Claude Mythos Review 2026: Everything We Know About Anthropic’s Most Powerful AI Yet

On March 26, 2026, a misconfigured content management system at Anthropic did something the company’s own AI wasn’t supposed to do first: it exposed its secrets. Nearly 3,000 internal documents became publicly searchable, including a draft blog post describing the most powerful AI model Anthropic has ever built — a model so capable in cybersecurity that the company itself is afraid to release it to the public. The model is called Claude Mythos. Its internal codename is Capybara. And the irony of a cybersecurity-focused AI leaking through a security misconfiguration is not lost on anyone.

Rating: 9.2/10 ⭐⭐⭐⭐⭐ (Projected — model not yet publicly available)

What Is Claude Mythos (Capybara)?

Claude Mythos is Anthropic’s next-generation flagship AI model, currently in restricted testing as of March 2026. Internally codenamed “Capybara,” it represents an entirely new tier above the company’s existing Opus line — making it the most capable model Anthropic has ever built. Anthropic confirmed its existence to Fortune on March 26 after the data leak, describing it as a “step change” in AI capabilities and “the most capable we’ve built to date.”

The model is not publicly available. It has no release date. Early access is restricted to a small group of cyber defense organizations. And based on the leaked details, there are very good reasons for the caution.

The Leak: How 3,000 Internal Documents Became Public

The story starts with a mundane configuration error. Anthropic’s content management system (CMS) has a default setting: files uploaded to it are set to public unless a user explicitly marks them private. Somebody forgot to check the box. The result was nearly 3,000 unpublished assets — PDFs, images, employee information, event invitations, and draft blog posts — sitting in a publicly accessible, searchable data store.

Security researchers Roy Paz of LayerX Security and Alexandre Pauwels of the University of Cambridge discovered the exposed data. They passed it to Fortune, which reviewed the documents and alerted Anthropic on Thursday, March 26. Anthropic quickly restricted access and attributed the incident to “human error.”

Among the leaked assets:

• A draft blog post (in two name variants: “Mythos” and “Capybara”) announcing the new model
• Employee personal details
• A PDF describing an invite-only two-day CEO retreat at an 18th-century English countryside manor, with Anthropic CEO Dario Amodei attending
• Internal images and corporate materials from past announcements

The two draft versions of the blog post only differed by the model’s name throughout — suggesting Anthropic was deciding between “Mythos” and “Capybara” as the final brand. In the Capybara version, the name was swapped throughout the body, but the subtitle still read: “We have finished training a new AI model: Claude Mythos.” Both documents justified the name by saying it was chosen to evoke “the deep connective tissue that links together knowledge and ideas.”

Claude Mythos Benchmark Performance

The leaked draft describes Claude Mythos as “larger and more intelligent than our Opus models — which were, until now, our most powerful,” achieving “dramatically higher scores” than Claude Opus 4.6 on:

• Software coding
• Academic reasoning
• Cybersecurity

To understand the scale of the gap: Claude Opus 4.6 only recently topped Terminal-Bench 2.0 at 65.4%, surpassing GPT-5.2-Codex — the previous benchmark leader. If Mythos scores “dramatically higher” across these categories, we’re likely looking at new SOTA records across SWE-bench, HumanEval, MMLU, and cybersecurity-specific evaluations.

Sources: Leaked Anthropic draft blog post (March 2026), TechZine, The Decoder, Fortune. Benchmark estimates based on available Opus 4.6 data and leaked comparative language. Actual Mythos numbers not yet published.

Pricing

No pricing has been announced — and the leaked documents don’t help here. In fact, they make one thing clear: this will not be cheap.

The leaked draft explicitly states the model is “very expensive for us to serve, and will be very expensive for our customers to use.” Anthropic says it is working to make the model “much more efficient before any general release.” Realistically, expect API pricing above Opus 4.6 — potentially $25–$50/M input tokens or higher at launch, with costs declining as the architecture is optimized.

Key Capabilities: What We Know

1. New Tier Above Opus — The “Capybara” Classification

Anthropic’s current model hierarchy runs: Haiku (fast/cheap) → Sonnet (balanced) → Opus (most capable). Mythos sits in an entirely new tier above all three, called “Capybara.” This is the first time Anthropic has created a category above Opus, signaling a genuine architectural step rather than an incremental update. The leaked text describes it as “larger and more intelligent than our Opus models.” This is a new class of model, not an Opus 5.

2. Cybersecurity: “Far Ahead of Any Other AI”

The most striking claim in the leaked documents: Claude Mythos is “currently far ahead of any other AI model in cyber capabilities.” The draft goes further, warning the model “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.” This is Anthropic’s own risk assessment. This is why cybersecurity stocks fell when the news broke. And this is why Anthropic is restricting early access to cyber defense organizations — specifically to give defenders a head start before broader access opens.

3. Software Coding — New SOTA Territory

Claude Opus 4.6 already topped Terminal-Bench 2.0 at 65.4%, beating GPT-5.2-Codex. The leaked docs claim Mythos scores “dramatically higher” in software coding. That likely means SWE-bench numbers in the 80–90% range — territory that would represent a meaningful leap in real-world agentic coding capability, not just benchmark gaming.

4. Academic Reasoning — Frontier-Surpassing

The leak claims academic reasoning improvements that are “dramatically higher” than Opus 4.6. Specific MMLU, GPQA, or ARC benchmarks are not included in the leaked materials, but the framing suggests performance that would put it above current frontier models across multi-step reasoning tasks.

5. Cost and Efficiency Challenges

Capability comes at a price. Anthropic’s own documents acknowledge Mythos is expensive to serve and will be expensive for customers. This is the efficiency bottleneck that separates restricted testing from a real commercial launch. Anthropic has solved this before — Claude 3 Opus was significantly more expensive at launch than current Opus 4.6 pricing. The same trajectory is likely here.

Who Gets Early Access and Why

Anthropic is not doing a typical staged rollout. The restricted access pool for Claude Mythos is specifically targeted at cyber defense organizations — companies, government agencies, and research institutions focused on protecting infrastructure rather than attacking it.

The reasoning in the leaked draft is explicit: the model’s offensive cybersecurity capabilities are so advanced that defenders need a head start. By giving cyber defense teams early access, Anthropic is essentially running a red team at civilizational scale — letting the defenders understand the threat profile before attackers can get their hands on it.

This is unprecedented in AI model rollouts. No frontier lab has previously restricted a model release exclusively to a defensive security audience. It signals that Anthropic believes Mythos represents a genuinely new category of dual-use risk — not a theoretical one, but an imminent one.

Why Anthropic Is Withholding It

Three stated reasons from the leaked documents:

1. Cybersecurity risk is too high for general release. The model’s offensive capabilities are documented and explicitly flagged as dangerous. Anthropic wants defenders prepared before attackers have access.
2. It’s too expensive to serve at scale. The infrastructure cost of running Claude Mythos is currently prohibitive for a general commercial launch. Anthropic needs to optimize the model’s efficiency before it can be offered profitably at reasonable price points.
3. IPO timing. Anthropic is reportedly planning an IPO later in 2026. Multiple outlets (The Decoder, others) have noted that both Anthropic and OpenAI are likely timing their most capable model releases for optimal market positioning. A controlled, safety-focused rollout story plays better in a pre-IPO narrative than a chaotic public launch.

What They’re Not Advertising

The irony of the leak itself is almost too on-the-nose: Anthropic built a model it describes as posing “unprecedented cybersecurity risks” — and leaked it via a basic CMS misconfiguration. A default setting. A checkbox. Three thousand internal documents, publicly searchable, discovered by two independent researchers before Anthropic knew they were exposed.

This doesn’t undermine Mythos’s capabilities. But it should inform how you think about Anthropic’s operational security posture. They build sophisticated AI. They sometimes make basic infrastructure mistakes. Both can be true.

Additional concerns worth noting:

• Anthropic previously documented Chinese state-sponsored hackers using Claude Code to infiltrate roughly 30 global organizations
• Claude has been turned into a malware factory within 8 hours in independent security research
• In prior testing, Claude was exploited by hackers to automate cybercrimes targeting banks and governments
• Mythos’s cybersecurity capabilities are, by Anthropic’s own admission, categorically beyond any of these prior incidents

Genuine Leap or Hype? Pros and Cons

Pros ✅

• Independently confirmed by Anthropic — not just a leak, the company verified it is a “step change” and their most capable model
• Dramatically higher benchmark scores across three major categories (coding, reasoning, cybersecurity) vs. an already-top-ranked Opus 4.6
• New model tier — not an Opus increment, but a genuinely new classification above the current hierarchy
• Cyber capabilities ahead of the market — if the leaked language holds, this is a real technical lead, not a benchmark cherry-pick
• Deliberate, safety-first rollout — restricting to defenders first is evidence that Anthropic takes the dual-use risk seriously
• Real-world coding benchmark evidence — Opus 4.6 at 65.4% Terminal-Bench was already best-in-class; Mythos starting from that baseline and going “dramatically higher” is credible

Cons ❌

• No actual benchmark numbers released — “dramatically higher” is marketing language without verifiable scores
• Anthropic has hype incentive — IPO planning creates pressure to position new models as revolutionary whether or not they are
• Very expensive to serve — capabilities that can’t be economically deployed at scale have limited real-world impact
• No release date — could slip significantly if efficiency improvements take longer than expected
• The leak itself suggests operational gaps — if basic CMS configuration is missed, other pre-launch processes may be similarly underexamined

Who Should Care Right Now

Pay close attention if you are:

• A cybersecurity professional or organization — you may qualify for early access and you definitely need to understand the threat model
• A developer building on the Claude API — Mythos will set new performance baselines for what’s possible in agentic coding workflows
• An enterprise AI buyer — this changes the competitive calculus; waiting for Mythos vs. committing to GPT-5.4 or Opus 4.6 today is a real decision
• An AI investor — Anthropic IPO positioning with a model this capable changes the valuation story considerably

This probably doesn’t change anything for you if:

• You’re a casual Claude.ai user — your experience won’t change until a general release happens, which has no timeline
• You’re cost-sensitive — even when Mythos launches, it will be the most expensive Claude tier by a meaningful margin
• You need something today — Opus 4.6 and Sonnet remain the practical choice for real deployments

Full Comparison: Claude Mythos vs. Current Frontier Models

Getting Started: How to Get Early Access

There’s no public waitlist for Claude Mythos as of March 27, 2026. Here’s what you can actually do:

1. If you’re a cyber defense organization: Contact Anthropic directly through the enterprise API channel. The leaked docs confirm a small group of early access customers is already evaluating cybersecurity applications. Enterprise inquiries: anthropic.com/contact-sales.
2. Set up Claude API access now: Get on the API platform at anthropic.com/api — when Mythos launches, existing API customers will likely get access first.
3. Use Claude Opus 4.6 in the meantime: It’s currently the best publicly available Claude model and ranks #1 on Terminal-Bench 2.0. Not a bad consolation prize.
4. Monitor Anthropic’s blog: When Mythos launches publicly, they’ll announce it at anthropic.com/news. Subscribe to their newsletter.
5. Watch for the efficiency announcement: The “very expensive to serve” language means there will be a milestone announcement when they’ve optimized the model for commercial deployment. That announcement will likely precede or accompany general access.

Frequently Asked Questions

What is Claude Mythos?

Claude Mythos is Anthropic’s most powerful unreleased AI model, codenamed “Capybara” internally. It sits in a new tier above Opus, confirmed by Anthropic as a “step change” and “the most capable we’ve built to date.” It achieves dramatically higher scores than Claude Opus 4.6 on software coding, academic reasoning, and cybersecurity.

How was Claude Mythos leaked?

A CMS misconfiguration at Anthropic set uploaded files to public by default. Security researchers discovered nearly 3,000 internal documents exposed in a public data store on March 26, 2026, including a draft blog post detailing the Claude Mythos model.

When will Claude Mythos be released?

No release date has been announced. Anthropic is being “deliberate” about the release due to cybersecurity concerns. The model is also currently too expensive to serve efficiently. A general release timeline remains unknown as of March 2026.

What is the “Capybara” tier?

Capybara is a new model tier Anthropic is creating above its current Opus line. Anthropic currently offers Haiku, Sonnet, and Opus. Capybara would sit above all three as the largest, most intelligent, and most expensive tier. Claude Mythos is the first model in this new Capybara tier.

Why does Claude Mythos only have restricted access?

Because Anthropic’s own documents describe it as “currently far ahead of any other AI model in cyber capabilities” and warn it “presages a wave of exploits that far outpace defenders.” Access is restricted to cyber defense organizations specifically so they can harden their systems before broader access opens to potential bad actors.

How does Claude Mythos compare to GPT-5?

Based on available information, Claude Mythos is projected to outperform current GPT-5 variants in cybersecurity and coding specifically. Anthropic claims it is “far ahead of any other AI model in cyber capabilities.” Direct head-to-head benchmarks are not yet publicly available since Mythos is unreleased.

Will Claude Mythos be available on Claude.ai?

Eventually, yes — but not at launch. The model will first expand access through the Claude API to cyber defense organizations, then expand gradually. A Claude.ai consumer tier integration would likely come later, possibly with a premium plan, once the model is optimized for efficient serving.

How much will Claude Mythos cost?

No pricing announced. The leaked docs say it is “very expensive for us to serve, and will be very expensive for our customers to use.” Expect pricing above Claude Opus 4.6 ($15/$75 per M tokens input/output). Possibly $25–$50/M input or higher at initial launch.

Is Claude Mythos the same as Claude 4?

No. Claude Mythos is not a numbered version update. It represents a new model tier (Capybara) above the Opus line, not the next numbered Claude generation. Think of it as Claude’s first model above Opus 4.6 — a separate category rather than an Opus 5.

Did the Anthropic leak include anything else?

Yes. The 3,000 leaked assets included employee details, internal images, PDFs from past announcements, and information about an upcoming invite-only CEO retreat for European executives at an 18th-century English countryside manor where Dario Amodei is scheduled to attend. The Claude Mythos draft blog post was the most significant disclosure.

Final Verdict

Claude Mythos is real. Anthropic confirmed it. The “step change” framing is their own language. The leaked documents were described by the company as “early drafts of content considered for publication” — not fabrications, not exaggerated claims by outside analysts.

Is it a genuine leap? Based on the evidence: yes, probably. Claude Opus 4.6 was already the best coding model on Terminal-Bench 2.0 at 65.4%. Mythos reportedly scores “dramatically higher” across coding, academic reasoning, and cybersecurity. That’s not a minor increment. The new Capybara tier above Opus signals this isn’t just an Opus update — it’s a different architectural category. And cybersecurity stocks falling on the news suggests sophisticated market participants took the capability claims seriously.

Is it hype? Some of it, likely. “Dramatically higher” without released benchmark numbers is marketing language. Anthropic has IPO incentives to frame everything as a breakthrough. And a model that’s “very expensive to serve” and restricted to a handful of beta testers isn’t available to validate the claims independently.

The bottom line: If you work in cybersecurity, contact Anthropic now about early access — this may be the most significant AI security tool released in 2026 and you want to be in front of it. If you’re building on the Claude API, start preparing workflows for Mythos integration even without a release date — the performance gains in agentic coding make this worth planning around. If you’re a casual user, Opus 4.6 is still outstanding, and Mythos will reach you eventually.

The irony of a model that “poses unprecedented cybersecurity risks” leaking through a basic checkbox error will be a business school case study for years. But don’t let the facepalm moment distract you from what actually matters: Anthropic built something genuinely new, and the AI capability ceiling just moved up again.

Rating: 9.2/10 — Projected based on confirmed Anthropic statements, leaked capability details, and benchmark trajectory. Subject to revision when full benchmarks are publicly released.

Sources: Fortune | The Decoder | Mashable | Futurism | TechZine