So, they make a model so good that they are not releasing it to the public? Claude mythos☠️
Anthropic's internal 'Glasswing' model achieves near-perfect coding scores, prompting internal warnings of a historic turning point.
Anthropic has reportedly developed a highly capable internal AI model, codenamed 'Glasswing,' which is not being released to the public. The model's significance was highlighted by an internal researcher who stated, "Glasswing is possibly the most consequential event in the AI industry I've seen up close since joining Anthropic almost 3 years ago. It feels like we're at a turning point in history." This language has sparked speculation about the model's power, with some interpreting it as a sign of nearing advanced, potentially risky AI capabilities.
The model's technical prowess is underscored by its reported 93.9% score on SWE-bench, a benchmark for evaluating AI on real-world software engineering tasks. This near-perfect performance suggests a massive leap in autonomous coding ability. Instead of releasing it, Anthropic is reportedly using Glasswing as a security tool, running it against critical infrastructure and other new frontier models to proactively identify vulnerabilities before they are deployed. This represents a strategic shift towards internal 'red teaming' with ultra-capable models to manage the risks posed by rapid AI advancement.
- Anthropic's internal 'Glasswing' model scores 93.9% on the SWE-bench coding benchmark, indicating superior performance.
- An internal researcher called it the "most consequential event" in AI they've witnessed, hinting at a historic capability leap.
- The model is being gatekept and used to test critical infrastructure and other AI models for security vulnerabilities pre-release.
Why It Matters
It signals a new era where AI capabilities may advance so rapidly that responsible deployment requires internal gatekeeping and security testing first.