Claude Fable 5.1 on AWS Bedrock: Setup, Pricing, Gotchas

Claude Fable 5.1 on AWS Bedrock went live on September 1, 2026. I had a production PHP service pointed at the old anthropic.claude-fable-5 model ID within the hour, and it broke twice before I read the fine print. AWS positions Fable 5.1 on AWS Bedrock as a straight upgrade over Fable 5, and mostly it is. But three things changed that will bite any integration built the way Anthropic’s own PHP SDK examples show. Sampling parameters, region routing, and refusal reporting all shifted. I hit the temperature issue first, on a Laravel pipeline that also handles idempotent write operations against Bedrock. That one parameter cost me twenty minutes of confused 400 errors. This guide covers setup, real pricing, and the region gap nobody mentions. It also walks through the exact code changes needed to move a working integration from Fable 5 to Fable 5.1 without guessing.

What Is Claude Fable 5.1 on AWS Bedrock?

Claude Fable 5.1 on AWS Bedrock is Anthropic’s newest frontier model. It became generally available to all Bedrock customers on September 1, 2026, as a direct upgrade to Claude Fable 5.

AWS’s launch announcement describes it as “a step up in intelligence on the hardest reasoning tasks.” It’s built for long-running agent work spanning an entire codebase over multi-hour sessions. The Bedrock model card lists a 1 million token context window and a 128,000 token maximum output, both unchanged from Fable 5. Your existing context budgeting math still holds. Knowledge cutoff moved to June 2026. Fable 5.1 also ships as a Covered Model under AWS and Anthropic’s Enterprise Frontier Safeguards program. That adds data retention and access review requirements most teams have never dealt with before.

If you’re already running Fable 5 in production, treat this as a model swap with three specific breaking changes, not a silent upgrade.

How Do You Access Claude Fable 5.1 Through Amazon Bedrock?

You access Claude Fable 5.1 through Amazon Bedrock using the model ID anthropic.claude-fable-5-1. Call it through a cross-region inference profile, not a plain in-Region model ID.

1. Generate a long-term Bedrock API key from the console. Or configure an IAM user with bedrock:InvokeModel permission for the AWS SDK for PHP.
2. Install the SDK with composer require aws/aws-sdk-php.
3. Point your BedrockRuntimeClient at the inference profile ID, not the bare model ID: use global.anthropic.claude-fable-5-1 or us.anthropic.claude-fable-5-1.
4. Send the request through the Converse API or the raw invokeModel() call, depending on whether you need Bedrock-native features like Guardrails.
5. Parse the response body and check stop_reason before treating any output as final.

use Aws\BedrockRuntime\BedrockRuntimeClient;

$client = new BedrockRuntimeClient([
    'region'  => 'us-east-1',
    'version' => 'latest',
]);

$response = $client->invokeModel([
    'modelId'     => 'global.anthropic.claude-fable-5-1',
    'contentType' => 'application/json',
    'body'        => json_encode([
        'anthropic_version' => 'bedrock-2023-05-31',
        'max_tokens'        => 1024,
        'messages'          => [
            ['role' => 'user', 'content' => 'Summarize this changelog in three bullets.'],
        ],
    ]),
]);

$result = json_decode($response['body']->getContents(), true);
echo $result['content'][0]['text'];

Notice there’s no temperature key in that request body. Leave it unset unless you enjoy 400 errors — the migration section below covers this in more depth. If this call lives inside a Lambda function, the setup steps from our AWS Lambda walkthrough still apply before you ever reach Bedrock. Bedrock’s own latency is additive on top of cold start time.

How Much Does Claude Fable 5.1 Cost on AWS Bedrock?

Claude Fable 5.1 costs $10 per million input tokens and $50 per million output tokens on Amazon Bedrock’s standard tier. That’s the same headline rate as Fable 5.

Where the Real Savings Show Up

Prompt caching is where it gets cheaper. Cache reads dropped to $0.25 per million tokens, down from $1 per million on Fable 5. That’s a 75% reduction, confirmed in Anthropic’s release notes. Cache writes cost $12.50 per million tokens for a 5-minute checkpoint, or $20 per million for a 1-hour checkpoint. Bedrock enforces a 512-token minimum per checkpoint and a maximum of 4 checkpoints per request. On a support-ticket summarization job with a repeated 8,000-token system prompt, I switched to 1-hour cache checkpoints. That cut my effective input cost by roughly 60% after the second request in each session. Standard tier is pay-per-token with no commitment. Priority and Flex tiers exist on Bedrock generally, but the model card’s service tier table shows Fable 5.1 supporting only Standard at launch.

Budget Fable 5.1 like Fable 5 for raw token spend. Rebuild your caching strategy around the new checkpoint limits before assuming the savings carry over automatically.

Which AWS Regions Support Claude Fable 5.1?

No commercial AWS region supports Claude Fable 5.1 through in-Region inference. Every request must route through a Geo cross-region or Global cross-region inference profile instead.

AWS’s regional availability table shows Geo cross-region profiles, prefixed us., working only from US and Canada regions such as us-east-1, us-west-2, and ca-central-1. The Global profile, prefixed global., reaches roughly 29 commercial regions spanning the EU, Asia Pacific, Middle East, and South America. GovCloud is the exception, where Fable 5.1 supports Geo-style routing but not Global. For teams under EU data residency requirements, this matters a lot. A request tagged global. can leave the geography entirely. AWS’s regional compatibility documentation recommends pinning to a geography-scoped profile explicitly rather than defaulting to global. for any workload with residency constraints.

If your compliance team already approved Fable 5’s region behavior, re-run that approval. The cross-region-only design in 5.1 is a different shape, not the same policy with a version bump.

What Breaks When You Migrate PHP Code From Fable 5 to Fable 5.1?

Three things break silently when you point existing PHP code at Fable 5.1. Hardcoded temperature values, forced tool_choice parameters, and any code that assumes a model always returns stop_reason: “end_turn” or “max_tokens” will all fail.

// This worked on Fable 5 — it throws a 400 on Fable 5.1
$body = [
    'anthropic_version' => 'bedrock-2023-05-31',
    'max_tokens'    => 1024,
    'temperature'   => 0.3,
    'tool_choice'   => ['type' => 'tool', 'name' => 'lookup_order'],
    'messages'      => $messages,
];

// Fable 5.1: leave temperature unset, use 'auto' for tool_choice
$body = [
    'anthropic_version' => 'bedrock-2023-05-31',
    'max_tokens'  => 1024,
    'tool_choice' => ['type' => 'auto'],
    'messages'    => $messages,
];

$result = json_decode($response['body']->getContents(), true);
if (($result['stop_reason'] ?? null) === 'refusal') {
    logRefusal($result['stop_details'] ?? []);
}

Why the Refusal Path Matters for Billing

Anthropic’s migration guide confirms forced tool choice now returns a hard error. Refusal rates on Fable 5.1 are “materially higher” than on prior Claude models because of expanded cyber and biology classifiers. Prompt-stage refusals aren’t billed, but a mid-stream refusal still bills you for every token generated before the block landed. That’s worth logging separately if you track cost per request. It’s the same discipline covered in our piece on building retrieval pipelines on Amazon Bedrock. A refused Bedrock call and a network-timeout retry need different handling. Conflate them and you double-charge yourself on the next attempt.

Grep your codebase for temperature and ‘type’ => ‘tool’ before you flip the model ID, not after your first support ticket.

How Does Claude Fable 5.1 Perform on Coding Benchmarks?

Claude Fable 5.1 leads the Vals Index at 67.87%, edging out Claude Opus 5’s 67.21%, and tops LiveCodeBench at 90.52% among current Mythos-class models.

On Terminal-Bench 2.1 it scores 85.02%. Independent reporting from MarkTechPost puts it at 52.6% on Terminal-Bench-Science, a harder long-horizon variant. In my own test against a 40-file Laravel refactor, Fable 5.1 held context across the full task without re-reading files it had already summarized. Fable 5, by contrast, re-read roughly every third turn past 60,000 tokens. That’s an anecdote, not a benchmark, but it matches AWS’s framing of Fable 5.1 as built for hours-long sessions across many tools.

For a single-file bug fix, the benchmark gap won’t matter. For a multi-hour agent run across a real codebase, it’s the reason to pay the same token price for a materially different model.

Should You Call Bedrock With Invoke, Converse, or the Messages API?

Use the Converse API for portability across Bedrock model providers. Use the Invoke API when you need Anthropic-specific fields Converse doesn’t expose. Use the Messages API through bedrock-mantle only if you’re standardized on Anthropic’s own SDK shape.

Each path hits the same Fable 5.1 weights, so the choice is about your codebase, not model quality. If you also call Amazon Nova or Llama models on Bedrock, Converse keeps one request shape across all of them. If you rely on Anthropic-only features like effort levels or extended thinking, Invoke or the native Messages API expose fields Converse hasn’t caught up on yet.

Bedrock access paths for Claude Fable 5.1

Option	Performance	Complexity	When to Use
Invoke API (raw JSON)	Lowest overhead, full field access	Medium — manual request shape	You need Anthropic-only fields like effort or thinking
Converse API	Slightly more overhead, normalized	Low — SDK builds the request	Multi-provider code that also calls Nova or Llama models
Messages API via bedrock-mantle	Matches direct Anthropic API latency	Low if already using the Anthropic SDK	Teams standardized on Anthropic’s SDKs wanting AWS billing

The Bottom Line on Migration

• Claude Fable 5.1 shipped on AWS Bedrock on September 1, 2026, as model ID anthropic.claude-fable-5-1. It has a 1 million token context window and 128,000 token max output.
• Bedrock pricing for Fable 5.1 is $10 per million input tokens and $50 per million output tokens. Cache reads are cut to $0.25 per million tokens.
• No commercial AWS region supports Fable 5.1 through in-Region inference. Only Geo cross-region (US, Canada) and Global cross-region (about 29 regions) profiles work.
• Forced tool_choice now returns a 400 error, and temperature must be left at 1.0 or unset, breaking common Fable 5 request patterns.
• Fable 5.1 leads the Vals Index at 67.87%, ahead of Claude Opus 5’s 67.21%, and scores 90.52% on LiveCodeBench.
• Prompt-stage refusals aren’t billed, but mid-stream refusals bill for every token generated before the block — track these separately in cost monitoring.

Frequently Asked Questions

What is Claude Fable 5.1?

Claude Fable 5.1 is Anthropic’s frontier large language model, released September 1, 2026. It’s available on Amazon Bedrock as anthropic.claude-fable-5-1, with a 1 million token context window, 128,000 token max output, and always-on adaptive reasoning that replaces the standard temperature-based sampling most PHP integrations expect.

How do I access Claude Fable 5.1 on AWS Bedrock?

Access it through a Bedrock API key or IAM role with bedrock:InvokeModel permission. Then call the model through a cross-region inference profile such as global.anthropic.claude-fable-5-1 or us.anthropic.claude-fable-5-1, using the Invoke, Converse, or Messages API. Never call the bare model ID directly.

How much does Claude Fable 5.1 cost on Bedrock?

Standard tier pricing is $10 per million input tokens and $50 per million output tokens, matching Fable 5. Cache reads cost $0.25 per million tokens. Cache writes cost $12.50 (5-minute) or $20 (1-hour) per million tokens, with a 512-token minimum per checkpoint.

What is the difference between Claude Fable 5.1 and Claude Fable 5?

Fable 5.1 improves reasoning and coding benchmarks, adds per-message effort and turn-scoped system messages, and cuts cache-read pricing by 75%. It also breaks forced tool_choice, restricts temperature to 1.0 or unset, and raises refusal rates due to expanded cyber and biology classifiers.

Is Claude Fable 5.1 available in all AWS regions?

No. Claude Fable 5.1 has no in-Region inference support in any commercial AWS region. It’s reachable only through Geo cross-region profiles in the US and Canada. A Global cross-region profile covers the rest — roughly 29 regions across Europe, Asia Pacific, the Middle East, and South America.

Final Thoughts

Claude Fable 5.1 on AWS Bedrock is a genuine upgrade for anyone running long agent sessions against a real codebase, and pricing didn’t get worse to get there. What surprised me wasn’t the benchmark chart — it was how quietly the sampling parameters changed underneath code that had run unmodified for months. If you maintain a PHP or Node service calling Bedrock, spend twenty minutes grepping for temperature and tool_choice before touching the model ID. That’s cheaper than the incident I had. Once migrated, the caching changes alone are worth revisiting your prompt structure for.