Key Summary
- METR evaluation: AI time horizon rises from 16 to 40 hours in six months
- Verifiable tasks like coding and optimization see rapid growth, while research judgment remains a human strength
- The starting point is dividing your organization's tasks into 'immediately verifiable' and 'requiring value judgment'
- AX transformation doesn't mean full automation—it starts with defining delegation scope by task type
According to METR's evaluation, the time horizon of cutting-edge AI rose from 16 hours in May 2026 to approximately 40 hours just six months later. AI safety evaluation organization METR released its R&D capability assessment report on November 20 (local time), revealing that OpenAI's GPT-6 Astra recorded a time horizon of about 40 hours, while Anthropic's Claude Opus 5.5 recorded about 39 hours.
Time horizon is a metric based on the time it takes a human expert to complete a specific task, measuring the level at which AI can successfully complete that same task with a 50% probability. According to METR, this evaluation expanded the scope by adding 36 more difficult tasks to the existing 222, broadening the assessment beyond coding and optimization to include the design of new research tasks and the ability to judge results. For this reason, METR explained that it is difficult to directly compare the previous 16-hour figure with the current 40-hour figure as an equivalent rate of performance improvement.
What Has Changed
METR estimated that the time horizon doubles on average every 3.61 months, or about 110 days. This is similar to the 105 days from the previous evaluation, indicating that the growth trend of capability doubling every 3 to 4 months continues. When compared against records from 15 programming optimization competitions, the top AI models scored higher than 96-98% of human participants, and in three separate tasks compared against specially recruited professionals, AI also achieved higher performance.
However, clear limitations remained in research judgment. METR had the AI design new evaluation tools for use in its own research, but none of the outputs proposed by the cutting-edge models were deemed useful enough by researchers to actually adopt. Still, METR noted that unlike models from a year ago, which fell well short of the evaluation standards, the latest models such as Claude Opus 5.5 came close to a usable level.
Implications for AX in Korean Companies
METR estimated that if the current growth trend continues, by July 2027 AI could achieve a 50% success rate even on tasks that take human experts about 200 hours to complete. It also noted that, as of earlier this month, Anthropic's AI model development speed was estimated to be about 1.9 times faster than it was the previous year.
What this metric shows is that AI is rapidly catching up to humans in tasks where results can be verified immediately, such as code verification and mathematical calculation checks. Conversely, tasks such as judging whether a particular research direction is worth pursuing or evaluating the significance of results remain firmly in the domain of human experts. For organizations considering AX transformation, the first step is to distinguish between tasks within their operations that are 'immediately verifiable' and those that 'require value judgment.' A practical approach suggested by these figures is to expand the scope of AI delegation starting with areas where the correct answer is clear—such as coding, optimization, and data verification—while maintaining human decision-making structures for the time being in areas that require task selection or strategic judgment.
Source: AI Task Capability Grows from 16 to 40 Hours in Six Months... "Judgment Still Has Limits"
