On June 9, 2026, Anthropic released Claude Fable 5. It is the first publicly available model from the company's Mythos-class tier, a level above the Opus class. The model scores 80.3% on SWE-Bench Pro, leads nearly every benchmark Anthropic tested, and costs $10 per million input tokens. It also ships with something no other frontier model at this capability level has: hard safety guardrails that automatically route high-risk queries to a weaker model.
The launch also introduced Claude Mythos 5, the same underlying model without the guardrails, restricted to approved partners through Project Glasswing. Together, the two releases represent Anthropic's attempt to make the most powerful AI it has ever built available to the public without handing the same capabilities to attackers.
Stop paying monthly for Testimonial Widgets.
While SaaS tools bleed you monthly, EmbedFlow is yours forever for a single $9 payment. Drop in a beautiful, fully responsive Wall of Love in minutes. Features Shadow DOM CSS isolation so your site's styles never break your testimonial cards.
What Makes Claude Fable 5 Different
Fable 5 is not an incremental upgrade. Anthropic calls it state of the art on nearly every benchmark the company tested. The lead grows as tasks get longer and more complex. On SWE-Bench Pro, the standard agentic coding benchmark, Fable 5 scores 80.3%. The next-best public model, Claude Opus 4.8, lands at 69.2%. GPT-5.5 sits at 58.6%. That is a 21-point gap over OpenAI's best public offering.
The model also excels at tasks that require sustained focus. In testing, Stripe used Fable 5 to run a codebase-wide migration across 50 million lines of Ruby in a single day. The same work would have taken a full engineering team over two months by hand. On Hex's internal analytics benchmark for long-running analytical tasks, Fable 5 became the first model to break 90%, a 10-point jump over Opus 4.8.
Vision is another area where Fable 5 pulls ahead. It can extract precise numbers from scientific figures, rebuild web app source code from screenshots alone, and beat Pokemon FireRed using only raw game screenshots with no navigation aids. Previous Claude models needed helper harnesses to play the same game.
How the Cybersecurity Guardrails Work
Fable 5 uses the same underlying model as Mythos 5. The difference is a set of safety classifiers that sit between the user and the model. When a query falls into a high-risk category, primarily cybersecurity, biology, chemistry, and model distillation, the system routes the request to Claude Opus 4.8 instead. The user gets an answer, just not from the most capable version of the model.
Anthropic says these safeguards trigger in less than 5% of sessions on average. The company deliberately tuned them to be conservative. This means some harmless requests will get flagged and downgraded. Anthropic acknowledges this is frustrating and plans to reduce false positives after launch as it refines the classifiers.
The stakes are real. On ExploitBench, the unblocked Mythos 5 scores 78.0% at finding and exploiting software vulnerabilities. Opus 4.8 manages 40.0%. GPT-5.5 scores 34.0%. Without guardrails, a model this capable could be used to discover zero-day vulnerabilities at scale. The guardrails exist to prevent that while preserving the model's usefulness for legitimate work.
Anthropic subjected the safeguards to both internal red teaming and an external bug bounty program spanning over 1,000 hours. No universal jailbreak techniques were found. The company did not specify whether partial bypasses were discovered.
Claude Mythos 5 and Project Glasswing
While Fable 5 is the public version, Mythos 5 is the same model with safeguards lifted in certain areas. It is available only to approved partners through Project Glasswing, Anthropic's cybersecurity initiative launched in April 2026 with roughly 50 initial partners including Apple, AWS, Cisco, CrowdStrike, Google, Microsoft, and Palo Alto Networks.
Project Glasswing has already produced results. Partners using Mythos Preview found more than 10,000 high- or critical-severity vulnerabilities across critical software infrastructure. The bottleneck is no longer finding flaws. It is verifying, disclosing, and patching them fast enough.
Anthropic recently expanded Project Glasswing to approximately 150 new organizations. New members include Dragos, Tenable, Netskope, BeyondTrust, Rubrik, and Hitachi. Each organization must meet Anthropic's security requirements before gaining access. The company plans to expand further through a structured trusted-access program that will eventually include federal agencies and organizations outside the US.
Pricing and Access Details
Both Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens. That is exactly double the cost of Opus 4.8, which runs $5 input and $25 output per million tokens. It is also less than half the price of the original Mythos Preview.
Batch API pricing is available at $5 per million input and $25 per million output. Prompt caching costs $12.50 per million tokens for a 5-minute write and $20 for a 1-hour write. Cache hits cost $1 per million tokens. A 90% input token discount applies to cached prompts.
On the Claude API and consumption-based Enterprise plans, Fable 5 is fully available. For subscription plans, access is rolling out in stages due to expected high demand. From launch through June 22, 2026, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost. On June 23, it moves to usage-based billing. Anthropic says it aims to restore Fable 5 as a standard subscription feature as quickly as capacity allows.
Fable 5 is available through the Claude API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. The model supports a 1M token context window by default and up to 128,000 output tokens per request.
Benchmark Comparison: Fable 5 vs GPT-5.5 vs Gemini 3.1 Pro
Fable 5 leads across most published benchmarks, but the gap varies by task. Here is how the three frontier flagships compare on key metrics.
| Benchmark | Claude Fable 5 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|
| SWE-Bench Pro (coding) | 80.3% | 58.6% | 54.2% |
| FrontierCode Diamond | 29.3% | 5.7% | — |
| GDPval-AA (knowledge ELO) | 1932 | 1769 | 1314 |
| GDPpdf (vision, no tools) | 29.8% | 24.9% | 16.7% |
| OSWorld-Verified (computer use) | 85.0% | 78.7% | 76.2% |
| Legal Agent Benchmark | 13.3% | 2.1% | 0.0% |
| Blueprint-Bench 2 (spatial) | 38.6% | 36.2% | 26.5% |
| ExploitBench (cybersecurity) | 78.0% | 34.0% | — |
| Input/Output per 1M tokens | $10 / $50 | $5 / $30 | $2 / $12 |
On independent hallucination testing, Fable 5 also performs well. The AA-Omniscience benchmark rates Fable 5 at 36.18%, compared to GPT-5.5 at 85.53% and Gemini 3.1 Pro at 49.87%. Lower is better on this metric.
GPT-5.5 holds an advantage in native multimodal processing, supporting text, image, audio, and video natively. Gemini 3.1 Pro is the cheapest option by a wide margin and integrates tightly with Google Cloud. But on pure benchmark performance, Fable 5 leads across nearly every published metric Anthropic and third parties have reported.
Who Should Use Claude Fable 5
Fable 5 is the right choice for teams running long-horizon, complex tasks where accuracy matters more than token cost. If you are building coding agents, running codebase migrations, doing deep research, or working with vision-heavy documents, the premium over Opus 4.8 is justified by the capability gap.
It is also a strong fit for enterprise knowledge work. The model's performance on finance benchmarks, legal reasoning, and document analysis makes it useful for professional services, trading firms, and data-heavy teams. IMC, a trading firm, reported that Fable 5 aced its trading-analysis evaluations across factual lookup, conceptual reasoning, root-cause analysis, and expected-value analysis.
If your primary constraint is token cost, Fable 5 is probably not the right pick. Opus 4.8 at $5/$25 per million tokens delivers strong performance for most tasks. Gemini 3.1 Pro at $2/$12 is the most cost-efficient Western frontier model and handles multimodal workloads natively.
If you are a cybersecurity professional or researcher, Fable 5 will not give you access to its full capabilities in that domain. For unblocked cybersecurity work, you need Mythos 5 access through Project Glasswing or the trusted-access program. The guardrails on Fable 5 are not optional and cannot be disabled by the user.
The Bigger Picture
Claude Fable 5 marks a turning point. It is the first time a model at this capability tier has been released to the general public with built-in safeguards. The approach, shipping the same model in two versions with different safety profiles, gives Anthropic a way to put frontier AI in more hands without giving attackers the same tools.
Whether this model becomes the default for high-end AI work depends on two things: whether the false positive rate on the guardrails drops quickly enough, and whether the performance gap over cheaper models justifies the cost for most teams. Right now, the gap is real and measurable. For teams that need the best, Fable 5 is it.
