Claude Fable 5: What Developers Need to Know Right Now (June 2026)

Anthropic Just Shipped Something Different - And Developers Are Paying Attention

Yesterday, June 9, 2026, Anthropic did something it had never done before. It took a model from its secret "Mythos" tier the one that until two months ago was only accessible to a handful of vetted cybersecurity firms and biology researchers and made it available to every developer with an API key.

The public version is called Claude Fable 5. The API model string is claude-fable-5. And if you build anything with Claude, you need to understand what changed overnight because this isn't an incremental update. Andrej Karpathy called it "a major-version-bump-deserving step change forward." That's not the kind of quote you see every release cycle.

This guide covers everything you actually need: what Fable 5 is, the benchmark numbers that matter for real work, pricing, the critical access window before June 23, and a practical decision framework for when to route to Fable 5 versus sticking with Opus 4.8 or Sonnet 4.6.

• What Claude Fable 5 is and how it differs from Mythos 5

• The exact API model string, context window, and pricing

• Benchmark comparison: Fable 5 vs Opus 4.8 vs GPT-5.5 vs Gemini 3.1 Pro

• The June 22 billing cliff you cannot miss

• When to use Fable 5 and when Opus 4.8 is still the right call

• What the safeguard fallback means for your production app

What Is Claude Fable 5? (And What Is Mythos?)

To understand Fable 5, you need the two-minute backstory on Mythos. In April 2026, Anthropic quietly released Claude Mythos Preview through a program called Project Glasswing a restricted access program involving partners like AWS, Microsoft, Apple, and CrowdStrike. The model was kept locked down because of how capable it is at finding and exploiting software vulnerabilities. It was genuinely too powerful to ship publicly without a safety layer.

Fable 5 is the answer to that problem. It's the same underlying Mythos-class model, but wrapped in safety classifiers that govern which kinds of requests it will fulfill versus hand off to Claude Opus 4.8. The naming is intentional Fable and Mythos share the same foundation, but Fable 5 is built for general use.

Here's the key thing developers need to internalize: Fable 5 sits above Opus in Anthropic's capability hierarchy. This isn't a sideways move or a specialized variant. It's a new top of the stack.

API Details - Everything You Need to Start Using It Today

Here's every practical spec in one place, sourced from Anthropic's official launch documentation:

• API model ID: claude-fable-5

• Context window: 1M tokens

• Max output tokens: 128k per request

• Input pricing: $10 per million tokens

• Output pricing: $50 per million tokens

• Batch pricing: $5 input / $25 output per million tokens

• Prompt caching discount: 90% off input tokens

• Data retention: 30-day mandatory (even for enterprise zero-retention plans)

• Amazon Bedrock model ID: anthropic.claude-fable-5 or regional prefix us.anthropic.claude-fable-5

The pricing is exactly double Opus 4.8's $5/$25 rate card. That's not a typo and whether that premium pays off depends entirely on what you're building. More on that decision in a later section.

The 30-day retention policy is new and worth flagging. Previously, many enterprise teams had zero-retention agreements with Anthropic. Those agreements do not cover Fable 5 or Mythos 5. If your app handles sensitive user data, you need to revisit your data agreements before routing production traffic through this model.

The quickest way to try it is a direct API call:

const response = await fetch("https://api.anthropic.com/v1/messages", {
  method: "POST",
  headers: {
    "x-api-key": process.env.ANTHROPIC_API_KEY,
    "anthropic-version": "2023-06-01",
    "content-type": "application/json",
  },
  body: JSON.stringify({
    model: "claude-fable-5",
    max_tokens: 4096,
    messages: [
      { role: "user", content: "Your prompt here" }
    ],
  }),
});

const data = await response.json();
console.log(data.content[0].text);

If you're already using the Anthropic SDK, swap your model string from claude-opus-4-8 to claude-fable-5. Nothing else changes in your integration the messages API is identical.

The Benchmark Numbers That Actually Matter

I'm going to be selective here because the benchmark table Anthropic published is long, and most of the rows won't affect your decision. Here are the ones that matter for developers building real applications:

Agentic Coding - The Headline Story

This is where Fable 5's lead is the most dramatic and the most relevant to anyone building with AI agents or using Claude Code.

• SWE-Bench Pro: Fable 5 80.3% vs Opus 4.8 69.2% vs GPT-5.5 58.6% vs Gemini 3.1 Pro 54.2%

• FrontierCode Diamond: Fable 5 29.3% vs Opus 4.8 13.4% vs GPT-5.5 5.7%

• Terminal-Bench 2.1: Fable 5 88.0% vs Opus 4.8 82.7% vs GPT-5.5 83.4%

The SWE-Bench gap of 11 points over the previous best and 21+ points over GPT-5.5 is not a rounding error. FrontierCode Diamond which tests deliberately brutal production-coding tasks shows Fable 5 scoring more than double Opus 4.8 and 5x GPT-5.5. These are different tiers, not competitive variants.

One nuance worth knowing: on Terminal-Bench, GPT-5.5 scores 83.4% using its Codex CLI harness. For terminal-driven agentic work specifically, GPT-5.5 with Codex remains competitive. Everywhere else in coding, the race is not currently close.

Knowledge Work and Reasoning

• GDPval-AA (economically valuable white-collar tasks): Fable 5 1932 vs Opus 4.8 1890 vs GPT-5.5 1769

• Spatial Reasoning: Fable 5 38.6%

• Tool Use: Fable 5 leads by 17.4 points over Opus 4.8

The Starred Benchmarks - Read This Carefully

Some numbers in Anthropic's table have asterisks (*). Those scores belong to Mythos 5, not Fable 5 the restricted version you cannot access. On domains like cybersecurity, biology, and certain chemistry tasks, Fable 5's safety classifiers trigger a fallback to Opus 4.8. For those specific use cases, you're getting Opus-level capability, not Mythos-level.

For the vast majority of business applications coding, analysis, content, agents, document work the safeguards never fire and you get full Fable 5 performance. This is only relevant if your app operates in security research, biotech, or similar domains.

The June 22 Deadline You Cannot Miss

This is the most time-sensitive information in this article. Anthropic's access structure for Fable 5 is staged, and the first window closes fast:

• Now through June 22, 2026: Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans. On Claude.ai it counts as 2x usage toward your plan limits, but there's no additional charge.

• June 23 onwards: Fable 5 moves to usage credits on subscription plans. Anthropic says they'll restore it as a standard feature "as soon as capacity allows," but there's no confirmed date.

• API access: Fully available today at $10/$50 pricing no window, no cliff. Just pay per token.

The practical implication: if you want to benchmark Fable 5 against your own production workloads before committing to the API pricing, do it this week. Run your hardest tasks through the model on claude.ai before June 22. That's free real-world testing you won't get after the window closes.

You can calculate the API cost difference between Fable 5 and Opus 4.8 for your specific usage using the LLM API Cost Calculator plug in your estimated monthly token volumes and the pricing numbers above to see the actual dollar delta before you commit.

Fable 5 vs Opus 4.8 vs Sonnet 4.6 - The Practical Decision Framework

The benchmark gap is real, but raw capability isn't the only variable. Here's the honest framework for choosing the right model for each task:

Use Sonnet 4.6 for:

• The majority of everyday Claude work chat, writing, quick summaries, content generation

• High-volume API calls where latency and cost matter more than maximum capability

• Tasks where you can easily evaluate the output quality yourself

Use Opus 4.8 for:

• Complex coding tasks where Sonnet 4.6 is falling short but Fable 5 feels like overkill

• Security-sensitive or biology-adjacent queries Fable 5 falls back to Opus here anyway

• Cost-sensitive production workloads that still need frontier-class reasoning

• Any task where you need zero-retention data handling

Use Fable 5 for:

• Long-horizon agentic coding tasks multi-file refactors, repository-level work, CI pipeline debugging

• Complex document understanding with large context the 1M context window + 128k output matters here

• Tasks where you'd otherwise need a senior engineer to review research synthesis, architecture decisions, legal document analysis

• Production AI agents where per-step accuracy compounds across many steps

That last point is worth expanding. A model that's 10 percentage points more accurate per step doesn't just perform 10% better across a 30-step agent workflow it fails catastrophically far less often. The compounding effect is why Fable 5's premium can net out cheaper on exactly the workloads where it looks most expensive.

The framing I keep coming back to: use Opus 4.8 by default, reach for Fable 5 when you'd otherwise spend a senior engineer's half-day on the problem.

How to Deploy llms.txt - Next.js, WordPress, and Static Sites

Once you've generated the file, deployment depends on your stack. Here are the three most common scenarios:

Next.js App Router (app/ directory)

Next.js has native support for metadata routes. Create the file at app/llms.txt/route.ts and return the content dynamically:

// app/llms.txt/route.ts
import { NextResponse } from "next/server";

export async function GET() {
  const content = `# Your Site Name

> Your one-sentence site description.

## Key Pages

- [Page Name](https://yourdomain.com/page): Short description.

## Data Usage Policy

Real-time AI search and citation is permitted.
AI model training requires written permission.
`;

  return new NextResponse(content, {
    headers: {
      "Content-Type": "text/plain; charset=utf-8",
      "Cache-Control": "public, max-age=86400",
    },
  });
}

This makes your file available at https://yourdomain.com/llms.txt and gets regenerated on each deploy. For a static version, just drop the generated llms.txt file into your /public folder Next.js serves everything in /public at the root path automatically.

The same applies to robots.txt if you're not already using app/robots.ts for dynamic generation, dropping robots.txt into /public is the quickest path.

WordPress

Upload llms.txt to your WordPress root directory (same level as wp-config.php) using your hosting file manager or FTP. That's it. WordPress doesn't interfere with static files at the root. Verify it's live by visiting https://yoursite.com/llms.txt in your browser.

Alternatively, if you're using Yoast SEO v23.5 or newer, it has a built-in llms.txt generator that auto-populates from your existing SEO metadata and regenerates weekly. Worth enabling if you're already on Yoast.

Static Hosting (Netlify, Vercel, GitHub Pages)

Drop the generated file into your project root (or /public for Vite/React apps, /static for Hugo/Eleventy). All major static hosts serve root-level text files at the bare path with no configuration needed. For Netlify specifically, verify there's no redirect rule accidentally catching the /llms.txt path in your _redirects file.

Common Mistakes That Make Your llms.txt Useless

I've reviewed a lot of llms.txt implementations over the past few months and the same mistakes show up repeatedly. Avoid these:

• Listing every single page. The whole point of llms.txt is curation. An AI model that reads 500 URLs with no descriptions learns almost nothing useful about your site. Pick the 5–15 pages that best define what you do. If a page was included on your sitemap, it doesn't need to be in llms.txt too the sitemap already handles discovery.

• Skipping descriptions. A bare list of links is marginally better than nothing. The real value comes from the one-line descriptions after the colon. These are what an AI model actually reads to understand your content. "Free tool for developers" is not a useful description. "Generate RFC-compliant cron expressions and see human-readable previews instantly" is.

• Using the wrong Markdown format. The spec requires a specific structure H1 site name first, blockquote description second, H2 section headings, then links with descriptions. Missing the blockquote or using incorrect heading levels will cause parsing failures in strict implementations.

• Accidentally blocking answer bots in robots.txt. This is the most damaging mistake. If you've added a blanket User-agent: * Disallow: / on a staging environment that got pushed to production, or if your CDN is serving a cached robots.txt from an old deployment, you're blocking everything including the bots that would cite your content.

• Setting an aggressive Crawl-delay. Some bots ignore crawl delay entirely, but for those that respect it, setting a delay over 10 seconds can effectively prevent real-time answer bots from fetching your pages before a query times out. Keep it at 1–2 seconds maximum, or remove it entirely for answer bots.

How to Verify Your llms.txt Is Working

Once deployed, verification is straightforward. First, check the file is accessible at the root path open a private browser window and visit https://yourdomain.com/llms.txt. You should see raw Markdown text, not an HTML page.

Second, check your robots.txt at https://yourdomain.com/robots.txt and confirm the AI bot rules are correct. If you're on Next.js, the Google Search Console robots.txt tester is useful for spotting syntax issues.

Third, use the Sitemap Validator to confirm your XML sitemap references are correct because a broken sitemap can indirectly hurt AI crawlability even if your llms.txt is perfect.

For deeper GEO (Generative Engine Optimization) analysis, tools like LLMrefs can show you how frequently your domain appears in AI-generated answers across ChatGPT, Perplexity, and Claude useful for measuring whether your AI SEO efforts are actually moving the needle.

Beyond technical verification, the real test is qualitative: ask ChatGPT or Perplexity about your site's primary topic and see whether your brand or a specific page gets cited. If it does and the description matches your actual content your implementation is working. If the AI still gets your site wrong or ignores it, revisit your blockquote description and section descriptions. That one-sentence summary is the most important line in the entire file.

llms.txt and AEO - The Bigger Picture

llms.txt is one piece of a broader shift called AEO Answer Engine Optimization. The idea is that ranking in Google's traditional blue-link results is no longer enough. You need to be cited in AI-generated answers, not just ranked in the index. Our guide on AEO vs SEO for 2026 covers the full picture, but llms.txt is essentially the technical foundation of any AEO strategy. It's the file that tells AI systems "I'm here, I'm trustworthy, here's what I know."

It also complements the broader trend of GEO Generative Engine Optimization which our post on whether SEO is dead in 2026 digs into in detail. If you want your developer blog or tools site to stay relevant as AI search becomes the default, this is the foundational technical work to do now before it becomes mainstream.

Quick-Start Checklist

Before we get to the FAQs, here's a condensed action list you can work through in one session:

• ✅ Generate your llms.txt and robots.txt using the free Robots.txt & LLMs.txt Generator

• ✅ Include 5–15 curated pages with one-line descriptions not a full sitemap dump

• ✅ Write a sharp one-sentence blockquote description this is what AI models read first

• ✅ Add a Data Usage Policy section (allow answer bots, decide on training bots)

• ✅ Configure robots.txt with explicit User-Agent rules for each major AI crawler

• ✅ Reference your llms.txt as a supplemental sitemap entry in robots.txt

• ✅ Deploy both files to your site root and verify they're accessible

• ✅ Test with ChatGPT or Perplexity ask about your site's main topic and see if you appear

Frequently Asked Questions

What is llms.txt and what is it used for?

llms.txt is a plain Markdown file hosted at your website's root (e.g., https://yoursite.com/llms.txt) that tells large language models and AI assistants which pages on your site are most important and what your site is about. Proposed by Jeremy Howard of Answer.AI in September 2024, it works alongside robots.txt and sitemap.xml robots.txt handles crawler permissions, sitemap.xml handles search engine page discovery, and llms.txt handles AI comprehension. It's designed to help AI tools like ChatGPT, Perplexity, and Claude accurately represent your content in generated answers.

Does Google use llms.txt for AI Overviews or ranking?

Google has not officially confirmed that it reads or uses llms.txt files for AI Overviews or traditional search ranking as of June 2026. Google's John Mueller has noted that major crawlers currently prioritise standard HTML over these files. However, the file takes about 10 minutes to implement, costs nothing, and is already adopted by Anthropic, Vercel, Stripe, and Cloudflare. Even if Google's crawler doesn't act on it today, other AI search engines and autonomous agents do and the standard is evolving quickly. The downside of not having it is higher than the effort required to add it.

What is the difference between llms.txt and llms-full.txt?

The standard defines two files. llms.txt is a lightweight curated index a list of your key pages with brief descriptions, usually under 5KB. llms-full.txt is the full Markdown content of every page listed in your index, concatenated into a single file. The full version is useful when you want an AI model to ingest your entire documentation or content base in a single fetch common for developer tool documentation sites. For most blogs and marketing sites, llms.txt alone is sufficient. Start with that, then add the full version later if needed.

How do I block AI training bots but still appear in AI search results?

Training bots (GPTBot, Google-Extended, Meta-ExternalAgent, Applebot-Extended) and answer bots (PerplexityBot, ClaudeBot, OAI-SearchBot) are different crawlers with different User-Agent strings. You can block training bots with a Disallow: / rule under their specific User-Agent in robots.txt, while leaving answer bots fully allowed. This prevents your content from being used in model training datasets while still letting real-time AI search assistants cite your pages. The Robots.txt & LLMs.txt Generator has a one-click "Block Training" preset that configures all of this correctly.

Where do I put llms.txt on my website?

It must be at the root of your domain accessible at https://yourdomain.com/llms.txt. For Next.js, put it in the /public folder or create a route handler at app/llms.txt/route.ts. For WordPress, upload it to the same directory as your wp-config.php file using FTP or your hosting file manager. For static hosts (Netlify, Vercel, GitHub Pages), place it in your project root or your static assets folder. Verify it's accessible by opening the URL directly in a browser you should see raw Markdown text.

Does llms.txt work with Disallow rules in robots.txt?

Yes, but they operate independently. robots.txt controls whether a bot can access a URL at all if a bot is blocked in robots.txt, it cannot fetch that URL regardless of what llms.txt says. llms.txt only helps bots that already have access understand your content better. Make sure the AI bots you want to read your llms.txt aren't accidentally blocked in your robots.txt first. Check your live robots.txt at /robots.txt and confirm that crawlers like PerplexityBot and ClaudeBot have Allow: / rules (or at least no Disallow rules) before you invest time on your llms.txt content.

Is the llms.txt generator on WebToolsHub free to use?

Yes. the Robots.txt & LLMs.txt Generator is completely free, requires no account, and runs entirely in your browser. No data is sent to any server. You can add your site name, description, important pages with custom descriptions, AI bot presets, and a data usage policy then download both files instantly. All tools on WebToolsHub work the same way: free, client-side, no sign-up.