Claude Mythos: What It Is, Launch & Why It Was Pulled

Last updated: June 14, 2026 · Originally published March 28, 2026 · Updated for the June 9 Mythos-class launch (Claude Fable 5 + Mythos 5) and the June 12 suspension under a US government export directive.

What is Claude Mythos in one sentence?

Claude Mythos is the codename for Anthropic’s Mythos-class tier, whose existence was first confirmed on March 27, 2026 through an accidental leak of approximately 3,000 internal CMS assets — including two draft blog posts, one branded “Mythos” and one branded “Capybara.” Anthropic confirmed the model to Fortune the same day, describing it as “a step change” and “the most capable we’ve built to date.” It moved from restricted preview to a public launch on June 9, 2026 (as Claude Fable 5 and Mythos 5) before Anthropic suspended both on June 12 under a US-government export directive.

The original draft framed Mythos as the first model in a brand-new Capybara tier sitting above Opus. That branding has not survived, but the capability tier has. Anthropic shipped Opus 4.7 on April 16, then Opus 4.8 on May 28, both inside the existing Opus line. In the Opus 4.8 announcement, the company explicitly separated generally available Opus from the restricted Claude Mythos Preview used in Project Glasswing and said Mythos-class models require stronger cyber safeguards before general release.

That makes the June 2026 reading sharper than the March leak: Capybara was likely a rejected marketing tier; Mythos is a real restricted capability class; Opus 4.8 is the public bridge that absorbs part of the Mythos work while Anthropic finishes safeguards for the rest.

What has happened since the leak? (June 2026 timeline)

The original March coverage answered what the leak said. The harder question, nine weeks later, is what actually shipped. Here is the verified timeline.

• March 26–27, 2026 Cybersecurity researchers Roy Paz (LayerX) and Alexandre Pauwels (University of Cambridge) discover ~3,000 unsecured Anthropic CMS assets. Fortune publishes the exclusive. Anthropic confirms the model’s existence and secures the data store.
• April 16, 2026 Anthropic releases Claude Opus 4.7 — the model ID is claude-opus-4-7, not claude-capybara-mythos. The release sits inside the existing Opus tier, contradicting the “Capybara above Opus” framing of the leaked draft. Opus 4.7 delivers ~3× SWE-bench gains over 4.6, an xhigh reasoning effort level, and a dramatic XBOW vision jump (54.5% → 98.5%).
• April 16, 2026 Anthropic’s own Opus 4.7 announcement references a “Claude Mythos Preview” — confirming the codename survived internally but is reserved for non-public access. Mythos is described as “more broadly capable but restricted.”
• May 28, 2026 Anthropic releases Claude Opus 4.8, the new public flagship. It improves coding, agentic work, professional knowledge tasks, honesty, and alignment at the same regular price as 4.7. The release also launches effort control, cheaper fast mode, Messages API system-entry updates, and Claude Code dynamic workflows.
• May 28, 2026 The same Opus 4.8 announcement says a small number of organisations are using Claude Mythos Preview through Project Glasswing for cybersecurity work, and that Anthropic expects to bring Mythos-class models to all customers “in the coming weeks” once stronger cyber safeguards are ready.
• June 9, 2026 Anthropic launches the Mythos-class tier: Claude Fable 5 (the first publicly accessible Mythos version) and Claude Mythos 5 (restricted). Anthropic calls Fable 5 the most capable model it has ever made generally available, with hard safety limits that block high-risk cyber/bio/chem requests and fall back to Opus 4.8.
• June 12, 2026 Anthropic disables both Fable 5 and Mythos 5 for all customers following a US-government national-security export directive that suspends access by any foreign national. Unable to filter foreign nationals from US users in real time, Anthropic shut both models down entirely; Opus 4.8 and all other models stayed online. Full story: Claude Fable 5 and Mythos 5: launched, then pulled.
• June 14, 2026 (today) Mythos-class access remains suspended. The launch confirmed the leak’s core claim — Mythos is a real tier above Opus — but a durable general release now depends on resolving the export-control constraint, on top of the EU AI Act GPAI regime. Opus 4.8 stays the model you can actually build on today.

What was Claude Mythos described as in the leak?

Per the two leaked draft blog posts (one Mythos-branded, one Capybara-branded — both dated for an unannounced release), Mythos was characterised as a fourth product tier above Anthropic’s existing Haiku → Sonnet → Opus hierarchy. Anthropic chose the name Mythos to evoke “the deep connective tissue that links together knowledge and ideas.” The Capybara version of the same draft kept nearly identical body text — including a telling inconsistency where the subtitle still read “We have finished training a new AI model: Claude Mythos,” even in the Capybara-branded version. The company was clearly deciding between two names for the same model when the leak occurred.

The leaked drafts described three capability areas with “dramatically higher” scores than Opus 4.6: software coding, academic reasoning, and cybersecurity. Anthropic did not publish absolute benchmark numbers in the drafts, but the company described the model as “currently far ahead of any other AI model in cyber capabilities” — and warned that it “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.”

An Anthropic spokesperson, replying to Fortune the day of the leak, said the company is “developing a general-purpose model with meaningful advances in reasoning, coding, and cybersecurity” — language that aligned with the leaked drafts but did not commit to the Capybara tier name or to a release date.

Mythos vs Capybara vs Opus 4.8: which name means what?

This is the single most common point of confusion in coverage of the leak. The terms refer to different things — model, tier, and shipped product. The table below maps what each name actually denotes as of June 2026.

The practical implication: if a vendor pitch in mid-2026 references the “Capybara tier,” treat it with significant scepticism. If it references Mythos, ask whether the access is actually Project Glasswing preview access, a forward-looking roadmap claim, or ordinary Opus 4.8 wrapped in Mythos language. As of today, Mythos has no active public API, no published pricing, no SLA, and no public model ID.

How did the leak happen?

The leak was a configuration failure, not a malicious breach. Anthropic’s content management system stored all website assets — blog posts, images, PDFs — in a central data store that was accessible without authentication. A configuration error meant that assets uploaded to this system were public by default unless explicitly marked private. Approximately 3,000 unpublished assets were discoverable to anyone with technical knowledge of how to query the system.

Among the exposed materials: two versions of the model-announcement draft (one Mythos-branded, one Capybara-branded), details about an invite-only European CEO retreat featuring Anthropic CEO Dario Amodei, internal images, and various PDFs. After Fortune informed Anthropic, the company secured the data and attributed the incident to “human error in the CMS configuration.”

The irony, widely noted at the time, was not lost on the AI community: a company whose latest model is touted for its cybersecurity prowess had its biggest announcement leaked through a basic CMS misconfiguration — the kind of vulnerability that AI coding tools, including Anthropic’s own Claude Code, increasingly make trivial to discover.

What did the benchmarks (claim to) show?

The leaked drafts gave no absolute numbers, only relative framing. Two months later, with Opus 4.7 and Opus 4.8 shipped publicly, we can read the leak with much more context. Anthropic’s own release sequence now provides the cleanest signal: Opus 4.8 is the best generally available model, but Anthropic still distinguishes it from the higher-capability Mythos Preview.

May 2026 capability snapshot

Even though Mythos has no public API, preview and partner data now lets us compare three layers: Mythos Preview, the new Opus 4.8 public release, and the older Opus 4.7 baseline.

Important: Anthropic’s Opus 4.8 benchmark table is partly image-rendered in the announcement, and many partner numbers are vendor-reported rather than independently audited. Treat the direction of travel as reliable; treat exact point differences as provisional until public leaderboards catch up.

Two things to flag about these numbers:

• Opus 4.8 is not Mythos, and Anthropic says so plainly. The company calls Opus 4.8 a modest but tangible improvement on 4.7, then separately says it is working on a new class of model with higher intelligence than Opus. That is the strongest public confirmation that Mythos is a tier above the current flagship, not a renamed 4.8 release.
• The new information gain is alignment, not just benchmark score. Opus 4.8’s most important public improvement is that it catches its own mistakes, flags uncertainty, and pushes back on weak plans. That matters for Mythos because the gated capability is cybersecurity-heavy: a stronger model that is also less sycophantic is easier to release safely.

How the May leaderboard maps to the March leak claims

One honest reading of both tables: Opus 4.8 absorbed another layer of the capability gains the leak attributed to Mythos, especially coding reliability and agentic collaboration. Mythos Preview still appears to be the higher-risk, higher-intelligence tier Anthropic is holding back for cyber-safety reasons. The key June update is not “Mythos disappeared”; it is the opposite: Anthropic launched Mythos-class models on June 9, then suspended them on June 12 — capability was never the question, compliance was.

Why are cybersecurity experts still concerned?

The cybersecurity dimension is what sets the Mythos leak apart from a routine model preview. According to the draft blog posts, Anthropic itself describes the model as “currently far ahead of any other AI model in cyber capabilities” — and warns that it “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.”

This creates a fundamental dual-use problem:

• Defensive side: a model with advanced vulnerability detection helps security teams find and patch weaknesses faster than ever. Anthropic’s existing Claude Code Security work, tested first on Opus 4.6 and then improved through Opus 4.7/4.8, has reportedly discovered over 500 high-severity exploits in open-source projects — in one widely cited case detecting a flaw in a PDF tool simply by analysing a developer comment in a changelog.
• Offensive side: the same capabilities — if accessed by malicious actors — could enable automated exploitation at a scale and speed that human defenders cannot match. This is not a hypothetical. In November 2025, Anthropic disclosed that a Chinese state-sponsored group exploited Claude’s agentic capabilities to infiltrate roughly thirty organisations, succeeding in a small number of cases. The June 12 export directive that pulled Mythos is the clearest sign yet that governments now treat frontier cyber capability as a controlled item.

The market reacted immediately to the original leak. Shares of major cybersecurity firms including CrowdStrike and Palo Alto Networks dropped more than 5% in the trading session that followed. The iShares Expanded Tech-Software Sector ETF (IGV) fell nearly 3%, and Bitcoin tumbled back to $66,000 as broader tech sentiment soured.

Nine weeks later, those stocks had largely recovered — and the June 12 suspension removed the most immediate competitive threat, since the Mythos-class models are no longer accessible to anyone. The defender’s dilemma has not gone away; it has simply moved behind an export-control wall while Anthropic resolves who is eligible to use a model this capable.

What does this mean under the EU AI Act?

The regulatory clock still matters. GPAI obligations under the EU AI Act have applied since August 2, 2025; the Digital Omnibus (provisional agreement May 2026) pushed the high-risk deadlines to December 2027 but left GPAI duties in force. So any durable European relaunch of a Mythos-class model is judged inside the GPAI regime from day one — on top of the US export constraint that suspended it in June.

Mythos, as described in the leak, would almost certainly be classified as a General-Purpose AI Model with Systemic Risk (GPAI-SR) under Regulation (EU) 2024/1689, Articles 51–55. That tier triggers extra obligations beyond the baseline GPAI rules:

• State-of-the-art adversarial testing (including red-teaming for cyber-misuse, the most relevant axis for Mythos given Anthropic’s own framing)
• Reporting of serious incidents to the EU AI Office within statutory windows
• Cybersecurity protections for model weights and training infrastructure
• Disclosure of energy consumption and training compute thresholds
• Compliance with the EU GPAI Code of Practice — voluntary today, but de facto required for a frictionless market launch

Anthropic’s staged rollout — starting with cyber defenders rather than open API access — is consistent with the Act’s risk-management philosophy. But the accidental public disclosure of the model’s existence before formal safety evaluations were complete, followed by a launch-then-pull driven by US export law, is exactly the kind of turbulence regulators designed the framework to surface. How Anthropic navigates Mythos’s eventual European launch will be a test case for the entire industry — and a useful preview of how the AI compliance workflow stack handles a frontier release inside the GPAI-SR regime.

What does this mean for developers building today?

For developers currently building on Claude’s API, the practical takeaways have shifted since March. The headline guidance is now clearer.

Don’t wait for Mythos. Build on Opus 4.8.

Opus 4.8 is the production-ready flagship today. It is the model Anthropic actually shipped and kept online, with a stable API contract, published pricing, platform availability, effort controls, and Claude Code dynamic workflows. The Mythos-class models (Fable 5, Mythos 5) launched on June 9 but were suspended three days later, so they have no dependable availability right now. Production teams should build against the Opus 4.8 contract and treat Mythos-class access as an upgrade path, not a dependency.

from anthropic import Anthropic

client = Anthropic()

# The model Anthropic actually ships today (June 2026)
response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=4096,
    thinking={
        "type": "enabled",
        "effort": "extra",          # use higher effort for hard agentic tasks
        "budget_tokens": 32000       # task budget cap
    },
    # Opus 4.8 also supports system updates inside messages for long tasks.
    messages=[{"role": "user", "content": prompt}]
)

# Speculative — DO NOT ship code that depends on this ID
# response = client.messages.create(
#     model="claude-capybara-mythos",   # not a real model ID as of June 2026
#     max_tokens=4096,
#     messages=[{"role": "user", "content": prompt}]
# )

The API contract will probably stay consistent

Anthropic’s model tiers share a unified API surface. When and if Mythos-class access returns, switching to it from Opus or Sonnet should be a single parameter change — the same way you can currently switch between claude-opus-4-8 and claude-sonnet-4-6. The cost of waiting is the cost of one config line, not the cost of rearchitecting your agentic workflow.

Cybersecurity tooling will be the first GA use case

The early-access program targets organisations focused on defensive security. If you work in AppSec, vulnerability research, or red-teaming, monitoring Anthropic’s early-access announcements is the highest-leverage move. The first sustained commercial deployment of Mythos-class capability will almost certainly be a cyber-defence pipeline under tight eligibility controls, not a general developer launch.

Plan for higher costs and restricted access

The leaked draft explicitly stated Mythos is “very expensive for us to serve, and will be very expensive for our customers to use.” The June launch-then-pull adds a second constraint: eligibility. Expect any Mythos-class return to start above current Opus 4.8 rates and to launch with access restrictions, higher minimum commitments, or security-and-eligibility gates. For most production workloads, effective prompt engineering on Sonnet or Opus will remain the more practical approach.

The 2026 Claude model lineup (June snapshot)

The diagram below shows where Mythos sits in Anthropic’s lineup today — and where the rejected Capybara tier would have sat in the leak’s framing. This is the most-asked question in support threads and Reddit comments.

How does Mythos fit the broader AI race?

The Mythos leak did not happen in isolation. The same week, reports emerged that OpenAI had finished pre-training its newest model, internally codenamed “Spud,” with CEO Sam Altman reportedly promising it could “really accelerate the economy.” Both companies are expected to time their strongest model releases around planned IPOs — Anthropic potentially as early as October 2026.

Reading the competitive dynamics in June, rather than in the heat of March:

• Anthropic’s commercial position has hardened. Throughout early 2026 the company has built significant momentum in B2B and developer markets, particularly through AI agents and coding tools. Claude Code has become the de facto standard for AI-assisted development at many organisations.
• Opus 4.8 now covers most ordinary Mythos-adjacent use cases. A tier above Opus still matters for cybersecurity and very long-horizon agents, but most coding, document, browser-agent, and knowledge-work teams should see the public bridge release as Opus 4.8, not as a reason to wait.
• The cybersecurity angle introduces a new competitive axis. Anthropic launched Claude Code Security in February 2026, and Mythos-class vulnerability-detection capabilities could turn the company into a major player in the security tooling market — currently dominated by CrowdStrike, Palo Alto Networks, and others. The 5%+ drop in cybersecurity stocks following the leak, and the US export directive that pulled Mythos in June, both suggest how seriously the frontier-cyber angle is taken.
• The infrastructure story matters. Anthropic’s reported deployments inside large enterprise environments — see our coverage of Anthropic’s higher-limit enterprise tier work with companies like SpaceX — suggest the company is preparing to serve a small set of very compute-hungry, eligibility-screened customers with Mythos-class models first.

Did Claude Mythos ship — and will it come back?

It shipped, then it stopped. Anthropic launched the Mythos-class tier on June 9, 2026 as Claude Fable 5 (general access) and Claude Mythos 5 (restricted), then disabled both on June 12 after a US-government export directive. So the question is no longer “when does it ship” but “when does access return.” Three factors set that window:

1. Export control is now the hard blocker. The June 12 directive suspends access by any foreign national, inside or outside the US. Until Anthropic can reliably filter eligibility in real time (or the directive is narrowed), a general relaunch is blocked — this is a compliance problem, not a capability one.
2. Cyber safeguards remain a parallel constraint. Even before the export order, Anthropic gated Mythos-class behind stronger cyber safeguards; Fable 5 shipped with hard limits that block high-risk cyber/bio/chem requests and fall back to Opus 4.8. Any relaunch keeps those limits.
3. EU AI Act GPAI obligations. Following the Digital Omnibus, high-risk deadlines moved to December 2027, but GPAI obligations (since August 2025) already apply — any durable European availability operates inside that regime. See our EU AI Act breakdown.

Realistic read: a return depends first on the export-control constraint being resolved, likely via eligibility-gated access (customer qualification, geo/identity filtering) rather than a wide-open consumer launch. Until then, Opus 4.8 is the most capable Claude you can actually build on.

The author’s own read — building on Claude in June 2026

I (Ignacy) build the articles on this site through a Claude Code workflow that has now moved from Opus 4.7 assumptions to Opus 4.8 as the right default target, with custom DTF skills loaded for the editorial gate, SEO validator, and source curation. I have not had access to the Mythos Preview. My honest read after the Opus 4.8 release and the brief Fable 5 / Mythos 5 launch:

• Opus 4.8 plus effort control and dynamic workflows covers more of the failure modes that production developers used to blame on “model not good enough.” The most important public gain is not raw intelligence; it is better judgment, self-critique, and long-task reliability.
• The cases where Mythos plausibly makes a real difference are narrow: deep red-team-grade vulnerability discovery, long-horizon agentic tasks where current models still drift after a few dozen turns, and high-stakes analytical work where the cost of being wrong is much higher than the cost of doubling token spend.
• The June launch-then-pull is the clearest lesson of all: even when the capability is ready, access can vanish overnight for regulatory reasons. Most teams should not optimise their roadmap around a Mythos-class release. The next 12 months will be defined by orchestration, dynamic workflows, observability, and agent architecture improvements as much as by the next model tier.