Published: 25 February 2026
What happened: Research institute METR (Model Evaluation and Threat Research) has published new benchmark data showing AI software capabilities doubling roughly every seven months. Its latest evaluation of Anthropic’s Claude Opus 4.6 broke all previous records — and the rate appears to be accelerating.
Why it matters: The METR chart has become a key reference for AI researchers and investors alike, with markets moving on its results. Yet METR’s own researchers now admit significant uncertainty in their measurements: AI is becoming so capable that finding tasks difficult enough to test it properly is increasingly hard.
Wider context: METR researcher Joel Becker told Sky News the situation is “serious, fast-moving, and appears not to be slowing down.” He also noted that AI tools are already “meaningfully speeding up” the rate at which AI professionals build better AI systems — a feedback loop that carries real weight even without full self-improvement. Demis Hassabis regularly states AI will have ten times the impact of the Industrial Revolution in a tenth of the timespan.
Background: METR’s benchmark measures how long a task an AI can complete 50% of the time. Even an 80% success rate falls short of what reliable corporate automation would require. Current UK and US employment data shows little sign of AI disruption — adverts for software engineering roles on Indeed are rising — but Becker cautions that the most dramatic capability gains have only emerged in the past few months, while economic statistics lag months behind.
AI is developing so fast it is becoming hard to measure, experts say — Sky News
Singularity Soup Take: When the people building the benchmarks admit they can’t keep up with what they’re measuring, it’s less a milestone report and more a quiet confirmation that we’ve entered genuinely uncharted territory.
Key Takeaways:
- Doubling Rate: METR data shows AI software capabilities doubling approximately every seven months — and the Claude Opus 4.6 evaluation suggests the rate is not stabilising but accelerating.
- Measurement Ceiling: METR researchers acknowledge growing uncertainty in their own evaluations; AI capability is outpacing the difficulty of available test tasks, making precise benchmarking increasingly unreliable.
- Automation Gap: METR’s headline metric measures 50% task completion rates; even an 80% benchmark wouldn’t support dependable automation in a corporate environment, leaving a significant gap between capability and deployment.
- Employment Lag: UK and US job statistics currently show no significant AI impact — software engineering adverts on Indeed are rising — but METR warns that the most dramatic gains have only occurred in the past few months, with economic data trailing behind.