Anthropic's Claude Mythos Preview: What We Know About the Mysterious Frontier Model
Anthropic quietly shifted the AI conversation in early April 2026. On April 7, 2026, it introduced Project Glasswing and centered the launch around Claude Mythos Preview, a new model it describes as extremely capable in cybersecurity and coding. But unlike a normal model launch, this one is not open to everyone.
The model is real, benchmarked, and documented in a full system card. But access is tightly restricted, and Anthropic has said it does not plan to make Mythos Preview generally available right now. If you are trying to understand what is hype vs what is actually confirmed, this breakdown is for you.
The most important source is Anthropic's own Project Glasswing announcement. The company says Mythos Preview is a general-purpose, unreleased frontier model with unusually strong vulnerability discovery capability. Anthropic also says the model has already found thousands of high-severity vulnerabilities, including issues across major operating systems and web browsers.
This was not announced as a consumer feature update. It was announced as a security initiative with named launch partners, including AWS, Apple, Google, Microsoft, NVIDIA, Cisco, Broadcom, CrowdStrike, Palo Alto Networks, JPMorganChase, and the Linux Foundation.
Anthropic also says it extended access beyond those launch partners to over 40 additional organizations that maintain critical software infrastructure.
Project Glasswing is Anthropic's umbrella program for Mythos Preview's security-focused rollout.
Usually with major model launches, developers can try the model quickly and compare notes publicly. That is not the case here.
Anthropic's system cards index lists Mythos Preview as an April 2026 model, and the Mythos system card itself is extensive. But Anthropic says Mythos Preview is not for general availability at this stage. Access is tied to a controlled cybersecurity program.
So the "mystery" is not that the model is undocumented. It is the opposite: documentation exists, but broad hands-on access does not.
My read (inference from Anthropic's public docs): this looks like a controlled pre-deployment phase for a higher-risk capability tier, where Anthropic wants real-world security value without broad offensive misuse risk.
Anthropic's system card reports a large step up from Opus 4.6 on several practical benchmarks. These are the numbers that stand out most:
| Evaluation | Mythos Preview | Opus 4.6 | Reported change | | --- | ---: | ---: | ---: | | SWE-bench Verified | 93.9% | 80.8% | +13.1 pts | | SWE-bench Pro | 77.8% | 53.4% | +24.4 pts | | Terminal-Bench 2.0 | 82.0% | 65.4% | +16.6 pts | | HLE (with tools) | 64.7% | 53.1% | +11.6 pts | | OSWorld | 79.6% | 72.7% | +6.9 pts |
A few caveats from the same source matter:
Even with those caveats, the direction is clear: Anthropic is signaling a model that is materially stronger on high-value engineering tasks, not just marginally better on one benchmark.
Two numbers are getting the most attention: 93.9% on SWE-bench Verified and 79.6% on OSWorld.
Both are impressive, but they are useful for different reasons.
SWE-bench Verified is still one of the most cited coding benchmarks because it uses real GitHub issues that human reviewers have validated as solvable. A jump from 80.8% (Opus 4.6) to 93.9% (Mythos Preview) is not a rounding artifact. That is a large practical gain, especially if your workflow depends on patching real code under constraints.
OSWorld is different. It is closer to "can this model actually operate software interfaces end to end?" The system card says Anthropic ran OSWorld with default settings at 1080p and up to 100 action steps per task. A 79.6% first-attempt success rate in that setup supports the view that Mythos is stronger not only in static coding tasks, but also in multi-step interactive computer use.
The right takeaway is not that benchmarks are perfect. The right takeaway is that Mythos appears to improve across multiple evaluation families at the same time, which is usually a stronger signal than a single leaderboard spike.
The most interesting thing here is not one headline score. It is the mix of scores across coding, terminal tasks, tool use, and computer interaction.
When the same model jumps across all three categories, it usually means better end-to-end execution, not only better text generation.
That is also why security teams are paying attention. A model that can chain reasoning + code + tools + environment actions can be useful for defensive triage and patching, but those same traits can be risky in offensive contexts if released without strong controls.
Project Glasswing is basically Anthropic saying: "We think this capability tier has too much dual-use risk for normal public rollout right now."
From Anthropic's own announcement:
That structure matters. It gives Anthropic real deployment data under constraints, while delaying mass access until safeguards are more mature.
The Mythos system card contains a nuanced safety story.
Anthropic says Mythos Preview is its best-aligned model so far on many internal measures. At the same time, it also says that because capability is so high, rare misaligned behavior can still be serious. In other words: the average case may be better, but tail risk still matters.
The card also describes mitigation patterns such as:
Anthropic also states in the Glasswing announcement that it plans to ship new safeguards first with an upcoming Claude Opus model, then iterate from there.
That is an unusual sequence, and it reinforces how cautiously Anthropic is treating Mythos-class behavior.
As of April 13, 2026, Anthropic's public position is clear: Mythos Preview is not for general availability right now.
At the same time, Anthropic also says its long-term goal is to safely deploy Mythos-class capability at scale. That strongly suggests Mythos is not a dead-end experiment. It is more likely a staging model used to stress-test safeguards, policy controls, and deployment procedures before wider release of similarly capable systems.
In practical terms, this likely means most developers should watch the next Opus-class release carefully. If Anthropic introduces new cyber and misuse protections there first, that release will probably be the bridge between today's restricted Mythos rollout and broader high-capability access.
If you build products with AI, Mythos Preview still matters to you even if you cannot use it today.
First, it is a directional signal. Frontier performance is moving toward stronger autonomous engineering capability, especially in code-heavy and environment-heavy tasks.
Second, it is a deployment signal. High-capability models may increasingly launch in restricted, policy-heavy tracks before broader API access. Teams that expect immediate public availability for every frontier release may need to adjust planning.
Third, it is a security signal. The practical bar for secure agent design just went up. If frontier models become stronger at tool use and exploit reasoning, organizations need better guardrails, better runtime monitoring, and better incident response for model-driven workflows.
A pragmatic takeaway:
If Anthropic follows its own stated timeline, these are the questions that will matter most by mid-2026:
These answers will tell us whether Glasswing is just a one-off security initiative or the start of a broader model-release pattern.
The "mysterious model" narrative is understandable, but the stronger framing is this: Anthropic is testing a powerful model under constrained deployment because it believes open release would currently carry too much downside.
The data Anthropic published suggests Mythos Preview is a meaningful jump over Opus 4.6 on coding and agentic workloads. The policy posture suggests Anthropic believes capability is advancing faster than safe default deployment norms.
If that assessment is correct, Project Glasswing may end up being remembered less as a one-time launch and more as an early template for how frontier models with serious dual-use potential get introduced.
For builders, the immediate move is straightforward: keep shipping, but raise your security posture now. The next generation of model behavior is already here.