After Claude Opus 4.8, What Is Anthropic’s Mythos-Class Model? Everything We Know

Every major Anthropic launch in 2026 has ended the same way: a footnote about Mythos. On May 28, when Anthropic released Claude Opus 4.8, the footnote got bigger. For the first time, the company said explicitly that Mythos-class models are coming to all customers “in the coming weeks.” That makes today the right time to understand exactly what Mythos is, why it has been locked away for months, and what to expect when it finally arrives.

How Mythos Came to Light: The Accidental Leak

Claude Mythos Preview did not have a quiet, planned reveal. On March 26, 2026, a configuration error in Anthropic’s content management system left draft blog posts and internal documents publicly accessible. Those documents described a model called Mythos — positioned as something far above the existing Opus tier.

Anthropic moved quickly. It patched the misconfiguration, confirmed the model was real, and scheduled a formal announcement. That announcement came on April 7, 2026 — accompanied by a 244-page system card, a red-team disclosure, and the launch of Project Glasswing. It was, by any measure, an unusual way to introduce a flagship model.

📌 Notable: Anthropic published a full system card for Mythos Preview before making the model publicly available — a first for any major AI lab. The 244-page document addresses safety evaluations extensively, though it does not disclose pretraining data composition.

What Is Claude Mythos? A New Model Tier Above Opus

Mythos is not a larger version of Opus. Anthropic introduced a new tier — internally codenamed Capybara — specifically because the model’s capabilities are qualitatively different from Opus-class models, not just quantitatively better. Industry estimates put Mythos at approximately 10 trillion parameters using a Mixture of Experts (MoE) architecture, where only a fraction of parameters are active during any single inference.

Anthropic’s own description from the red team disclosure is precise: Mythos Preview is “strikingly capable at computer security tasks.” But the cybersecurity angle, while dramatic, is a downstream consequence of a more fundamental truth: Mythos is a better general-purpose AI model across the board.

The Benchmark Numbers: A Genuine Capability Leap

Below are verified scores from Anthropic’s system card and third-party analyses. All figures are as published; asterisked estimates are from corroborated third-party benchmarks.

Benchmark	Mythos Preview	Opus 4.8	vs Opus 4.6 baseline
SWE-Bench Verified	93.9%	88.6%	80.8% (Opus 4.6)
SWE-Bench Pro	77.8%	69.2%	53.4% (Opus 4.6)
Terminal-Bench 2.0	82.0%	~74–75%*	65.4% (Opus 4.6)
CyberGym (Cybersecurity)	83.1%	N/A (public)	66.6% (Opus 4.6)
Cybench (CTF)	100% (saturated)	N/A (public)	<1% (prev. best)
USAMO 2026 (Maths)	97.6%	N/A	Major leap vs 4.6
GPQA Diamond	94.6%	~90%*	Near-saturated

*Terminal-Bench 2.0 score for Opus 4.8 is estimated based on its Terminal-Bench 2.1 score of 74.6% and the harness difference. | GPQA Diamond score for Opus 4.8 is approximate from third-party sources.

The gap on SWE-Bench Pro is particularly telling: Mythos at 77.8% versus Opus 4.8 at 69.2% is an 8.6-percentage-point lead on the hardest real-world coding benchmark available. On cybersecurity-specific tests, the distance is even larger — and more consequential.

Why Anthropic Is Holding It Back: The Cybersecurity Problem

The numbers above explain why Anthropic made an unprecedented decision: publish the system card, announce the model, but do not release it.

On Cybench — a benchmark of 35 capture-the-flag challenges from major security competitions — the previous best AI model scored below 1%. Mythos scored 73% (Cybench pass@1), eventually achieving 100% on the full benchmark, saturating it entirely. On CyberGym’s vulnerability reproduction benchmark, developed at UC Berkeley, Mythos scored 83.1% compared to 66.6% for Opus 4.6.

These are not theoretical risks. During Project Glasswing evaluations, Mythos identified thousands of high-severity vulnerabilities across every major operating system and web browser — including flaws that had survived decades of expert human review. It also chained multiple Linux kernel flaws to demonstrate a full privilege-escalation path from ordinary user access to full machine control.

⚠️ Anthropic’s own statement: “Although Mythos is currently far ahead of any other AI model in cyber capabilities, it presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders. In the short term, this could be attackers, if frontier labs aren’t careful about how they release these models.”

The dual-use nature of these capabilities — the same model that finds a zero-day for a defender can find it for an attacker — led Anthropic to conclude that no company, including itself, had developed safeguards strong enough for a public release. That is why Mythos went to Project Glasswing first.

Project Glasswing: The $100M Defensive Moat

Project Glasswing is Anthropic’s structured response to the problem it created. The $100 million initiative gives exclusive Mythos access to a consortium of organisations for defensive cybersecurity work on internal systems and open-source code.

Founding partners include Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks — plus approximately 40 additional vetted critical-infrastructure operators.

The logic is asymmetric deployment: give defenders a head start in a faster cyber fight. Partners use Mythos to scan their own codebases and open-source dependencies for vulnerabilities before attackers can weaponise the same capabilities in future, less-restricted models.

Results so far (as of May 2026): Glasswing partners have reported 530 high or critical vulnerabilities, of which 75 have been patched and 65 have public advisories. The low fix rate reflects the 90-day coordinated vulnerability disclosure window still being in progress. Claude Security — a lighter, Opus 4.7-powered version available to Enterprise customers — has separately patched over 2,100 vulnerabilities in its first three weeks of public beta.

India relevance: India’s large IT services sector — with major presence in US and European critical infrastructure codebases — stands to be directly affected by the vulnerability landscape Mythos is uncovering. Indian enterprises on Claude Enterprise can already access Claude Security in public beta today.

The Full Timeline: From Leak to “Coming Weeks”

Date	Event
Mar 26, 2026	Mythos Preview accidentally leaked via misconfigured CMS; draft blog posts exposed publicly
Apr 7, 2026	Anthropic officially announces Mythos Preview and Project Glasswing; 244-page system card published
Apr 7, 2026	red.anthropic.com publishes findings: “strikingly capable at computer security tasks”
Apr 8, 2026	Project Glasswing partners confirmed: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks
Apr 30, 2026	Claude Security (powered by Opus 4.7) launched in public beta for Enterprise customers — a Glasswing preview for the public
May 26, 2026	gHacks reports Anthropic plans public release of Mythos-class models once safeguards are ready
May 28, 2026	Opus 4.8 launch post: “expect to be able to bring Mythos-class models to all our customers in the coming weeks”
Coming weeks	Mythos-class public release — timeline unconfirmed, estimated mid-to-late June 2026 based on Anthropic cadence

When Will Mythos Go Public? What We Know

As of May 29, 2026, Anthropic has not committed to a specific date. The phrase “coming weeks” — used in the Opus 4.8 blog post — is the company’s clearest public signal yet. Here is what the available evidence suggests about the rollout sequence:

• Enterprise API access first — consistent with every prior Claude launch
• ai (Pro, Team, Max) — likely shortly after API
• Amazon Bedrock and Google Vertex AI — typically within days of the main API launch
• Free tier access — uncertain; Mythos pricing will likely be premium only at launch

On pricing: Project Glasswing partner access costs $25 per million input tokens and $125 per million output tokens — five times the standard Opus 4.8 rate. A public Mythos tier will almost certainly be priced as a premium above Opus, though Anthropic may introduce a standard and fast mode (as it did with Opus 4.8) to broaden accessibility.

💡 India cost note: At Glasswing partner pricing (₹2,375 per million input tokens), Mythos would be significantly more expensive than Opus 4.8 (₹475 per million input tokens). For Indian startups and SMBs, the practical entry point will likely be Claude Security (public beta now) or the eventual Mythos fast mode. Prompt caching and batch pricing will be critical to watch.

What Mythos Means for AI Development Beyond Cybersecurity

The cybersecurity angle has dominated coverage, but Mythos matters more broadly. Three things stand out:

1. Mathematical Reasoning at Near-Human Expert Level

A score of 97.6% on USAMO 2026 — the United States Mathematics Olympiad — is not a coding benchmark. It requires creative problem-solving at the level of the best university mathematics students in the world. This capability is directly relevant to quantitative research, financial modelling, drug discovery, and any domain where reasoning under uncertainty matters.

2. A New Baseline for Agentic Work

Mythos’s 93.9% on SWE-Bench Verified and 77.8% on SWE-Bench Pro suggest a model that can handle almost any real-world software engineering task autonomously. When Mythos capabilities trickle into future Opus and Sonnet models — as they historically do — the practical productivity ceiling for AI-assisted development will move significantly.

3. Alignment That Scales With Capability

Unusually, Anthropic’s alignment assessments show Mythos scores close to Opus 4.8 on prosocial traits — not worse. The standard assumption that more capable models are harder to align appears not to hold here, which is significant evidence for Anthropic’s approach to Constitutional AI and RLHF at scale.

Frequently Asked Questions

What is Claude Mythos?

Claude Mythos Preview is Anthropic’s most powerful AI model, announced April 7, 2026. It sits above the Opus tier in a new category called Capybara, with scores of 93.9% on SWE-Bench Verified, 97.6% on USAMO 2026 mathematics, and 83.1% on CyberGym cybersecurity benchmark.

When will Claude Mythos be released publicly?

Anthropic confirmed on May 28, 2026 that Mythos-class models are coming to all customers “in the coming weeks.” Based on their release cadence, mid-to-late June 2026 is the most likely window. Enterprise API access will come before claude.ai availability.

Why is Claude Mythos restricted?

Mythos can autonomously identify and exploit zero-day vulnerabilities in major operating systems and browsers. Anthropic concluded that no company — including itself — has developed safeguards strong enough to prevent misuse at this capability level. Project Glasswing gives defenders access first.

What is Project Glasswing?

A $100 million AI cybersecurity initiative by Anthropic, giving exclusive Mythos access to 12 founding technology companies and approximately 40 critical-infrastructure operators for defensive vulnerability research. Partners include AWS, Apple, Google, Microsoft, and NVIDIA.

How does Claude Mythos compare to Claude Opus 4.8?

Mythos leads Opus 4.8 on every published benchmark: SWE-Bench Verified (93.9% vs 88.6%), SWE-Bench Pro (77.8% vs 69.2%), and Terminal-Bench 2.0 (82.0% vs ~75%). Mythos also has exclusive cybersecurity benchmark scores with no Opus 4.8 equivalent.

What will Claude Mythos cost?

Project Glasswing partner pricing is $25/M input and $125/M output — five times the standard Opus rate. Public pricing is unconfirmed. Expect a premium tier above Opus 4.8’s $5/$25 pricing.

Bottom Line

Claude Mythos Preview is the most capable publicly benchmarked AI model in existence — but it is also the first frontier model held back specifically because it is too powerful to release responsibly without new safeguards. That is a historically significant moment in AI development, regardless of where you stand on the risks.

For WorthView readers: the practical question is not whether Mythos is impressive — the benchmarks settle that — but whether the public version, when it arrives, will retain the cybersecurity capabilities that make the restricted version so significant. Anthropic’s answer to that will define what Mythos actually means for developers, enterprises, and India’s fast-growing AI adoption curve.

Watch this space. “Coming weeks” is now a countdown.