Claude Fable 5: API, Benchmarks, Pricing & How to Use It

Claude Fable 5: API Access, Benchmarks, Pricing & How to Use It

On June 9, 2026, Anthropic did something it had never done before: it handed the public a model from its top-secret "Mythos" tier, the class of models that, until now, only cyber-defense partners and a handful of biology researchers were allowed to touch. The public-safe version is called Claude Fable 5, and it doesn't sit in the Opus family. It sits above it.

If you've spent the last year watching models inch forward half a point at a time on a benchmark, Fable 5 is a jolt. It posts 80.3% on SWE-Bench Pro while the next-best model sits 11 points behind. It finished a migration in a 50-million-line codebase in a day. And Andrej Karpathy called it "a major-version-bump-deserving step change forward." This guide covers what it is, how it benchmarks, what it costs, and exactly how to put it into production, with every figure traced back to Anthropic's launch materials.

What is Claude Fable 5?

Think of Anthropic's lineup as a ladder: Haiku for speed, Sonnet for balance, Opus for hard problems, and now Mythos-class for the genuinely brutal ones. Fable 5 is the first Mythos-class model released for general use. It's designed for ambitious, long-running, asynchronous tasks: large-scale code migrations, multi-day agentic sessions, deep research, and dense knowledge work that previous models simply couldn't hold together.

Anthropic's own framing is the key insight: "the longer and more complex the task, the larger Fable 5's lead over our other models." On a quick one-shot question, you might not notice the difference. Hand it a problem that takes hours and dozens of steps, and the gap becomes obvious.

One wrinkle worth understanding up front: Fable 5 shares its underlying weights with Claude Mythos 5, a restricted release for vetted cybersecurity and biology partners that runs the same model with certain safeguards removed. Fable 5 is the safe-for-everyone twin. (Anthropic notes the name comes from the Latin fabula, "that which is told", a cousin of the Greek mythos. The safeguards are the only thing separating the two.)

Key specifications

Claude Fable 5 benchmarks: the full table

Here's the complete benchmark set Anthropic published at launch. A few things to read carefully:

• The scores shown are the higher of Mythos 5 / Fable 5, they're within 1-3 points of each other on most tests.
• The starred (*) rows are the exception: on cybersecurity, biology, and some reasoning tasks, Fable 5's blocking safeguards kick in, so real-world Fable 5 performance on those rows lands closer to Opus 4.8 (more on that below). Don't quote the starred numbers as "Fable 5" without that caveat.

What actually stands out

• Coding is the headline. +11 points over Opus 4.8 on SWE-Bench Pro, and more than 2x on FrontierCode (Diamond), a benchmark for difficult tasks held to production-codebase standards, where Fable 5 leads even at medium effort. That last part matters: it means strong results without burning maximum reasoning tokens.
• Spatial reasoning nearly tripled over Opus 4.8 (38.6% vs 14.5%).
• Legal reasoning went from "barely passing" to category-leading (13.3% vs GPT-5.5's 2.1% and Gemini's 0.0%).
• The Stripe story puts numbers in context: in a 50-million-line Ruby codebase, Fable 5 ran a codebase-wide migration in a single day, work Stripe estimated at over two months for a full team.
• Vision went SOTA. As a flex, Fable 5 beat Pokémon FireRed using only raw game screenshots, no maps, no navigation aids, no helper harness that earlier models needed.

Andrej Karpathy's launch-day reaction captured the mood, and notably backed up the "step change" framing with a qualitative read, not just the scores:

"The benchmarks are great and it's SOTA on everything by a margin but I'll add that qualitatively also, this is a major-version-bump-deserving step change forward (imo of the same order as Claude 4.5 was in November), peaking especially for long problem-solving sessions on very difficult problems. You can give it a lot more ambitious tasks than what you're used to, the model 'gets it' and it will just go..."

He also flagged the catch that early users are running into: the launch-day safeguards are "configured to be a little too trigger happy," something Anthropic itself acknowledges and says it will tune over time. Which brings us to the part of Fable 5 every API builder needs to understand.

The safeguard fallback (read this before you build)

Because Mythos-class capabilities carry real misuse risk, Fable 5 ships with classifiers covering cybersecurity, biology/chemistry, and distillation. When a request trips one, it's automatically answered by Claude Opus 4.8 instead, and the user is told. Anthropic reports this happens in fewer than 5% of sessions; over 95% of Fable sessions involve no fallback at all.

This is also exactly why the starred benchmark rows above matter: on those cyber/bio tasks, the published high score is Mythos 5's, while Fable 5, with safeguards active, performs closer to Opus 4.8.

What you need to know as a developer:

• If your workload touches security research, bio/chem, or anything the classifiers read as distillation, expect a slice of responses to come from Opus 4.8.
• API customers must configure Anthropic's new Fallback API, on the API it isn't fully automatic the way it is in the Claude apps.
• You aren't charged Fable prices for rerouted requests.
• There's a 30-day data-retention requirement on all Mythos-class traffic, used only for safety monitoring (not training).

A gateway helps here: you can log which requests fell back, route them consistently, and keep the behavior uniform across your app.

Claude Fable 5 pricing

Fable 5 is a premium model, roughly 2x the price of Claude Opus 4.8. Full token economics:

The nuances that decide your real bill:

• Prompt caching gives a 90% discount on cached input, design for it if you reuse large system prompts or context.
• Token efficiency offsets some of the price. Anthropic and early customers report Fable 5 finishing tasks in fewer turns and tokens, so a job at 2x the per-token rate can land closer than the sticker price suggests, on the right tasks.
• US-only inference is available at 1.1x pricing for data-residency needs.
• On Claude.ai subscription plans, Fable 5 counts as 2x usage. It's free on Pro, Max, Team, and seat-based Enterprise from June 9-22, 2026; after June 23 it needs usage credits until capacity allows it to return as a standard feature.

Because Fable 5 is both expensive per token and token-hungry on long tasks, cost control isn't optional. Routing it through a gateway lets you cap spend, cache aggressively, and reserve Fable 5 for the jobs that actually justify it.

How to access the Claude Fable 5 API

Fable 5 is available today on:

• Claude API (Claude Platform), model string claude-fable-5
• Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry
• TrueFoundry AI Gateway

Option A, Call the Claude API directly

Grab an API key from the Claude Console and call the model with the string claude-fable-5. Anthropic's quickstart walks through authentication and your first request: see the Claude API docs.

Calling the API directly is perfect for a quick test or a single app. The moment Fable 5 goes into production across more than one team, you need access control, budgets, and fallbacks, and that is where a gateway comes in.

Option B, Call Fable 5 through the TrueFoundry AI Gateway (recommended for production)

Direct API access is enough to prototype. Running a premium, token-hungry model like Fable 5 across an organization is a different problem: you have to control who can use it, cap spend, handle the safeguard fallbacks consistently, and keep one view of cost across providers. A gateway sits between your apps and the model and handles all of that.

How enterprises are actually using it, and why a gateway is better

• Cost governance on an expensive model. At $10/$50 per million tokens, an ungoverned Fable 5 rollout can run up a bill fast. Teams set per-team and per-application budgets, rate limits, and virtual keys so spend is capped and attributable before the first invoice, not after.
• Smart routing between Fable 5 and cheaper models. Most enterprises do not send everything to Fable 5. They route the hard, long-horizon jobs to Fable 5 and default routine traffic to Opus 4.8 or a smaller model, getting frontier quality where it matters without paying frontier prices everywhere.
• Automatic fallbacks and high availability. Launch-week demand for Fable 5 is expected to be high and hard to predict. Teams configure automatic fallback to Opus 4.8 or another provider on errors, rate limits, or capacity issues, so the app never goes dark.
• One integration across every provider. Fable 5 is available on Anthropic, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. Through the gateway it is one OpenAI-compatible endpoint, so switching providers or models is a one-line change with no rewrite and no lock-in.
• Observability and governance built in. Every call is logged with token usage, latency, and cost per request, per team, per model, plus guardrails and audit trails, which is what security and platform teams require before a new model is approved for production.

In practice that means a platform team can turn Fable 5 on for the whole company in minutes, with spend controls and fallbacks already in place, instead of each team wiring up the raw API and hoping the bill stays sane.

Step 1: Add Claude Fable 5 to your gateway

In the TrueFoundry gateway, open your connected Anthropic provider, search for claude-fable-5 on the Models Selection screen, enable it (its $10 / $50 pricing is shown inline), and use Access Control to decide which teams and virtual keys can call it.

Step 2: Call it with the auto-generated snippet

Open the Playground, pick claude-fable-5, and copy the ready-made Usage code snippet. TrueFoundry generates it for OpenAI, LangChain, Node.js, cURL, LlamaIndex, and more, in both streaming and non-streaming modes.

The win: switching between Fable 5, Opus 4.8, GPT-5.5, or any other model is a one-line change, and every call is governed, logged, and observable. Add the X-TFY-LOGGING-CONFIG and X-TFY-METADATA headers to tag and track spend per team or feature.

Ready to put Fable 5 into production with cost controls and fallbacks from day one? Try Claude Fable 5 on the TrueFoundry AI Gateway.

When should you use Claude Fable 5?

Use Fable 5 when the task is genuinely hard and long: large migrations, multi-stage agent workflows, deep research, complex document analysis, the jobs where Karpathy's "give it a lot more ambitious tasks" actually pays off. For shorter, latency-sensitive, or high-volume work, Opus 4.8 (half the price) or a smaller model is usually the smarter call. The pattern most teams settle on: route by task complexity, hard, high-value jobs to Fable 5, everything else to a cheaper model, all from one gateway.

Frequently asked questions

What is the Claude Fable 5 API model name? claude-fable-5, available via the Claude API and major cloud marketplaces.

How much does Claude Fable 5 cost? $10 per million input tokens and $50 per million output tokens, with a 90% prompt-caching discount on input, about double Claude Opus 4.8.

What's the context window? 1,000,000 tokens, with text, image, and file inputs.

Is Claude Fable 5 better than GPT-5.5? On SWE-Bench Pro, Fable 5 (80.3%) outperforms GPT-5.5 (58.6%) and leads on nearly every benchmark Anthropic published. Validate on your own workload before committing.

How is Fable 5 different from Mythos 5? Same underlying model. Fable 5 is the safe-for-general-use release; Mythos 5 lifts certain safeguards and is restricted to vetted partners.

Why did my request return an Opus 4.8 response? Fable 5's safeguards route flagged cybersecurity/biology/chemistry/distillation queries to Opus 4.8 (under 5% of sessions), and you aren't charged Fable prices for them.

‍