Claude Fable 5 Is Back: API Routing Guide for Mythos Class

Anthropic’s most capable publicly available model is back. Claude Fable 5 — the first Mythos-class model released for general use — had access restored on July 1 after a 19-day suspension triggered by US export controls. If you haven’t updated your routing strategy yet, now is a good time: the model sitting at the top of Code Arena with a 1,665 Elo and leading SWE-bench Verified at 95% is fully available on the Claude API again, with a new safety classifier in place.

Here’s what you actually need to know before switching your production calls to claude-fable-5

.

What Mythos Class Means for Your Code #

Mythos is Anthropic’s new top tier — sitting above Opus in the model hierarchy, built specifically for long-horizon agentic work and complex reasoning. Fable 5 is the safeguarded version of that tier: Mythos 5 exists but doesn’t include safety classifiers; Fable 5 does, which is why it’s the one you can actually ship in a product.

The API model ID is claude-fable-5

(no date suffix). It supports a 1M token context window and up to 128k output tokens per request. One immediate structural difference: extended thinking is always active and cannot be disabled. Every response contains a thinking block followed by a text block — update your response parsing accordingly.

For API calls, use streaming for any non-trivial task. Non-streaming calls on complex prompts can take 60+ seconds before returning. That’s not a bug; that’s the model reasoning.

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY from env
message = client.messages.create(
    model="claude-fable-5",
    max_tokens=4096,
    messages=[{
        "role": "user",
        "content": "Refactor this module into 4 testable units. Return a diff."
    }]
)
print(message.content[0].text)

The Benchmark Numbers Worth Knowing #

Fable 5 leads Code Arena at 1,665 Elo — roughly 100 points ahead of second place. On SWE-bench Verified it hits 95%. On math, it scores 97.6 aggregate versus Opus 4.8’s 53.3 — a gap that matters if your workflows involve any numerical reasoning. The Every team’s Senior Engineer benchmark put it at 91/100, placing it in the range of human senior engineers.

The gap is task-specific. Fable 5’s advantage is largest on multi-file refactors, cross-module debugging, architectural decisions requiring first-principles reasoning, and long autonomous runs where the model needs to plan, execute, and iterate without supervision. On a single-function fix or a well-scoped algorithm, the gap over Opus 4.8 shrinks considerably. The model is not uniformly better at everything — it’s substantially better at the hard stuff.

The Pricing Math Is More Interesting Than It Looks #

Fable 5 is nominally 2x the cost: 0/0 per million input/output tokens versus Opus 4.8’s /5. That’s the number most people stop at. The number they miss: Fable 5 uses approximately 25% of the context window that Opus 4.8 requires to complete equivalent complex tasks. A workflow that takes 200k tokens on Opus 4.8 typically closes in around 50k on Fable 5 — a 4x token efficiency gain that completely offsets the 2x price difference.

A practical rule: if a task typically requires three or more Opus 4.8 iterations to get right, routing it through Fable 5 on the first pass is likely cheaper. For simple, well-scoped tasks, Opus 4.8 remains the cost-efficient default.

The ZDR Problem Nobody Briefed Their Legal Team On #

This is the one that will catch enterprise teams off-guard. Fable 5 mandates 30-day data retention on all prompts and outputs — a trust-and-safety requirement Anthropic has made non-negotiable. Opus 4.8 supports Zero Data Retention (ZDR). Fable 5 does not.

The critical part: existing ZDR agreements do not apply to Fable 5 traffic. There is no configuration toggle, no platform exception, and no enterprise carve-out. This applies whether you’re calling through AWS Bedrock, Google Cloud Agent Platform, Microsoft Foundry, or directly through the Claude API. If you have a Claude Enterprise setup with ZDR, that agreement still doesn’t cover Fable 5 calls.

Anthropic hasn’t published guidance on how this interacts with GDPR’s data minimization and storage limitation principles — a gap compliance teams will need to flag. The practical fix: a routing classifier that tags each request by data sensitivity, then routes non-sensitive reasoning workloads to Fable 5 and regulated or high-sensitivity data to Opus 4.8.

What Happened with the Export Controls #

The 19-day suspension happened because Amazon researchers found a prompting technique that bypassed Fable 5’s safety classifiers and produced working exploit code. The US government applied export controls on June 12; because Anthropic had no reliable way to verify nationality in real-time, they suspended access for all users globally. Export controls were lifted June 30, access restored July 1.

The new classifier blocks the specific technique in over 99% of test cases. The trade-off: it flags benign coding and debugging requests more often than before. When a request is blocked, it falls back automatically to Opus 4.8 rather than returning an error — so existing integrations won’t break, they’ll just get a slower, less capable response on those edge cases.

The Routing Strategy: Use Both #

The framing of “Fable 5 or Opus 4.8” is the wrong question. The right answer is a routing layer that sends each task to the right model.

Fable 5 is a genuine capability step up for the hardest class of developer tasks. The 30-day retention requirement is a real constraint that needs a routing strategy, not a workaround. Get the classifier in place, update your response parsing for thinking blocks, use streaming, and check the official Anthropic announcement for the latest on platform availability. The model is worth the integration work.