Claude Fable 5 dropped yesterday (June 9, 2026), and it’s not just another incremental upgrade. This is Anthropic’s first Mythos-class model — a tier above Opus — and the benchmarks are genuinely staggering. We’re talking 95% on SWE-bench Verified. That’s not a typo.
I’ve spent the last 24 hours putting it through its paces, reading every scrap of documentation, and digging into the safeguard system that’s already generating controversy. Here’s everything you need to know.
What Is Claude Fable 5?
Claude Fable 5 is Anthropic’s new flagship model, sitting above Opus 4.8 in their model hierarchy. The API model ID is claude-fable-5, and it introduces a new model class called “Mythos” — which has implications we’ll get into later.
The key specs:
- Context window: 1 million tokens
- Max output: 128K tokens
- Extended thinking: Supported
- Model class: Mythos (above Opus)
- Release date: June 9, 2026
If you’ve been following the Claude lineup from Opus 4.7 through 4.8, Fable 5 represents a significant architectural leap. This isn’t just a fine-tuned Opus — it’s a fundamentally different capability tier.
Benchmark Results: The Numbers That Matter
Let’s cut straight to the data. Here’s how Fable 5 stacks up against Opus 4.8 and GPT-5.5:
| Benchmark | Claude Fable 5 | Opus 4.8 | GPT-5.5 |
|---|---|---|---|
| SWE-bench Verified | 95.0% | 88.6% | — |
| SWE-bench Pro | 80.0% | 69.2% | — |
| FrontierCode Diamond | 29.3% | 13.4% | — |
| Every Senior Engineer | 91/100 | 63/100 | 62/100 |
The SWE-bench Verified score of 95% is remarkable. That means Fable 5 can solve 95 out of 100 real-world GitHub issues autonomously. For context, Opus 4.8 was already considered excellent at 88.6%.
But the number that should really grab your attention is the Every Senior Engineer benchmark: 91/100 vs 63 for Opus and 62 for GPT-5.5. This benchmark tests the kind of holistic engineering judgment that separates senior devs from mid-level — architecture decisions, code review quality, debugging complex systems. Fable 5 doesn’t just edge ahead; it obliterates the competition.
The FrontierCode Diamond result (29.3% vs 13.4%) is also notable. This benchmark tests the hardest competitive programming problems, and more than doubling Opus’s score shows genuine reasoning improvement, not just better pattern matching.
Pricing: What It Costs
Here’s where things get real:
| Input | Output | Batch Input | Batch Output | |
|---|---|---|---|---|
| Fable 5 | $10/M tokens | $50/M tokens | $5/M | $25/M |
| Opus 4.8 | $5/M tokens | $25/M tokens | $2.50/M | $12.50/M |
Fable 5 is exactly 2x the price of Opus 4.8 across the board. That’s a significant premium. For a deep dive into how this compares with the broader market, check out our AI API pricing comparison.
Is it worth 2x? That depends entirely on your use case. If you’re running coding agents on complex codebases where the difference between 88% and 95% success rate saves you hours of manual debugging — absolutely. For general chat or simple tasks? Stick with Opus.
Free access period: Fable 5 is available at no extra cost on Pro, Max, Team, and Enterprise plans through June 22, 2026. After June 23, it requires usage credits.
How to Access Claude Fable 5
You have several options:
Claude.ai: Available now on Pro, Max, Team, and Enterprise plans. Free through June 22.
API: Use model ID claude-fable-5 in your API calls. Works with the standard Messages API.
Third-party routers: If you’re using OpenRouter, expect Fable 5 to be available there shortly.
Coding tools: Works with Aider, Claude Code (see our cheat sheet), and any tool that supports the Anthropic API.
import anthropic
client = anthropic.Anthropic()
message = client.messages.create(
model="claude-fable-5",
max_tokens=8192,
messages=[
{"role": "user", "content": "Your prompt here"}
]
)Extended thinking is supported, so you can enable it for complex reasoning tasks:
message = client.messages.create(
model="claude-fable-5",
max_tokens=16000,
thinking={
"type": "enabled",
"budget_tokens": 10000
},
messages=[{"role": "user", "content": "Complex problem here"}]
)The Safeguard System: What You Need to Know
This is where Fable 5 gets interesting — and controversial. Because it’s a Mythos-class model, Anthropic has implemented a new safeguard layer that didn’t exist on Opus.
How it works: An AI classifier evaluates incoming queries. If it detects sensitive topics — specifically cybersecurity exploitation, biological research, or model distillation — it falls back to an Opus 4.8 response instead of using Fable 5’s full capabilities.
On the API: The default behavior is BLOCK — the request is rejected outright. Developers can opt into the fallback behavior instead, which returns an Opus 4.8 quality response rather than an error.
Impact: Anthropic says fewer than 5% of sessions are affected. In practice, most developers writing normal application code, doing data analysis, or building products will never hit the safeguards.
But there’s a more controversial aspect: hidden competitor blocking. Fable 5 reportedly limits its effectiveness for frontier LLM development tasks — pretraining pipelines, distributed training architectures, ML accelerator design — without informing the user. It uses techniques like prompt modification, steering vectors, or PEFT to subtly degrade response quality for these topics. You won’t get an error; you’ll just get a less helpful answer.
This is a significant trust issue and something the developer community is actively debating. For a deeper dive, read our full safeguards explainer.
The Mythos 5 Connection
Here’s something important: Claude Fable 5 and Mythos 5 are the same underlying model. The difference is that Mythos 5 has the safeguards removed.
Mythos 5 is deployed through Project Glasswing, a US government cyber defense initiative. Approximately 150 organizations in over 15 countries have access to the unrestricted model for national security purposes.
This means Fable 5 is essentially the “civilian” version of a model that governments are using for offensive and defensive cyber operations. The safeguards exist specifically to prevent the public release from being used for those same purposes.
For the full story on Mythos 5, read our dedicated explainer.
30-Day Data Retention Policy
All Mythos-class model traffic (including Fable 5) is subject to a mandatory 30-day data retention policy. Your prompts and responses are stored for 30 days.
Important caveats:
- Anthropic states they will not train on this data
- The retention is for safety monitoring and abuse prevention
- This applies to all access methods (API, Claude.ai, third-party)
If you’re working with sensitive data, review our AI GDPR developers guide to understand the compliance implications.
When to Use Fable 5 vs Opus 4.8
Use Fable 5 when:
- Working on complex, multi-file codebases where higher accuracy saves debugging time
- Tasks requiring senior-engineer-level judgment (architecture, system design)
- Long-context analysis where the 1M window and superior reasoning combine
- Hard algorithmic or competitive programming problems
- You need the absolute best output quality and cost is secondary
Stick with Opus 4.8 when:
- Simple to medium complexity tasks where 88% accuracy is sufficient
- High-volume batch processing where 2x cost matters
- General conversation, writing, or analysis
- You’re using local models for initial iterations and only escalating to cloud for final review
- Your workflow already works well with Opus (if it ain’t broke…)
For coding agents specifically, the SWE-bench jump from 88.6% to 95% means fewer failed attempts and less human intervention. If you’re running Claude Code routines on production codebases, the reliability improvement might justify the cost.
Context Engineering with Fable 5
The 1M token context window combined with 128K output makes Fable 5 particularly powerful for context engineering workflows. You can load entire codebases, documentation sets, and conversation histories while still getting substantial outputs.
Extended thinking adds another dimension — the model can “think through” complex problems before responding, which particularly helps with:
- Multi-step refactoring across many files
- Debugging complex race conditions or distributed systems issues
- Architecture decisions that require considering many constraints simultaneously
Frequently Asked Questions
Is Claude Fable 5 free to use?
Yes, through June 22, 2026. It’s available at no additional cost on Pro ($20/mo), Max, Team, and Enterprise Claude plans. After June 23, it requires usage credits. On the API, standard pricing applies from day one: $10/M input tokens, $50/M output tokens.
What’s the difference between Fable 5 and Mythos 5?
They’re the same underlying model. Fable 5 is the public version with safety safeguards active. Mythos 5 is the unrestricted version available only to approved government and defense organizations through Project Glasswing. The safeguards in Fable 5 redirect certain sensitive queries (cyber, bio, distillation) to Opus 4.8-level responses.
Will my data be used for training?
No. While there is a mandatory 30-day data retention policy for all Mythos-class traffic, Anthropic explicitly states this data will not be used for model training. The retention is for safety monitoring purposes only.
Can I use Fable 5 with coding tools like Aider or Claude Code?
Yes. Any tool that supports the Anthropic API can use Fable 5 by specifying claude-fable-5 as the model ID. This includes Aider, Claude Code, Continue, Cursor, and others. Check our Claude Code vs Codex CLI vs Gemini CLI comparison for tool recommendations.
What happens when the safeguards trigger?
On the API with default settings, you get a blocked response (error). If you opt into fallback mode, you receive an Opus 4.8-quality response instead. On Claude.ai, the fallback happens transparently — you get a response, just not at Fable 5’s full capability level. Less than 5% of sessions are affected.
Is Fable 5 worth 2x the price of Opus 4.8?
For complex coding tasks, yes. The jump from 63 to 91 on Every Senior Engineer and from 88.6% to 95% on SWE-bench represents a qualitative leap in capability. For simple tasks, general writing, or high-volume processing, Opus 4.8 remains excellent value. See our detailed comparison.
Bottom Line
Claude Fable 5 is the most capable coding model available today, full stop. The benchmarks don’t lie — 95% on SWE-bench Verified and 91/100 on Every Senior Engineer put it in a category of its own. The safeguard system adds complexity and the hidden competitor blocking raises legitimate trust concerns, but for most developers writing application code, this is simply the best AI assistant money can buy.
The question isn’t whether Fable 5 is good. It’s whether your workflow demands the best, or whether “very good” (Opus 4.8 at half the price) is good enough. For many of us, the answer will be: use Fable 5 for the hard stuff, Opus for everything else.