Anthropic has restricted pre-release testing of Claude Mythos 5.1, its frontier artificial intelligence model designed for high-risk applications, exclusively to vetted American organizations, marking the first time the UK AI Safety Institute has been excluded from early evaluation cycles. The decision, announced in September 2026, limits access to specialized US government programs and defensive consortiums while leaving international evaluators sidelined, raising questions about how frontier AI models will be assessed globally as capabilities advance.

Why Did Anthropic Exclude International Evaluators?

Anthropic has not publicly detailed the specific regulatory or policy drivers behind the decision to bypass British evaluators. However, the timing suggests a connection to US government export-control directives. The company previously paused access to frontier models across June 2026 following those directives before restoring deployment on July 1, 2026. The exclusion represents an abrupt shift from prior practice; the UK institute previously conducted extensive technical testing on Claude Mythos Preview in April 2026, recording a 73 percent solve rate on expert-level cybersecurity exercises and documenting the first complete solve of a 32-step cyber attack range.

AI safety researcher Miles Brundage called the exclusion a worrying precedent, though he acknowledged a nuanced reality about technical capacity.

"The British institute retains greater overall testing capacity than the US Center for AI Safety Institute, though the American center requires more funding, stable leadership and communication rules, and higher salaries," Brundage stated.

Miles Brundage, AI Safety Researcher

What Makes Claude Mythos 5.1 Different From Other Claude Models?

Anthropic launched Claude Mythos 5.1 alongside Claude Fable 5.1 on September 1, 2026, with both models sharing the same underlying weights but operating under fundamentally different safety regimes. The distinction matters significantly for how these models handle sensitive requests. Claude Fable 5.1, available to commercial customers, includes automated classifiers that intercept sensitive requests related to cybersecurity or biology. When those filters flag risky prompts, requests route down to earlier, less capable models like Claude Opus 4.8 and Claude Opus 5.5.

Claude Mythos 5.1, by contrast, operates without those filtering classifiers, exposing the raw model to trusted partners for evaluation and specialized use cases. This unfiltered access allows researchers and security professionals to assess the model's true capabilities and potential risks without safety guardrails intervening. The trade-off is tighter access controls; organizations approved for Mythos must accept a default 30-day data retention policy for safety monitoring.

How to Access Claude Mythos 5.1 and Related Programs

• Cyber Verification Program: Anthropic provides Mythos access through this specialized vetting channel run alongside the American government, designed for cybersecurity professionals and researchers.
• Life Sciences Verification Program: A parallel program for researchers working in biology and molecular design, also run with American government coordination.
• Project Glasswing: A defensive consortium established in April 2026 that includes approximately 150 organizations across more than fifteen countries, though current Mythos 5.1 participation remains limited to American institutions.

What Are Claude Mythos 5.1's Actual Capabilities?

Anthropic documented substantial capability jumps in the 5.1 release compared to earlier versions and competing models. On Terminal-Bench 4.0, a coding benchmark, Mythos 5.1 achieved a score of 60.9 percent with its safeguards turned off, compared to 55.8 percent for Fable 5.1 and 37.3 percent for GPT-5.6 Sol. In molecular biology evaluations, the model designed protein binders that reached a hit rate near 50 percent across 12 target structures, compared to typical field rates of 10 to 15 percent. These performance gains underscore why access controls matter; the model's improved capabilities in sensitive domains like synthetic biology create higher stakes for evaluation and oversight.

Pricing for Mythos 5.1 matches Fable 5.1 at $10 per million input tokens and $50 per million output tokens, making them cost-equivalent despite their different safety profiles. However, Anthropic cut cache-read pricing by 75 percent to $0.25 per million tokens, reducing costs for prolonged automated agent workflows where models process large amounts of cached context repeatedly.

What Do Experts Say About the Precedent?

The exclusion of the UK AI Safety Institute has triggered concern among researchers monitoring frontier model deployment practices. Industry observers continue to monitor whether British authorities will revise future memorandums of understanding to require domestic evaluations on upcoming frontier models. The decision highlights a tension between national security concerns and the collaborative international approach that has characterized AI safety evaluation to date. As frontier models grow more capable in dual-use domains like cybersecurity and synthetic biology, governments face pressure to restrict access while researchers argue that diverse, independent evaluation strengthens overall safety outcomes.

The practical implications extend beyond evaluation politics. Restricting Mythos access to American institutions may slow the pace at which international researchers can identify risks and contribute to safety improvements. Conversely, tighter controls may reduce the risk of sensitive capabilities spreading to actors without appropriate safeguards. The outcome of this policy choice will likely shape how other AI labs approach frontier model deployment in the coming years, particularly as export-control frameworks continue to evolve.