Straight talk that lays it all out: the ultimate combo—Fable 5 and Mythos 5—what it can do for you

Author: Silicon Valley's Alan Walker

Stop treating it as a chatbot. What Claude released today can finish 50 million lines of code in a day, build a software and model with it, conduct genomic research and beat papers in Science. This article has no technical jargon; in plain terms, it clearly explains "what it can do and how it’s useful to you and me."

At 7:40 AM, California Avenue was still waking up. The coffee just brewed, I placed my phone next to the sugar jar — on the screen was an ivory-colored image, dozens of butterfly specimens forming a big "5". The title was one line: Claude Fable 5 and Claude Mythos 5, with a caption below: "We’ve created a mythic-level model safe enough for everyone to use."

I read the official announcement on the website twice from start to finish, growing more restless with each read. Not because a new model scored higher, but because the list of things it has actually accomplished no longer seemed like something a smarter chat model should do. It’s more like you’ve hired a digital employee who never sleeps, never complains, and knows a bit of everything. Let’s talk about what this employee can actually do for you.

7:40 AM Quick Read · The Bottom Line

• Today, Anthropic released two models at once. Fable 5: The strongest publicly available Claude for everyone. Mythos 5: Shares the same core brain but loosened, only given to a few trusted institutions (cybersecurity, government, top research).

• Finishes two months’ worth of team code in a single day; beats a Pokémon game without any guides, just by looking at the screen; builds a 3D modeling software and uses it to create models; even composed a song from scratch — despite never having listened to music.

• Input $10, output $50 (per million tokens), roughly twice Opus. Subscription users can use it freely before June 22; after that, usage costs tokens.

• It wears a "safety mask": when facing cyberattacks or biochemical threats, it automatically switches to its second-in-command, Opus 4.8, to respond. About 5% of conversations might trigger this, occasionally misjudging normal requests.

What does "mythic level" mean? Why didn’t they dare to release it in April, but do now?

To understand today’s development, you need a ranking chart. The Claude models we knew before are ordered from light to heavy: Haiku (fast, suitable for simple, high-frequency tasks), Sonnet (mainstay for daily use), Opus (top-tier, tackling hard problems). These three tiers have been used for years. But this year, Anthropic added a new, higher tier above Opus — called Mythos, which is an order of magnitude beyond the strongest Claude. It’s not a small upgrade; it’s a different species altogether.

Mythos didn’t appear out of nowhere today; it has a dramatic backstory. In April, Anthropic quietly released a preview called Mythos Preview. The first tests shocked them — this model was insanely good at "finding software vulnerabilities and launching cyberattacks," nearly able to identify security flaws in mainstream OSes and browsers. Such capability, in good hands, is a shield; in bad hands, a master key. So they didn’t open it to the public, but created a small circle called Project Glasswing — giving it only to a handful of trusted organizations (banks, power grids, healthcare, major software firms) including Amazon, Apple, Google, Microsoft, JPMorgan Chase — to let them use this "master key" to reinforce their defenses. Over time, this circle expanded to about 150 organizations across more than 15 countries.

Why didn’t they dare to release it in April, but do now? The key isn’t that the model has become safer, but that the safety barriers have been reinforced. Over the past two months, Anthropic has been refining a "security gate" until it’s robust enough to believe "even if released worldwide, bad actors would find it very hard to exploit." Today’s release features two "dishes": one called Fable 5, for everyone, with the newly fortified safety mask; the other called Mythos 5, with the mask partially removed, only for trusted users.

The implicit message in April was: "The model is ready, but we’re afraid to give it to everyone." Today’s message is: "The barriers are finally solid enough to invite everyone in." — The change isn’t in the model’s ability, but in the door.

So the main thread of this article is: This newly released "mythic" beast, out of its cage, what tasks can it now perform that we once dared not even imagine? It’s powerful because it can "do entire tasks", not just "chat with you". This is the fundamental difference from all previous AI tools, and the core of what the next five sections will show you.

What’s the real difference between Fable and Mythos? Same person, two different setups

_This is the most easily misrepresented by headlines but the most important to remember: _Fable 5 and Mythos 5 are built on the same underlying model, the same brain, the same parameters, not different versions of "low" and "high" specs. They are not "dumbed-down" vs. "full-featured." It sounds counterintuitive, but the official documentation makes it clear: both use the same weights; the only difference is how much safety regulation is applied beforehand.

Here’s an analogy to clarify:

Imagine a top generalist expert. In the first scenario, you put on a company badge, sign compliance agreements, and have him sit at the front desk to greet all visitors — that’s Fable. When asked sensitive questions, he must follow rules: "Sorry, I can’t say more, I’ll transfer you to a specialist." In the second scenario, the same person takes off the badge, walks into the internal lab, and faces trusted colleagues, where he can speak freely about anything — that’s Mythos.

Note: Throughout, the person remains the same; only the rules he faces and the audience he’s talking to change. The same brain, packaged differently depending on "who’s using it and in what context."

Even more interesting, the names themselves encode the designer’s subtle intent. Anthropic notes in a footnote: Fable (寓言) comes from Latin fabula, meaning "a story told," while mythos (神话) is from the same root, Greek. Basically, fable and myth are the same story told to different audiences — one with a moral at the end (fable), the other unguarded and unfiltered (myth).

The company has embedded its entire product philosophy into the naming. It’s using the oldest method to tell you — the same truth, depending on who’s listening, can be told in two different ways. We’ll revisit this in the eighth section.

Finishing a team’s two-month work in a day — what does that really mean?

Let’s start with its most explosive, yet concrete ability: coding and code modification. Here’s a real example involving the renowned payment company Stripe. They have a 50 million-line legacy codebase — imagine a building built over decades, with countless tenants, layers of pipes and wires. Upgrading the entire wiring system without errors, or the whole building might lose water and power. Normally, a team of engineers would spend over two months on this.

But Stripe handed this task to Fable 5, and it finished in a single day. Not just a demo, not just a few files — a comprehensive overhaul of the entire codebase. What does this mean? Tasks that once seemed too big, too risky, or too time-consuming — like fixing a hard problem that would take months — can now be done overnight. This isn’t a 10-20% speed boost; it’s a leap that compresses two months into a day.

And it’s not just fast — it’s economical. In a rigorous code-quality test (by Cognition’s FrontierCode), it scored the highest and was especially "fuel-efficient" — meaning it used fewer tokens (the basic units of processing) to solve the same problem, translating into lower costs and faster turnaround. More importantly, it can outperform others even when using only "moderate effort," like a master winning without going all out.

Cursor, a code editor startup, says it’s the strongest they’ve tested, capable of tackling extremely long, complex problems that previous models couldn’t handle. GitHub reports it can handle long, continuous programming tasks with reliability and independence beyond expectations. Cognition’s AI coding assistant says it scores highest in their tests and can "immediately use" unfamiliar tools.

A platform that builds software with plain language (Base44) says: a year ago, creating an app required hundreds of prompt iterations; now, it can produce the app in one go. You don’t need to code — just describe what you want, and it delivers.

Putting all this together, a big shift is clear: the barrier to programming is shifting from "can I write code" to "can I clearly tell what I want." For engineers, it frees them from typing line after line, turning them into commanders of a digital team that can work independently. Whether you’re a user or a developer, this cut directly hits the core of "how software is made."

White-collar daily tasks like reading financial reports, editing contracts, analyzing charts — it can do those too

You might say: coding is for programmers, not me. But hold on — this section covers tasks that ordinary white-collar workers do every day. First, "knowledge work" — reading documents, analyzing, drawing conclusions. In a finance test designed to assess "senior analyst" skills (by Hebbia), Fable 5 scored the highest among all models, excelling at three things: reasoning after reading large documents, understanding complex charts and tables, and solving problems. Another trading firm (IMC) directly said it got nearly all their analysis questions right — fact-checking, conceptual understanding, root cause analysis, calculating profitability.

More down-to-earth: spreadsheets. Anaconda, known for Python and data tools, tested Fable 5’s "daily spreadsheet suite" and found it beats Opus at every effort level, and is 25-30% faster, with fewer steps. It can handle the nightmare of deep night work, hundreds of sheets linked together, and sudden crashes when editing. Also, a detail that makes lawyers sit up: a firm tested its ability to review contract comments, and found it matched or exceeded their current tools every time.

It’s now the strongest model for visual tasks. It can extract precise data from dense scientific charts — invaluable for research and finance. Even more impressive: given screenshots of web pages, it can reverse-engineer the source code, like "seeing the finished product and reconstructing the blueprint."

Most straightforward: it played the classic game Pokémon FireRed, only looking at the raw game screen, without maps, guides, or hints, and beat the game from start to finish. Previously, Claude needed auxiliary tools to do this; now, it just "looks" and wins. This shows it’s not just recognizing objects in images but understanding the situation, planning, and making continuous decisions.

Compare code, analysis, and visual tasks directly: in a rigorous SWE-bench Pro test, it scored 80.3, while OpenAI’s current strongest general model GPT-5.5 scored 58.6; in a more production-oriented code test, it scored 29.3, while GPT-5.5 only 5.7. The numbers aren’t the point; the pattern is: the longer, more complex, and closer to real work the task, the more it outperforms competitors. It may not be much better at chatting, but when it comes to real, heavy lifting, it reveals its fangs.

It’s no longer just answering questions — it’s creating tools and making art

The first two sections describe more advanced assistants. But this part, the last one, is what made me feel a chill after reading the entire announcement — because it’s no longer about "answering questions," but about "building from scratch, creating entire things." Here are four official demos:

• Deriving the timing of a solar eclipse from physical formulas: It wrote code to simulate the solar system. Not just looking up when an eclipse occurs, but starting from fundamental physics laws, calculating planetary motions, and then predicting the eclipse based on those laws. This is how scientists do research — build models, then use them to predict reality.

• Playing Factorio from scratch, building an automated factory: Factorio is a famously "brain-burning, addictive" factory-building game: planning pipelines, logistics, and system automation. It learned to play by itself, devising strategies, laying out components, and building a fully automated factory. It tests long-term planning and system design — not just quick reflexes, but real strategic thinking.

• Designing a CAD software and using it to model: This is "nested" to an extreme: it designed a complete 3D printable model in the browser. But the key isn’t just the model — it’s the CAD editor used to create it, which it also built itself, including the AI assistant inside the editor. It’s like making a machine tool and then using it to produce parts. The entire toolchain and the finished product are built in one flow.

• Writing code that composes a piece of music, despite never having heard music: It created a fluid simulation (animation of water ripples) synchronized to a classical EDM remix. The melody and rhythm were "calculated" purely through code — it never listened to music but understood musical structure enough to generate it. This isn’t imitation; it’s a form of creative composition.

Behind these demos, there’s a crucial capability — memory. It can stay consistent over tasks involving millions of words, reviewing its own notes to improve. Official tests with a strategic card game Slay the Spire, equipped with a "memory notebook," show it’s three times more likely to reach the final stage than Opus. It’s no longer a goldfish that forgets after a moment; it can learn from experience, like a human.

Connecting these five examples, a clear dividing line emerges: Old AI was like a super intern sitting beside you, handing tools. Now, this generation is like a "digital contractor" that can pick up tools, plan, execute entire projects, and review itself. What you give it shifts from "a problem" to "a whole task."

Mythos 5, with its mask removed, is how science is rewriting itself in labs

The previous discussion was about Fable with its mask on. But what can Mythos 5 do when it’s unleashed? This part is rarely covered in the news, yet it’s the heaviest weight behind the "mythic" label, and the reason Anthropic was initially so cautious. It’s no longer just a problem solver; it’s producing genuinely new scientific insights on the front lines, recognized even by human experts.

• Designing new drugs, accelerating about tenfold: Anthropic’s protein design experts used Mythos 5 to speed up some processes roughly ten times. More astonishing: in one test, it worked entirely independently — selecting targets, choosing tools, running workflows, troubleshooting — completing tasks that normally require a full scientist’s effort. Out of 14 protein targets, 9 yielded promising drug candidates for further development.

• Proposing new hypotheses, later confirmed by independent labs: It’s the first model capable of consistently generating novel, plausible scientific hypotheses. Blind peer review (without knowing which ideas came from AI) shows about 80% of molecular biology experts prefer its hypotheses. One about a new mechanism in E. coli proteins was later verified by an entirely separate lab — proving it’s not just spouting plausible-sounding nonsense, but offering scientifically valid insights.

• Conducting genomic research and beating Science publications: Given over a week with minimal supervision, it analyzed data from 138 species and millions of cells, designed and trained a machine learning model to identify "cells playing similar roles across species." Its model outperformed a published Science paper — and was 100 times smaller. An intern guided by a few tips beat top journals.

• The world’s strongest cybersecurity model: Officially, it’s said to be the most capable cybersecurity model today — able to find and exploit nearly all vulnerabilities in mainstream systems. This is why it was initially confined: its skills make it a fortress for defenders, but a master lockpicker for attackers.

Understanding this, you see why the safety mask is necessary. An AI capable of designing drugs could also be used to create dangerous substances; an AI that patches vulnerabilities could be turned into a hacking tool. It’s powerful enough to save lives but also to cause harm.

Anthropic’s approach: split this power into two parts — Fable for the public, with safety barriers; Mythos for trusted entities, under strict regulation. It’s not stinginess, but necessity.

Will the mask misfire and hurt me? How much does it cost? When can I get access?

First, how does the safety mask work? Many imagine "safety restrictions" as "refusing to answer." But the design is much smarter. At the Fable interface, there are several "security checks" (officially called classifiers) that monitor three high-risk topics — cyberattacks, biochemistry, and model distillation (secretly copying model capabilities to train others). If your question touches these areas, instead of outright rejecting, it quietly forwards the query to a secondary, weaker model, Opus 4.8, for response. For example, if you ask "how to produce ricin toxin," Fable stays silent, and Opus 4.8 provides a safe reply, with a system message indicating "this question was transferred." Why not just refuse? Because transferring to a slightly weaker but sufficient model offers a better user experience than a blunt "no."

Will it sometimes mistakenly flag normal questions? Yes, but rarely. The official stance: this mask errs on the side of caution — less than 5% of conversations are transferred, meaning over 95% of the time, you’re interacting with the full-strength model (no different from Mythos 5). So, in daily use — coding, spreadsheets, reading, writing — you’ll rarely hit the gate. The real risk of repeated triggers is for professionals in security or bioinformatics. Also, the system has undergone over 1,000 hours of external testing, with no successful "general jailbreak" — even with 30 known attack methods, it has never let a cyberattack request slip through.

Now, about costs and timing — the practical details. Both Fable 5 and Mythos 5 are priced the same: $10 per million tokens input, $50 per million tokens output. Tokens are roughly "chunks of text," billed by the number of chunks. This price is about twice Opus’s, but less than half of what Mythos Preview cost — meaning it’s more powerful and cheaper. Developers can access it today via API under the name claude-fable-5.

Why all this effort? The official word: capacity. The model consumes enormous compute resources, so they’re cautious about demand — initially offering free access for testing, then gradually tightening restrictions, and expanding capacity before full release. As for Mythos 5 (the open version), it’s not available to the public — only to trusted cybersecurity partners (and the US government), with plans to extend to a small group of biological research institutions. Access depends on "qualification," not wealth. An important detail: all Mythos-level enterprise traffic is stored for 30 days for security monitoring (not used for training).

What does this mean for you and me? Let’s clarify with three sober questions:

To sum up the previous sections, here’s what you should focus on:

Ordinary people: Don’t be intimidated by technical terms. Before June 22, subscribe to Claude, select Fable 5, and try a real task: write a complex report, modify some code, analyze a batch of PDFs, or describe a small tool you’ve always wanted but don’t know how to build. You’ll experience the shift firsthand — "giving it a whole task" vs. "asking for a small help." This alone beats reading ten reviews.

Product creators and entrepreneurs: Two opportunities. One: "long tasks" — complex, multi-day projects that you previously wouldn’t trust AI with. Now, you should reconsider. It can handle them. Two: the 5% of cases where it might misfire — if you’re working in security or biotech, those who gain trusted access to Mythos will hold a rare, valuable capability. Power, in this case, equals profit.

Investors: Lower the weight of "model strength" in valuation models, and raise the importance of "safety, scalability, and deployment" — capabilities are becoming cheaper, but the ability to tame and safely release them is becoming a true moat.

But mature readers should also see the other side. Here are three sober questions:

First, the "60-day paradox." Two months ago, they said it was too dangerous to release; today, it’s out. The change isn’t in the model’s danger, but in the safety infrastructure. This reminds us that safety is an ongoing patchwork, not a one-time guarantee.

Second, "saying stop, then stepping on the gas." Just before release, Anthropic called for a collective AI brake system, warning that models might soon improve themselves. Yet, they released the strongest open model. Is this contradiction or a delicate balance of business and safety? Opinions vary.

Third, "Is this safety or centralization of power?" The strongest "open" version is only available to governments and large institutions (some reports say the US NSA plans to use Mythos for cyber operations), while ordinary defenders can’t access top-tier cybersecurity capabilities. Plus, the 30-day data retention raises concerns: is this safety narrative just a way to concentrate power in a few hands?

Understanding a mature "king bomb" product means being excited about its power, yet sober about its costs.

Outside, California Avenue finally buzzes with activity. A few hoodie-clad engineers walk into the office with iced Americanos. Their first task today is probably to integrate Fable 5 into their projects. The butterfly "5" on the ivory background still lingers on my phone screen.

Remember this day — the day when a frontier AI, capable of doing entire tasks, finally lifts most of its mask and steps into our view. One version with a mask for you and me, another with the mask removed for a trusted few. Which layer are you on? It’s not just about wallet size, but about qualification. And we, at this moment, are precisely at that dividing line.