On April 7, Anthropic officially announced something unusual: it released a new model called Claude Mythos, while simultaneously stating the model would not be open to the public.
Not "available later," but "currently only for 40 organizations; everyone else waits."
The project is called Project Glasswing.
How strong is Mythos?
First, the data:
| Benchmark | Claude Opus 4.6 | Claude Mythos | Change |
|---|---|---|---|
| SWE-bench Verified | 80.8% | 93.9% | +13.1% |
| SWE-bench Pro | Not disclosed | 77.8% | Current top |
The current second place on SWE-bench Pro is Zhipu GLM-5.1 (58.4%), and third is GPT-5.4 (57.7%). Mythos leads the second-place model by nearly 20 percentage points at 77.8% — a significant gap.
But Anthropic isn't just topping leaderboards. They had Mythos scan real codebases and found:
- Thousands of zero-day vulnerabilities in mainstream operating systems and browsers
- A bug in FFmpeg that had been hiding for 16 years — a vulnerability that had been missed by over 5 million passes of traditional security scanning tools
Cisco Chief Security Officer Anthony Grieco said: "This represents a profound paradigm shift and a clear signal that the old approach to hardening systems is no longer sufficient."
What is the logic behind Project Glasswing?
What Anthropic is giving these 40 organizations is preview access to Mythos, for one purpose only: defensive cybersecurity work.
Current partners with access include: Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks.
Anthropic is also providing:
- $100 million in API credits covering security testing and research for participants
- $4 million in direct donations to open-source security organizations
Why this "limited edition" approach?
The official explanation is straightforward: Mythos "currently far surpasses any other AI model" in cybersecurity capabilities, but precisely because of that, it could also be used to accelerate large-scale cyberattacks. Anthropic believes the risk of releasing this model publicly without sufficient safeguards is too high.
So they chose a compromise: give it to defenders first, let security teams patch vulnerabilities in advance, and then consider broader release.
Is this strategy sound?
It's worth considering.
The traditional AI model release logic is: release, test, find problems, patch, repeat. This process is public, and risk is diffused — attackers and defenders get new tools simultaneously.
Glasswing's approach reverses that: give defenders a time window first, let them use Mythos to proactively find and fix vulnerabilities, and then consider wider access.
From a game theory perspective, this design makes sense — if defenders move first, the cost for attackers to exploit the same model becomes significantly higher.
But actual effectiveness depends on two things: whether these 40 organizations can actually patch critical vulnerabilities within this window, and whether the model's capabilities themselves are at risk of leaking (Mythos's existence was already disclosed by media following a data leak).
What level of model is Mythos?
Anthropic has explicitly stated this is "the most powerful model to date," not an iterative extension of the Opus 4.x series, but a new generation.
Weeks ago, media outlets exposed Mythos's existence following a data leak. Fortune reported that Anthropic internally described it as "a step-change in capability." Now both the name and capabilities are officially confirmed, but not through a public release — rather through a targeted partnership program.
What Anthropic is waiting for hasn't been made clear. But based on the SWE-bench Pro 77.8% figure, this model's coding ability is already on another level.
Sources: Anthropic debuts preview of powerful new AI model Mythos in new cybersecurity initiative (TechCrunch); Anthropic debuts Project Glasswing, CocoLoop, leveraging its powerful Mythos model to reinforce software security (SiliconAngle); Anthropic is giving some firms early access to Claude Mythos to bolster cybersecurity defenses (Fortune); Exclusive: Anthropic Mythos AI model representing step change in power revealed in data leak (Fortune)