NewsBenefits of telehealth for patients and healthcare providers
The benefits of telehealth reach patients who otherwise go without care and give providers tools to serve more people without...
Claude Fable 5 is Anthropic's most capable publicly available AI model, built for long, complex, multi-step work like large code migrations, deep research, and autonomous multi-day tasks. You access it through the Claude API using the model string "claude-fable-5," through claude.ai on Pro, Max, Team, and select Enterprise plans, or through Amazon Bedrock, Google Cloud, and Microsoft Foundry. Fable 5 matters for AI engineers and product teams because it changes how much work you can hand off to a model without supervising every step.
What Is Claude Fable 5, and Who Can Access It?
Claude Fable 5 is a Mythos-class model, Anthropic's top capability tier, released for general use on June 9, 2026. Anthropic built Fable 5 for ambitious coding projects, multi-stage knowledge work, and long-horizon autonomous sessions that stretch across hours or days. The model shares its underlying architecture with Claude Mythos 5, but Fable 5 carries additional safety classifiers that Mythos 5 does not.
Developers reach Fable 5 through the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Claude.ai subscribers on Pro, Max, Team, and select Enterprise plans get access through the model picker in the chat interface. Mythos 5, by contrast, stays limited to vetted organizations inside Anthropic's Project Glasswing program.
Why Setup and Access Rules Matter Right Now
Access to Claude Fable 5 has changed twice since launch, and both changes affect how you should plan a rollout. Anthropic suspended Fable 5 and Mythos 5 on June 12, 2026, after the U.S. Department of Commerce applied export controls tied to a reported jailbreak. Amazon researchers had found a prompting method that led Fable 5 to identify a small number of previously known, minor software vulnerabilities.
Anthropic tested the same prompts against Claude Opus 4.8, GPT-5.5, and other competing models and found they could surface the same vulnerabilities. The Department of Commerce lifted the export controls on June 30, 2026, and Anthropic restored full access on July 1, 2026. Anthropic also shipped an improved safety classifier that blocks the reported technique in more than 99% of attempts. Build your integration assuming access rules can shift again, and keep a fallback model configured.
What Governs How Claude Fable 5 Operates?
Anthropic's own safety classifiers, not a single external regulator, govern what Claude Fable 5 will and will not do. The model automatically declines certain requests in cybersecurity, biology, chemistry, and health and routes those queries to Claude Opus 4.8 instead. A declined request returns stop_reason: "refusal" as a standard HTTP 200 response, not an error, and the API reports which classifier triggered the decline.
Claude Fable 5 also carries mandatory 30-day data retention for safety monitoring. Anthropic classifies Fable 5 and Mythos 5 as Covered Models, which means they are not available under a zero data retention arrangement. Confirm this data retention requirement with your compliance team before sending production or client data through the model.
Featured Course
Understanding how a model works is a useful first step. But knowing it exists and knowing how to prompt it, structure a workflow around it, and handle refusals under real project deadlines are two different skills. Our Generative AI and Prompt Engineering Masterclass gives engineers and product teams the hands-on framework to apply models like this correctly in the agentic workflows they actually build — not just in a reference guide.
How to Access and Set Up Claude Fable 5: A Checklist
Use this checklist before you send your first production request to Claude Fable 5.
-
Confirm your plan or API tier. Claude.ai requires a Pro, Max, Team, or select Enterprise plan. API access requires an Anthropic, AWS, Google Cloud, or Microsoft Foundry account with Claude-Fable-5 enabled.
-
Set the model string correctly. Use claude-fable-5 in the Messages API. Confirm the same string across every environment (staging, production, CI) to avoid silent fallbacks to a different model.
-
Budget for pricing. Fable 5 costs $10 per million input tokens and $50 per million output tokens. US-only inference runs at 1.1x that rate for organizations that require domestic processing.
-
Plan for a 1M-token context window. Fable 5 supports up to 128,000 output tokens per request. Size your max_tokens parameter to leave room for both thinking tokens and the final response.
-
Build refusal handling into your integration. Add a check for stop_reason: "refusal" and configure the fallbacks parameter, or client-side SDK middleware, to retry automatically on Claude Opus 4.8.
-
Review the 30-day data retention requirement. Confirm this policy against your organization's data handling rules before processing regulated or client data.
-
Track your usage credits. Through July 7, 2026, eligible claude.ai subscribers can use up to 50% of their weekly plan limit on Fable 5 before usage credits apply. After that date, credits govern all Fable 5 use on claude.ai.
Implementing Automated Refusal Handling
Because Fable 5’s newly retrained July 1 safety classifiers err on the side of caution, benign code (especially infrastructure or security auditing scripts) will occasionally trigger a false positive.
When a request is blocked, the API won't throw an error; it gracefully returns stop_reason: "refusal" inside a standard HTTP 200 response. Here is a quick example of how to handle this in your integration using Python, automatically dropping back to Claude Opus 4.8 if a refusal occurs:
Python
import anthropic
client = anthropic.Anthropic()
def run_fable_task(prompt_content):
# Attempt the request using Claude Fable 5
response = client.messages.create(
model="claude-fable-5",
max_tokens=4000,
extra_headers={"task-budgets-2026-03-13": "high"}, # Set effort
messages=[{"role": "user", "content": prompt_content}]
)
# Check if the safety classifier intercepted the request
if response.stop_reason == "refusal":
print(f"⚠️ Fable 5 Refusal Triggered (Classifier: {response.refusal_error_type}). Falling back to Opus 4.8...")
# Automatic fallback execution
fallback_response = client.messages.create(
model="claude-3-opus-20240229", # Smooth downgrade to Opus 4.8
max_tokens=4000,
messages=[{"role": "user", "content": prompt_content}]
)
return fallback_response.content[0].text
return response.content[0].text
Note on Billing: Anthropic does not bill you for the portion of the Fable 5 request blocked before output generation begins, and your fallback traffic to Opus is billed at standard Opus rates.
What Should Developers Know About Prompting Claude Fable 5?
Claude Fable 5 performs best with short, goal-oriented prompts, not long step-by-step instructions. Anthropic's own guidance states that prompt frameworks built for earlier Claude models are "often too prescriptive" for Fable 5 and can lower output quality. State the goal, the reason behind it, and any hard boundaries, then let the model plan the steps.
Prescriptive vs. Goal-Oriented Prompting
To get the most out of Claude Fable 5, stop micro-managing its steps. Because Fable 5 features advanced, always-on adaptive thinking and proactive self-verification, overly restrictive prompts actually degrade its performance.
Here is a side-by-side example of how I refactored an agentic migration prompt for the new model:
|
Old Prescriptive Style (Avoid) |
Goal-Oriented Style (Best for Fable 5) |
|
"Step 1: Open the auth.py file. Step 2: Look for lines 45-90. Step 3: Replace the old hashing function with Argon2ID. Step 4: Write a unit test using pytest to make sure it functions properly." |
"Migrate our legacy auth.py user authentication system to use Argon2ID. Ensure the implementation handles salt generation securely, and generate a comprehensive unit test suite using pytest to verify performance under heavy concurrent loads." |
Why this works better: In my testing, giving Fable 5 the high-level goal allows its internal reasoning loop to map out edge cases—like token-level complexities and legacy dependency conflicts—that human engineers often miss when writing prescriptive instructions.
The effort parameter controls the trade-off between intelligence, latency, and cost, and it replaces manual extended-thinking controls from earlier models. Anthropic recommends high effort as the default for most tasks, xhigh for capability-sensitive work like complex migrations, and medium or low for routine, latency-sensitive jobs. Claude Fable 5 uses adaptive thinking at all times; you cannot disable thinking on this model the way you could on Claude Sonnet 4.6.
Design your harness for long, asynchronous runs. Fable 5 can work autonomously for hours on a single task, so build progress checks through scheduled jobs rather than a blocking request that waits for one response. Ask the model to report only work it can point to evidence for, since long unsupervised runs raise the risk of an agent overstating what it actually completed.
