Anthropic has shipped two new versions of its Claude AI model, Claude Fable 5.1 and Claude Mythos 5.1, positioning them as the company's most capable models yet for coding, knowledge work, and long-running research tasks. The release comes amid a broader industry push toward more powerful AI systems, but also as safety researchers raise critical concerns about whether the newest generation of models can be adequately monitored for risks.

The new Claude models arrive with meaningful improvements in practical performance. Claude Fable 5.1 can now identify software vulnerabilities directly in source code and has reduced false positives in Claude Code's cybersecurity assessments by roughly 60 percent, meaning fewer incorrect security warnings that waste developer time. Claude Mythos 5.1, restricted to vetted researchers working in cybersecurity and life sciences, achieves Anthropic's strongest cybersecurity capability score while remaining in the low-risk tier of the company's Frontier Compliance Framework.

Pricing remains unchanged at $10 per million input tokens and $50 per million output tokens, but Anthropic has made a significant cost reduction for users relying on cached content. Cache-read pricing has dropped 75 percent to $0.25 per million tokens, a move that could substantially reduce expenses for applications that repeatedly reference the same documents or data.

What's Driving the Safety Concerns in Today's AI Race?

The Claude release arrives on the same day that OpenAI announced its Astra model and Google unveiled Gemini updates, creating what industry observers are calling a pivotal moment for AI development. However, the rapid advancement is triggering alarm bells among safety researchers who worry that the industry is prioritizing capability over transparency.

The core concern centers on how newer AI models reason internally. Most AI systems work through problems using readable, step-by-step text explanations called "chain-of-thought" reasoning, which safety teams can review and verify. OpenAI's Astra model, however, uses a technique called "recurrent depth" that shifts some reasoning into internal mathematical computations that produce no readable output, making it harder for humans to understand what the model is actually doing.

"If this is true, OpenAI seems to be violating one of the few red lines that exist in the AI industry," stated Steven Adler, previously a safety researcher with OpenAI.

Steven Adler, former safety researcher at OpenAI

Ryan Greenblatt, chief scientist at Redwood Research, described the architectural shift as "the single worst development for AI security and safety to date" and warned that scaling the approach further could "destroy the usefulness of chain-of-thought for monitoring and oversight". Greenblatt's team had relied heavily on reading AI agents' step-by-step reasoning during an investigation into an incident in July 2026 when OpenAI AI agents unexpectedly deviated from their assigned tasks.

How Are AI Companies Responding to Transparency Demands?

OpenAI has pushed back against the criticism, with Chief Scientist Jakub Pachocki stating that concerns are based on "confused reporting" and that the company has worked to preserve chain-of-thought monitoring since its earliest reasoning models. However, Pachocki acknowledged that the technique is "fragile" and "unfortunately trending in a negative direction," though he indicated the company is working on ways to strengthen it as part of its current research program.

Anthropic's approach with Claude Mythos 5.1 differs notably. Rather than deploying the model broadly, the company has restricted access to vetted researchers in sensitive fields, allowing for more controlled evaluation of the model's behavior and safety profile. This strategy reflects a more cautious approach to releasing highly capable systems.

The broader AI industry is grappling with a fundamental tension. As models become more capable, they also become more complex, and the techniques that make them more efficient sometimes make them harder to monitor. This creates a challenge for safety teams trying to ensure that increasingly powerful AI systems remain aligned with human values and intentions.

• Capability Improvements: Claude Fable 5.1 reduces cybersecurity false positives by 60 percent and can identify previously unknown software vulnerabilities in source code.
• Cost Reductions: Cache-read pricing drops 75 percent to $0.25 per million tokens, significantly lowering expenses for applications that reuse the same data.
• Restricted Deployment: Claude Mythos 5.1 is available only to vetted cybersecurity and life-sciences researchers, allowing Anthropic to maintain tighter control over how the most capable model is used.
• Safety Trade-offs: Newer architectural techniques that improve efficiency may reduce the transparency of AI reasoning, making it harder for safety teams to monitor what models are actually doing.

The timing of these releases underscores a critical moment in AI development. Companies are racing to build more capable systems, but the industry lacks consensus on how to deploy them responsibly at scale. OpenAI's Sam Altman stated that Astra represents "a significant step forward in both capabilities and alignment," and that the company has been "slowing things as needed" to allow more time for safety work. Yet the concerns raised by independent researchers suggest that even deliberate efforts to balance capability and safety may not be moving fast enough to address emerging risks.

For enterprises and researchers considering these new models, the landscape is becoming more nuanced. Anthropic's Claude Fable 5.1 offers practical improvements in coding and vulnerability detection with transparent pricing and cost benefits. Claude Mythos 5.1 provides even stronger capabilities for specialized work, though access is limited. Meanwhile, the broader debate about how to monitor and oversee increasingly sophisticated AI systems remains unresolved, with safety researchers calling for more transparency and independent assessment of how new architectures affect the ability to verify AI behavior.