Claude Fable 5.1 on Amazon Bedrock: Why Fable 5 Was Banned, the Setting You Must Flip First, Model ID, Pricing and the Refusals That Return HTTP 200

Claude Fable 5.1 is Anthropic's top-priced frontier model, released on September 1, 2026, and it runs on Amazon Bedrock as global.anthropic.claude-fable-5-1 or us.anthropic.claude-fable-5-1 at $10 per million input tokens and $50 per million output tokens, with cache reads now at $0.25. Before your first call works, you must flip a Bedrock setting most people have never heard of, the data retention mode, to aws_review; until you do, every request fails with a ValidationException. Two more things the launch coverage left out. Its predecessor, Fable 5, was switched off worldwide for nearly three weeks in June after a US export-control directive, which is why so many people searched "why is Claude Fable banned". And Fable 5.1 does not always answer with an error when it declines: a blocked request comes back as a normal HTTP 200 with stop_reason: "refusal", and refusals are materially more frequent than on previous Claude models. Code that only checks for errors will treat a refusal as success.

Jake runs a phone repair shop, and his nephew Sam builds small web apps for local businesses. In September Sam read that Fable 5.1 was the best model money could buy and pointed his invoice assistant at it on Bedrock. The first day every call failed. The second day, after Ethan, who looks after AWS for a few of Sam's clients, found the retention switch, the calls succeeded, but some answers came back empty and the app happily saved blank summaries. Jake's question was simple: "If it is the best one, why does it keep saying nothing?" This page is the answer: what Fable 5.1 is and what changed from Fable 5, why Fable 5 was banned and whether it is back, the benchmarks, how it compares with Opus 5.5 and Mythos 5.1, the Bedrock model IDs and Regions, the retention switch, the real price, how to handle refusals in code, and every error you are likely to meet.

If Bedrock is new to you, our plain-English guide to what Amazon Bedrock is takes five minutes and makes every section below easier.

What Claude Fable 5.1 is, and what changed from Fable 5

Fable 5.1 is Anthropic's frontier model for ambitious coding, long-horizon agents and enterprise knowledge work. It handles software projects that span an entire codebase, runs multi-hour jobs across many tools inside agent harnesses, and reads dense filings, charts and diagrams well. On price it sits at the top of the Claude family, above Opus 5.5, Sonnet 5.5 and Haiku 5.5, and it is the generally available twin of Claude Mythos 5.1: the two are the same model with different safeguard levels.

What changed from Fable 5:

• Lower real cost. The list prices stay at $10 and $50, but cache reads fell 75%, from $1 to $0.25 per million tokens, and Anthropic estimates "about 25% lower cost for typical token-billed workloads, and up to about 45% lower for highly agentic work," because the model gets more done with fewer tokens.
• Effort that pays off. At low or medium effort, Fable 5.1 performs similarly to or better than Fable 5 at much lower cost. Claude Code defaults to high; Claude.ai and Claude Cowork default to medium.
• Fewer false alarms. Claude Code users see about 60% fewer cyber safeguard interventions per session, and biology safeguards fire 85% less often on benign elementary biology and medical questions. Fable 5.1 can now help find software vulnerabilities, though it still will not help develop exploits.
• One new limit. New API accounts "can no longer manually edit prior context while keeping the transcript of Claude's earlier thinking," a change Anthropic made to resist distillation.

Moving from Fable 5 to Fable 5.1: the checklist

1. Swap the model ID from anthropic.claude-fable-5 to anthropic.claude-fable-5-1, keeping the same global. or us. prefix.
2. Leave the retention mode alone; Fable 5.1 needs the same aws_review setting Fable 5 did.
3. Run your own test set at medium effort as well as high; Anthropic reports the lower settings now match or beat Fable 5, and the saving is in output tokens.
4. Cache more. With hits at $0.25 instead of $1, long shared prompts such as a codebase or a policy manual are worth marking for caching where they were not before.
5. Keep your refusal handling. The classifier is more precise than in June, with about 60% fewer cyber false positives, but refusals still arrive as HTTP 200.

One thing stays the same and costs money quietly: Claude 4.7 and later models, Fable included, use a newer tokenizer that produces about 30% more tokens for the same text. If you are moving a workload from an older Claude model, count tokens again before you budget.

Why Claude Fable 5 was banned, and is it back?

Searches like "why claude fable 5 is banned" and "is claude fable coming back" come from a real event. In order:

Why it happened: the government believed it had found a way of "jailbreaking" Fable 5. Amazon researchers had found a bypass: prompted the right way, the model identified a number of software vulnerabilities, and in one case it produced code demonstrating how one of them could be exploited. Anthropic disputed the weight of it at the time, writing that "we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model," and that its own testing found less capable models could identify the same vulnerabilities. It could not comply selectively, because there was no reliable real-time way to check every user's nationality, so it switched the models off for everyone.

The fix that brought Fable back is the part that still affects you. Anthropic added a new safety classifier that blocks the reported behavior "in over 99% of cases," and says researchers from the Commerce Department's Center for AI Standards and Innovation tested both the old and new safeguards and "agree that they are extraordinarily strong." The cost was stated plainly: the classifier "comes at the cost of flagging benign requests more often during routine coding and debugging tasks." That is the direct ancestor of the refusal behavior described later on this page.

So, is Claude Fable back? Yes. Fable 5 and Fable 5.1 are both listed by Anthropic as open to all Amazon Bedrock customers today. If Fable "is not working" for you right now, it is almost certainly one of the three causes in the troubleshooting section, not the ban.

Jake: "So the government pulled it off the shelf like a recalled phone charger?"

Ethan: "Close. A charger gets recalled because it can catch fire. This was recalled because someone showed it could help find the weak spot in a lock. The maker added a guard and put it back on the shelf less than three weeks later. The guard is why it sometimes refuses things that look harmless."

Is Claude Fable 5.1 good? The benchmarks

At launch, Fable 5.1 was benchmarked against Fable 5, Opus 5 and OpenAI's GPT-5.6 Sol:

Two footnotes matter. Fable 5.1 and Fable 5 scored zero on OSWorld 2.0 tasks where safeguards intervened, so the safety classifier costs real points on some tasks; the 5.1-point gap to Mythos 5.1 on Terminal-Bench 4.0 is the same effect, and Anthropic expects it to shrink. And Terminal-Bench-Science carries a standard error of 3.5 to 4.5 points per model, so small gaps there are noise. The large ones are not: Fable 5.1 more than doubles Fable 5 on scientific terminal work and nearly doubles it on AutomationBench.

Now the awkward part, and the most useful fact on this page. Three weeks after this table, Anthropic released Claude Opus 5.5 with a new table of its own, and on every one of its nine rows Opus 5.5 scores higher than Fable 5.1: 66.4% against 55.8% on Terminal-Bench 4.0, 1846 against 1735 on GDPval-AA v2.1, 57.8% against 51.8% on CursorBench 4.0, 58.7% against 52.6% on Terminal-Bench-Science, and 67.7% against 65.6% on Humanity's Last Exam with tools. Some benchmark versions differ between the two launches, and both models lose points where safeguards step in, but the direction is the same on every line. Opus 5.5 costs $4 and $20 per million tokens, 40% of Fable's price.

So is Fable 5.1 worth it? On the newest numbers, not as a default. Treat it as a model to test against Opus 5.5 on your hardest real tasks, and pay its price only where your own results show it winning. Our Claude Opus 5.5 guide has that model's details.

Claude Fable 5.1 vs Opus 5.5 vs Mythos 5.1: which one to call

Fable 5.1 costs two and a half times Opus 5.5 on every token line, needs a retention setting Opus does not, and refuses more often. And on the newer benchmarks, Opus 5.5 scores higher. The sensible pattern is to run Opus 5.5 or Sonnet 5.5 by default and send work to Fable 5.1 only where your own tests show it doing better. If you need the cyber or biology capability that Fable holds back, that is Mythos 5.1, and our Claude Mythos 5.1 guide covers how verified teams get it. For the everyday flagship, see our Claude Opus 5.5 guide.

Claude Fable 5.1 on Amazon Bedrock: model IDs, endpoints and Regions

On Bedrock's main endpoint, bedrock-runtime, Fable 5.1 supports Anthropic's Messages API, the Converse API and InvokeModel. It does not support the OpenAI-style Responses or Chat Completions APIs. There is no in-Region option anywhere on bedrock-runtime, so you always call an inference profile:

There is no eu., jp., au. or apac. profile for Fable 5.1. A team in Frankfurt or Mumbai that must keep data inside its geography cannot use Fable 5.1 on Bedrock today and should look at Opus 5.5, which has eu., au. and jp. profiles. Everyone else uses global., which is also the cheapest.

The Mantle endpoint is the confusing part. On bedrock-mantle, Fable 5.1 runs with the Messages API only, and in-Region only in GovCloud (US-West); Anthropic's Messages-on-Bedrock offering also has a regional endpoint in us-east-1. If you build on Mantle, ask the endpoint itself: GET /v1/models on your Region's Mantle URL lists what your account can call, with a status for each model. For a plain explanation of the two endpoints, read our Mantle vs bedrock-runtime guide.

Features on bedrock-runtime: response streaming, prompt caching (at least 512 tokens per checkpoint, up to 4 checkpoints, 5-minute or 1-hour TTL, on system, messages and tools), abuse detection, Guardrails, Knowledge Bases, Agents, Flows, prompt management and model evaluation. The only service tier is Standard: no Priority, Flex, Reserved or Batch on Bedrock.

Your first Fable 5.1 call on Bedrock, in five steps

1. In the Bedrock console's Model catalog, select an Anthropic model and submit the use case form once for the account; access is granted on submission.
2. Give your IAM role bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream on global.anthropic.claude-fable-5-1 and on the foundation model, with a wildcard Region.
3. Set the data retention mode to aws_review in every Region you will call from, using the command in the next section.
4. Run the one-line CLI check from the code section; an "OK" means the account, IAM, profile and retention mode all line up.
5. Before real traffic, add the stop_reason check and the Opus fallback from the refusals section, so a refusal never passes for an answer.

The switch you must flip first: data retention mode aws_review

Fable 5.1's Bedrock card is direct about it: "To use this model, you must opt in to AWS review by setting your data retention mode to aws_review via the Data Retention API." Fable is one of four models Anthropic calls "Covered Models," whose capabilities "represent a substantial step up from prior generations and create elevated risk if misused," and some misuse only shows up across many requests, so the safety checks need a window of retained data.

On Bedrock that retention happens inside AWS. Under aws_review, prompts and outputs for models that require review are retained within the AWS boundary for up to 30 days and may be reviewed by AWS, and your content is not shared with Anthropic. Until you set the mode, the model shows as unavailable on the Mantle endpoint and every bedrock-runtime call fails with a ValidationException.

There is no console button for this yet. It is one call per Region, because the setting "does not propagate to other Regions":

# Check the current mode
curl https://bedrock.us-east-1.amazonaws.com/data-retention \
  -H "Authorization: Bearer $AWS_BEARER_TOKEN_BEDROCK"

# Opt in to AWS review for Fable 5 and Fable 5.1
curl -X PUT https://bedrock.us-east-1.amazonaws.com/data-retention \
  -H "Authorization: Bearer $AWS_BEARER_TOKEN_BEDROCK" \
  -H "Content-Type: application/json" \
  -d '{ "mode": "aws_review" }'

The person running it needs bedrock:PutAccountDataRetention. Repeat it in each Region your code calls from; with the global profile, retained data sits in the destination Region that served the request. If your account was already set to the older provider_data_share mode, nothing to do: it sits above aws_review on AWS's ladder and satisfies Fable too.

Two edges. Setting aws_review does not make other models start retaining data; each model declares what it allows, and Opus 4.8, for example, permits zero retention whatever you set. And if your company forbids retention, eligible customers can use Fable 5 and 5.1 with zero data retention for internal use through December 31, 2026, as a bridge to Anthropic's Enterprise Frontier Safeguards, which keeps monitoring data in your own cloud account and rolls out in phases this fall. Ask your AWS or Anthropic account team about eligibility.

Jake: "So Sam's app was turned away at the door on day one because of a form?"

Ethan: "Because of a signature. Fable only works if you agree that AWS may keep your requests for up to 30 days in case someone is misusing it. One command, once per Region. After that it opened."

Claude Fable 5.1 pricing on Bedrock, and a worked job

These are Anthropic's list rates; regional and multi-region endpoints on Bedrock carry a 10% premium over global endpoints, which is the us. column. Thinking tokens are billed as output, and they count toward max_tokens. On the Claude API the Batch API halves input and output to $5 and $25, but Bedrock offers Fable 5.1 on the Standard tier only. The charge shows on your AWS bill under AWS Marketplace, listed by the provider, which our Cost Explorer walkthrough shows how to find.

A worked example, with every number an assumption to replace with your own: an agent job that reads 3 million input tokens, of which 2.5 million are cache hits, 300,000 are 5-minute cache writes and 200,000 are fresh, and writes 150,000 output tokens.

Read that table honestly. With identical token counts, the cache price alone makes Fable 5.1 about 12% cheaper than Fable 5. Anthropic's larger 25 to 45% estimate comes from Fable 5.1 needing fewer tokens to finish the same job, which this example cannot show; measure it on your own workload. And the same job on Opus 5.5 costs well under half. Lowering effort to medium is the biggest lever you control, since thinking is billed as output at $50 per million.

Refusals that return HTTP 200, and the fallback Bedrock will not do for you

This is what bit Sam's app. Fable 5.1's Bedrock card has a section called Content Restrictions, and it is worth quoting whole: the model "includes blocking classifiers for dual-use content in cybersecurity and life sciences. When a classifier blocks a request, the API returns a standard HTTP 200 response with stop_reason: "refusal" and a stop_details object containing the restriction category. Refusal rates on this model are materially higher than on previous Claude models." AWS then tells you how to treat it: "Customers should handle stop_reason: "refusal" as a primary response path."

Billing follows the timing. "Prompt-stage refusals (blocked before inference begins) are not billed. Mid-stream refusals (blocked after partial output) are billed for tokens generated before the block." So a refusal that arrives halfway through an answer still costs you the half you received.

Anthropic's help center explains what trips the classifiers: offensive cybersecurity such as exploits, malware or attack tooling; many dual-use biology topics including virology, toxicology, drug design and molecular design; attempts to extract the model's summarized thinking; and a narrow set of frontier AI development tasks. Crucially, the checks "also review everything the model reads," including files and tool results, "so a block can come from content you didn't type." A security log pasted into an invoice summary can trigger it.

In Anthropic's own apps, a blocked request is re-run on an Opus model, with a notice that the model switched: Opus 4.8 for offensive cyber and Opus 5 for biology. On the Claude API that switching "isn't active by default" and must be configured. On Bedrock the server-side version is not available at all: the fallbacks parameter is unsupported, so you use the client-side fallback pattern instead. Here is that pattern with the Anthropic SDK on bedrock-runtime:

from anthropic import Anthropic
from aws_bedrock_token_generator import provide_token

client = Anthropic(
    base_url="https://bedrock-runtime.us-east-1.amazonaws.com/anthropic",
    api_key=provide_token(region="us-east-1"),
)

PRIMARY = "global.anthropic.claude-fable-5-1"
FALLBACK = "global.anthropic.claude-opus-5-5"

def ask(messages, max_tokens=8000):
    r = client.messages.create(model=PRIMARY, max_tokens=max_tokens, messages=messages)
    if r.stop_reason == "refusal":
        # Log the category so you can see what tripped the classifier
        print("Fable declined:", getattr(r, "stop_details", None))
        r = client.messages.create(model=FALLBACK, max_tokens=max_tokens, messages=messages)
    return r

reply = ask([{"role": "user", "content": "Summarize this invoice: ..."}])
text = next((b.text for b in reply.content if b.type == "text"), "")
if not text:
    raise RuntimeError(f"No answer; stop_reason={reply.stop_reason}")

The last three lines are the guard Sam's app was missing: never save an empty answer as if it were a result. Log the stop_details category for a week and you will learn exactly which of your inputs trip the classifier; often it is a pasted log or a security-sounding phrase that can be trimmed before the call.

Jake: "So the blank summaries were the model saying no, politely, and the app not listening?"

Ethan: "Exactly that. It answered with a shrug instead of an error, and the app read the shrug as a finished job. Now it checks for the shrug and asks Opus instead."

Working code: Messages API, Converse, InvokeModel and the CLI

Set the retention mode first, then any of these will work. The Anthropic SDK with the effort setting:

pip install -U anthropic aws-bedrock-token-generator

response = client.messages.create(
    model="global.anthropic.claude-fable-5-1",
    max_tokens=16000,
    output_config={"effort": "medium"},
    messages=[{"role": "user", "content": "Plan the migration of this service to Python 3.13: ..."}],
)
print(response.stop_reason)
print(next((b.text for b in response.content if b.type == "text"), ""))

Bedrock's Converse API with boto3:

import boto3

client = boto3.client("bedrock-runtime", region_name="us-east-1")
response = client.converse(
    modelId="global.anthropic.claude-fable-5-1",
    messages=[{"role": "user", "content": [{"text": "Explain this chart in plain English: ..."}]}],
    inferenceConfig={"maxTokens": 16000},
)
print(response["stopReason"])
for block in response["output"]["message"]["content"]:
    if "text" in block:
        print(block["text"])

InvokeModel with the Anthropic body, for older codebases:

import json, boto3

client = boto3.client("bedrock-runtime", region_name="us-east-1")
response = client.invoke_model(
    modelId="global.anthropic.claude-fable-5-1",
    body=json.dumps({
        "anthropic_version": "bedrock-2023-05-31",
        "max_tokens": 16000,
        "messages": [{"role": "user", "content": "Review this pull request: ..."}],
    }),
)
print(json.loads(response["body"].read()))

And a one-line check from the AWS CLI that the model, the profile and the retention mode all line up:

aws bedrock-runtime converse \
  --region us-east-1 \
  --model-id global.anthropic.claude-fable-5-1 \
  --messages '[{"role":"user","content":[{"text":"Reply with the single word OK."}]}]' \
  --query 'output.message.content[0].text'

Your IAM role needs bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream on both the inference profile and the foundation model; since the global profile can run in any Region, use a wildcard Region in the foundation-model ARN, arn:aws:bedrock:*::foundation-model/anthropic.claude-fable-5-1. Leave out temperature, top_p and top_k, give max_tokens room for thinking, and pick the answer by block type rather than position, since the first block is often a thinking block.

Using Fable 5.1 in Claude Code on Bedrock

Claude Code can run on Bedrock with Fable 5.1 as its model: start it with claude --model global.anthropic.claude-fable-5-1 or set ANTHROPIC_MODEL to that ID. Set the retention mode in your Region first. At default high effort and $50 per million output tokens, a long day of agent work on Fable adds up fast, so most teams keep Opus 5.5 or Sonnet 5.5 as the default and switch to Fable with /model for the hard problems. In Claude Code, Anthropic reports about 60% fewer cyber safeguard interventions per session on Fable 5.1 than on Fable 5, and a setting under Config, "Switch models when a message is flagged," controls whether a blocked message moves to another model. Our Claude Code on Bedrock guide has the full setup and the model pins that keep the bill sane.

Is Claude Fable 5.1 free, and can you run it locally?

No, Fable 5.1 is not free anywhere. In Anthropic's apps it is not on the Free plan; Pro gets it through usage credits; Max 5x and 20x and Team premium seats can spend up to 50% of their weekly limits on it; Enterprise includes it. On Bedrock you pay per token from the first one, and there is no Fable-specific free tier.

You also cannot run it locally. Fable is a closed model: Anthropic has not released its weights, so there is nothing to download into Ollama, LM Studio or llama.cpp, and any "Claude Fable 5.1 free download" is not what it claims to be. If you want a strong model on your own machine, the open models are the route; our guides to running Qwen3.8 locally on Windows and Kali and the 27B version on an 8 GB GPU show what a home PC can do today. They are not Fable, but they cost nothing per token.

Can Claude Fable 5.1 generate images? No. It reads images, and AWS calls its vision "best-in-class" for dense filings, charts and diagrams, but it only writes text.

Claude Fable 5.1 not working on Bedrock: troubleshooting by symptom

ValidationException on every call

The Region's data retention mode is below aws_review. Run the GET from the retention section; set the mode with the PUT; repeat for every Region you call from.

"Invocation with on-demand throughput isn't supported"

You called the bare anthropic.claude-fable-5-1. Use global.anthropic.claude-fable-5-1 or us.anthropic.claude-fable-5-1; our on-demand throughput guide explains why Bedrock requires a profile.

Calls work, but some answers are empty

Those are refusals: HTTP 200, stop_reason set to "refusal". Check stop_reason on every response, log stop_details, and retry on Opus with the pattern above. Look at what the request carried, including files and tool output, since the classifier reads those too.

The us. profile fails from Europe or Asia

The US geo profile can only be called from US and Canadian Regions. Use the global. profile, which works from 33 Regions.

AccessDeniedException

Usually IAM: the policy names the model but not the inference profile, or the reverse, or locks the foundation-model ARN to one Region. Our Bedrock AccessDeniedException guide walks through each case.

400 error about temperature, top_p or top_k

Fable accepts temperature only at 1.0 and top_p only at 0.99, never both, and no top_k. Remove them, including defaults a framework adds for you.

"thinking.type.enabled is not supported"

Fable rejects the old budget_tokens style. Use thinking: {"type": "adaptive"} or leave it out, and set depth with output_config.effort.

Answers stop mid-sentence

Thinking used up max_tokens. Raise it, or lower effort to medium or low, which Anthropic reports keeps quality at or above Fable 5 for much less.

Batch, Flex or Priority request rejected

Bedrock offers Fable 5.1 on the Standard tier only. Drop the service_tier field or set it to default.

Throttling under load

Spread work across the global profile, back off and retry, and request a quota increase in Service Quotas; our "Too many requests" guide has the retry code.

Sam's invoice assistant now runs on Opus 5.5 by default, sends the two kinds of document it struggles with to Fable 5.1 at medium effort, checks stop_reason on every reply, and has not saved a blank summary since. The retention setting took one command; the refusal handling took twelve lines.

It is easy to feel behind when the most expensive model in the lineup comes with a ban story, a retention form and a habit of saying no. You are not behind. Fable 5.1 is a genuinely excellent model wrapped in rules that make sense once you see them, and every one of those rules takes a few lines to handle. Sam learned that the most expensive model is not the default model, Jake learned that a polite refusal is still a refusal, and Ethan added one more check to his setup list. Flip the switch, check the stop reason, and save Fable for the work that deserves it.

Revision note. Written October 10, 2026, thirty-nine days after Claude Fable 5.1 reached Amazon Bedrock. "Look before you leap," says the old proverb; with Fable, look at the stop reason before you save the answer.