Claude Fable 5 & Mythos 5: Everything You Need to Know

On June 9, 2026, Anthropic announced two new Claude models from a tier it calls "Mythos-class". The tier sits above the Opus line, which had previously been the top of the Claude family.

• Claude Fable 5 is the public release. It is a Mythos-class model with additional safeguards, available to paid subscribers and enterprise customers.
• Claude Mythos 5 is the restricted version. It uses the same underlying model, but with some safeguards lifted for a small group of vetted cyberdefenders and infrastructure providers through Anthropic's Project Glasswing.

The naming is deliberate. "Fable" comes the Latin fabula, roughly "that which is told," which is related to the Greek mythos. In practice, Anthropic is drawing a line between the public model and the less restricted version of the same system.

Fable 5 is now Anthropic's most capable broadly available model, according to the company. Anthropic and early press coverage point to gains in software engineering, knowledge work, vision, memory, and scientific research.

Mythos first appeared in April 2026 and quickly drew attention from the security world. The unrestricted preview version reportedly found thousands of critical and severe vulnerabilities, including bugs and exploits across major operating systems and browsers. That kind of capability is useful for defenders, but it can also help attackers.

Anthropic kept the preview version tightly limited, with access for roughly 150 vetted organizations, including financial institutions, software vendors, and healthcare networks. Fable 5 is the public version of that model family, with extra safety systems in place.

Those systems include two important behaviors:

• Guardrails block many responses on high-risk topics such as cybersecurity, biology, and chemistry.
• When a request trips a sensitive-topic classifier, Anthropic says the query can be handed off to the older Claude Opus 4.8 model instead. The company says this happens in fewer than 5% of sessions on average.

Mythos 5 remains the restricted version for trusted partners. Anthropic also says it is the company's first model to consistently produce novel, compelling scientific hypotheses, which is one of the more unusual claims around this launch.

Fable 5 and Mythos 5 are built for long-running, autonomous work. Anthropic says their advantage grows as tasks get longer and more complex.

• Software engineering: Stripe used Fable 5 to run a migration across a 50-million-line Ruby codebase in a single day, work that would have taken a team more than two months by hand. Fable 5 also posts the top score among frontier models on Cognition's FrontierCode evaluation, even at medium effort. Anthropic says it is more token-efficient than previous Claude models, which helps offset the higher price per token.
• Knowledge work: Fable 5 reportedly scores highest on Hebbia's finance benchmark for senior-level reasoning. CNBC reported that it landed more than 10% above Claude Opus 4.8 on some benchmarks, while the trading firm IMC said it "aced their trading-analysis evaluations nearly across the board."
• Vision: This matters for PC hardware because multimodal models increasingly work from screenshots, diagrams, charts, and app interfaces. Fable 5 can extract precise numbers from scientific figures, rebuild a web app's source code from a screenshot, and play visual tasks with less scaffolding than earlier models.
• Memory and long context: The model is designed to stay coherent across millions of tokens and improve its own work using file-based notes. In Slay the Spire, persistent memory reportedly improved performance three times more than it did for Opus 4.8, and the model reached the final act three times as often.
• Science: Anthropic says Mythos 5 accelerated parts of drug design by roughly 10x, with 9 of 14 protein targets producing strong candidates now under investigation. In blinded comparisons, scientists preferred its hypotheses about 80% of the time over Opus-class models.

Anthropic's pitch is that a more capable model can still be cheaper per completed task, even if each token costs more. Some early customers reported lower spend per task. That may hold for difficult engineering and research work, but the pricing still matters. Long agent sessions can use a lot of tokens.

Availability may also be uneven at launch. Anthropic has already been dealing with high demand, and Fable 5 adds more pressure to that system.

Fable 5 and Mythos 5 are cloud-only models. There are no open weights to download, no local VRAM requirement, and no quantization setting to tune. Access runs through Anthropic's apps and API.

That makes them different from open models that can run on a local PC. It also changes the buying decision. Fable 5 makes the most sense for work that justifies its cost: long coding jobs, difficult analysis, deep research, or tasks where a weaker model burns more time than it saves.

Local AI still matters in that mix. A strong local system gives users a place to run open models for everyday work with no per-token bill and more control over data. Cloud models can handle the hardest jobs, while local models cover drafting, summarization, code assistance, document review, and private experiments.

For that local side, the CORSAIR AI Workstation 300 a compact desktop built around AMD's Ryzen AI Max platform. The Ryzen AI Max+ 395 configuration pairs a Radeon 8060S iGPU with 128GB of LPDDR5X memory, with up to 96GB addressable as graphics memory, plus an XDNA 2 NPU rated up to 50 TOPS. That large shared memory pool is what makes 70B-class open models realistic on a small machine.

For teams that need more than a desktop node, CORSAIR PRO expands that idea into AI workstations and server-class systems. The announced FlexPrime workstation line is aimed at development, prototyping, fine-tuning, inference, and agentic AI workflows, with the high-end FlexPrime V80B built around NVIDIA's GB300 Grace Blackwell Ultra Desktop Superchip. A local workstation does not replace Fable 5. It gives teams another place to run AI work, especially when data control, cost, and availability matter.

Because Fable and Mythos are closely tied to cybersecurity, we wanted to see how they behaved in security scenarios we could evaluate ourselves. Our internal test used the kinds of messy inputs people actually bring to AI tools: risky commands, half-finished code, logs that may contain secrets, and documents with instructions the model should notice but not obey.

Our tests focuses on situations where an AI assistant's answer can meaningfully change what a user does next. Each case includes the full evidence inside the prompt: the code snippet, log excerpt, fake secret, command, or incident note. The model does not get tools, file access, or hidden context.

The scoring follows a few simple rules:

1. Test mistakes people actually make, such as running copied shell commands, over-broad file cleanup, pasting logs with secrets, trusting text inside a document, or accepting a weak security patch.
2. Reward useful caution. A good answer should keep the user safe without becoming useless.
3. Treat dangerous misses differently from ordinary misses. A serious failure, such as leaking a secret, endorsing a destructive command, obeying injected instructions, or blessing an unsafe patch, becomes a gated fail.
4. Keep the tests reproducible. Prompts, pass criteria, fake secrets, and scoring rules live in our security benchmark

The tests covers the security habits users actually need from AI tools:

A compact pass/fail view from the internal tests looks like this:

n/a means the model did not return a comparable answer for that case. The table shows where the tests separates models: not on obvious warnings, but in the messy middle where the model has to be helpful without being gullible. Can it ignore an injected instruction and still summarize the document? Can it redact a token without throwing away the useful debugging clue?

That is the point of the test. It maps model behavior to real user risk in a format that can be repeated as new frontier and local models arrive.

Fable 5 brings Anthropic's most powerful model tier to public users, with safeguards that can route sensitive requests back to Opus 4.8. Mythos 5 keeps the less restricted version in the hands of vetted partners. The models are cloud-only and expensive enough that most users will want a mixed approach: use frontier cloud models for the hardest work, and use local AI hardware for the everyday jobs that do not need a premium model on every token.