropic took both models offline the same day. Requests failed, sessions pinned to the model stalled, and teams spent the rest of the week re-pointing traffic at Opus 4.8 or Sonnet 4.6.

The freeze lasted nearly three weeks. On June 30, the Commerce Department lifted the controls, a reversal Axios reported after weeks of negotiation between Anthropic and the administration. On July 1, Fable 5 came back online across every first-party surface. The turnaround from lifted controls to restored access took about a day, which suggests Anthropic had the retrained classifier staged and waiting for the regulatory green light.

What changed under the hood

The suspension centered on one reported jailbreak technique that could bypass Fable 5’s safeguards on cybersecurity topics. Anthropic’s fix was targeted rather than broad. The company retrained a safety classifier specifically against that technique, and it says the new classifier now blocks the bypass more than 99% of the time. The Hacker News has a solid technical summary of the restoration and the classifier work behind it.

The second change is what happens when the classifier fires. Any flagged Fable 5 request is automatically rerouted to Claude Opus 4.8, and you get a notification that the reroute happened. You still get an answer; it comes from a different model. In practice this should be rare. Anthropic says more than 95% of Fable sessions involve no fallback at all, so most teams will never see the behavior. We break down the full safeguard stack, including what triggers a reroute and what the notification looks like, in our guide to Fable 5’s safety safeguards.

It’s worth pausing on what this design means for you as a developer. Anthropic could have shipped a blunter fix, a blanket restriction on security-adjacent content that would have caught legitimate work in the net. Instead, the classifier targets one known technique, and the reroute keeps flagged requests alive on a capable model rather than returning an error. Your integration degrades to Opus 4.8 instead of failing. That’s a meaningful difference for anyone running Fable 5 in a pipeline where a hard refusal means a broken run.

Nothing else about the model changed. Same weights, same 1M context window, same 128k output ceiling, same model ID. If your prompts worked in early June, they work now.

How to start using Fable 5 again

The model is live on four surfaces as of July 1. Here’s the state of each one.

On Claude.ai and paid plans

Pro, Max, Team, and select Enterprise subscribers get Fable 5 included for up to half their weekly usage limits through July 7, 2026. Pick the model from the model selector and you’re running; there’s nothing to re-enable on your account. The half-of-weekly-limits cap is worth keeping in mind if you plan to lean on Fable 5 hard this week. Once Fable 5 usage crosses that line, you’re back on other models until your limit resets.

On the API

The model ID is unchanged: claude-fable-5. Any request shape that worked before the suspension works now, and our API walkthrough covers auth, streaming, and long-output handling in depth. A minimal Messages API call looks like this:

import anthropic

client = anthropic.Anthropic()

response = client.beta.messages.create(
    model="claude-fable-5",
    max_tokens=16000,
    betas=["server-side-fallback-2026-06-01"],
    fallbacks=[{"model": "claude-opus-4-8"}],
    messages=[
        {"role": "user", "content": "Summarize the breaking changes in our v2 API spec."}
    ],
)

print(response.model)  # tells you which model served the request

One detail matters here. The fallbacks parameter shown above is the API-side version of the rerouting behavior: it’s in beta on the Claude API and on AWS, and it lets a flagged request re-run on Opus 4.8 inside the same call instead of stopping. We cover the parameter’s semantics, billing, and streaming behavior in a dedicated fallbacks guide.

Pricing is unchanged from launch: $10 per million input tokens, $50 per million output tokens, with a 90% prompt-caching discount that brings cache hits down to $1.00 per million tokens. That’s roughly 2x Claude Opus 4.8 per token and about 3x Sonnet 4.6, so caching discipline matters more here than on any other Claude model. Structure your prompts with the stable content first and verify cache reads in the usage block on every deploy.

In Claude Code

Fable 5 is selectable again in Claude Code as of July 1. Switch to it with the model command and long-horizon sessions run the way they did before the outage. Nothing about the integration changed during the suspension; the model was removed, not reworked. Our Fable 5 in Claude Code guide covers model switching, effort settings, and the kinds of agentic work where Fable 5 earns its price over Opus 4.8.

On AWS, Google Cloud, and Microsoft Foundry

Cloud availability is the one open question. Anthropic says Fable 5 will return to AWS, Google Cloud, and Microsoft Foundry “as quickly as possible,” but as of publication the per-cloud status is still evolving. Don’t assume your Bedrock or Vertex integration is live again. Check your provider’s model catalog before re-enabling traffic, and treat the first-party API as the reliable path this week. If your architecture allows it, a temporary dual-path setup, cloud where available and first-party API as the fallback route, keeps you unblocked while the providers catch up.

The July 7 usage-credits deadline

There’s a clock running. Through July 7, 2026, Pro, Max, Team, and select Enterprise subscribers get Fable 5 included for up to half their weekly usage limits. After July 7, Fable 5 access shifts to usage credits.

That gives you a window of less than a week to evaluate the restored model on plan-included capacity. If you were on the fence about re-adopting, run your heaviest evaluation workloads now, while they draw from your existing limits instead of a separate credit balance. Teams that wait until after the deadline will pay for the same evaluation twice: once in credits, once in the time spent setting up credit billing. We break down how the credit system works, what it costs, and how to estimate monthly spend in our usage credits explainer.

Two practical notes on the deadline. The included window covers subscription plans, not the API; pay-as-you-go API usage bills at standard token rates on both sides of July 7. And “up to half your weekly usage limits” is a ceiling, not a guarantee, so heavy users will hit it well before the calendar deadline does its work for them.

If you moved workloads off Fable 5 during the outage

Plenty of teams re-pointed production at Opus 4.8 or Sonnet 4.6 on June 12 and left it there. Moving back deserves more care than a one-line model swap, for three reasons.

First, verify behavior before you re-point. Nineteen days is long enough for your own prompts, tools, and downstream parsers to have drifted. Re-run your saved test scenarios against claude-fable-5 in Apidog and compare the outputs against your Opus 4.8 baseline before production traffic touches the model again. Pay attention to output length and structure; the two models format long responses differently, and a parser tuned to one can choke on the other.

Second, account for the new rerouting behavior. If your workload sits anywhere near security topics, expect the occasional Opus 4.8 fallback and make sure your response handling reads the served-model field instead of assuming Fable 5 answered. Logging which model served each request costs one field and saves you a confusing debugging session later.

Third, sequence the cutover. Interactive and evaluation traffic first, batch and production second, cloud-hosted workloads last, once your provider confirms availability. Our switch-back guide has a step-by-step runbook, including cache re-warming and cost checkpoints for each stage.

FAQ

Is Claude Fable 5 available on Amazon Bedrock, Vertex AI, or Microsoft Foundry yet?

Not confirmed as of publication. Anthropic says cloud availability will follow “as quickly as possible,” and the per-cloud status is evolving. Check your cloud provider’s model catalog directly rather than relying on the July 1 announcement, which covers first-party surfaces only.

What’s the difference between Fable 5 and Mythos 5?

Mythos 5 is the same underlying model without the safety classifiers. It’s a limited release through Project Glasswing, a program for cyberdefenders and infrastructure providers, with select biomedical researchers gaining access later through a trusted access program. Our Fable 5 vs Mythos 5 comparison covers eligibility and the practical differences. For almost everyone, Fable 5 is the answer; Mythos 5 requires program participation.

How often will my requests get rerouted to Opus 4.8?

Rarely, unless your workload touches the flagged categories. Anthropic says more than 95% of Fable sessions involve no fallback at all. When a reroute does happen, you’re notified, and the response tells you which model served it, so your code can branch on that if it needs to.

Did Fable 5’s pricing change after the relaunch?

No. Pricing is the same as at launch: $10 per million input tokens, $50 per million output, and a 90% prompt-caching discount ($1.00 per million on cache hits). The one pricing-adjacent change is the July 7 shift from plan-included usage to usage credits for subscribers.

What to do this week

Fable 5 is back, and the restored version is a safer bet than the one that launched: same capability, plus a classifier that closes the hole that got it suspended. Your move this week is straightforward. Verify the model against your own test suite, spend the plan-included capacity before July 7, and re-point production once the outputs check out. If you don’t have a regression suite for your LLM endpoints yet, this outage is the argument for building one. Download Apidog and turn your old Fable 5 requests into a repeatable test scenario, so the next disruption costs you an afternoon instead of a migration.