Anthropic's Claude Opus 4.8 Advances AI Honesty, 'Mythos' Model Nears

Anthropic has unveiled Claude Opus 4.8, the latest iteration of its AI model, promising significant improvements in performance and, notably, a greater degree of honesty. The company also teased an upcoming, even more advanced AI model codenamed 'Mythos'.

Opus 4.8: Enhanced Performance and Honesty

Claude Opus 4.8 has reportedly surpassed its predecessor, Opus 4.7, in various benchmarks. In the "Agentic Code" test on SWE Bench Pro, Opus 4.8 achieved a score of 69.2 percent, up from 64.3 percent. Anthropic claims the new model also outperforms competitors in areas like computer use, financial analysis, and knowledge work. While GPT-5.5 reportedly edged out Opus 4.8 in "Agentic Terminal Coding," early testers suggest Opus 4.8 demonstrates superior decision-making, asks more pertinent follow-up questions, and is better at recognizing its own errors.

Anthropic emphasized that Opus 4.8 has been trained to be more "honest." This means the model is less prone to making unsupported claims or jumping to conclusions when evidence is thin. Instead, it's designed to highlight uncertainties about its own work, providing a more transparent view of its capabilities and limitations.

Availability and Developer Options

Claude Opus 4.8 is now available across all Anthropic products for subscribers. Pricing remains unchanged from the previous version. Developers can access Opus 4.8 through settings like Claude Code and can opt to use more tokens for complex coding tasks, a feature Anthropic recommends for achieving better results. To accommodate this increased token usage, Anthropic has raised the usage limits for Claude Code.

The Road to 'Mythos'

Looking ahead, Anthropic is working on a more cost-efficient model with performance comparable to Opus. More excitingly, they are developing a successor to Opus 4.8, internally codenamed 'Mythos,' which aims to be even more intelligent.

A select group of companies are already testing 'Mythos' under the codename 'Project Glasswing'. Anthropic is focusing on establishing robust safety protocols before its public release. According to the company, "Mythos-level models" are expected to be rolled out to customers within the coming weeks, suggesting rapid progress on this next-generation AI.

Context:

The AI landscape is fiercely competitive, with major players like OpenAI, Google, and Anthropic constantly pushing the boundaries of model capabilities. The focus on "honesty" and self-awareness in AI models is a critical area of research, directly addressing concerns about AI hallucination and reliability. For European users and developers, the increasing power of these models, coupled with Anthropic's emphasis on safety, signals a maturing AI ecosystem that may eventually align with the EU's stringent regulatory frameworks, such as the AI Act.

What this means for you:

If you're a subscriber to Anthropic's services, you can immediately start using Claude Opus 4.8 for potentially more accurate and reliable AI assistance. For developers, the increased token limits and improved error recognition could streamline complex coding projects. Keep an eye out for the public release of 'Mythos' in the coming weeks, as it promises a significant leap in AI intelligence, which could impact everything from creative work to complex problem-solving.

What's still unclear:

• The exact pricing structure for 'Mythos' and how it will compare to Opus 4.8.
• The specific safety measures being implemented for 'Mythos' and how they will be communicated to the public.
• The full range of capabilities and benchmarks for 'Mythos' beyond its claimed superior intelligence.

Why this matters:

Anthropic's Claude Opus 4.8 advances AI honesty, with 'Mythos' model imminent. This dual focus on improved reliability and groundbreaking intelligence positions Anthropic as a key player in the AI race, driving the industry towards more transparent and capable artificial intelligence systems.