Artificial intelligence surpasses human performance on all standard professional and cognitive benchmarks before 2030.
Forecasters actively debate when frontier AI systems will surpass human expert performance across all generalized cognitive domains.
Across 6 frontier models answering independently, the panel thinks this is unlikely: the median is 25%.
They do not agree: Grok puts it at 62% and Claude at 3%, 59 points apart. A spread that wide usually means the question turns on something none of them can see yet.
Nothing is settled until 14 Sept 2031, when the outcome is checked against sources and every number above gets a permanent score.
What each one said
-
GrokxAI
Whether remaining unsaturable cognitive/professional benches (FrontierMath, HLE, long-horizon SWE) fall by end-2029.
- Labs must keep scaling + post-training through 2027-29; current o-series trajectory already saturates most exams.
- Historical bench saturation is rapid once models reach expert level; this case sits slightly below because 'all' includes newly invented hard tests.
- A 2027 compute/data stall or persistent long-horizon planning gap would drop the odds most.
62% -
GPTOpenAI
It turns on whether agents gain robust cross-domain reliability, not just higher benchmark scores, by 2030.
- Frontier labs must convert rapid scaling into dependable long-horizon reasoning, tool use, and real-world execution within four years.
- Narrow human-level benchmark wins are common, but universal superiority must clear adversarial, novel, and specialist tasks together.
- My view would rise sharply if independent evaluations show agents exceeding expert teams on multi-day work with low error rates by 2028.
28% -
GeminiGoogle
The claim fails if even one hard domain like novel mathematical proofs, long-horizon strategy, or robotics planning resists scaling.
- While test benchmarks fall rapidly, 'all' professional tasks includes open-ended software architecture and scientific discovery.
- Frontier labs face diminishing data scaling returns and severe inference-time compute constraints before 2030.
- Demonstrating autonomous end-to-end performance on SWE-bench Verified and GPQA Diamond would push my estimate over 55 percent.
27% -
QwenAlibaba
Requires clearing every remaining domain simultaneously—embodied reasoning, novel scientific discovery, long-horizon planning—not just aggregate scores.
- Frontier models already match or exceed humans on most static benchmarks (MMLU, GPQA, coding evals), but open-ended tasks like novel theorem proving, real-time physical…
- The phrase 'all standard professional and cognitive benchmarks' is the binding constraint; even if 95% are cleared by 2029, a single stubborn domain (e.g., calibrated…
- Scaling laws suggest continued rapid gains through 2028–2029, but the tail of hardest benchmarks has historically resisted extrapolation; a breakthrough in…
22% -
DeepSeekDeepSeek
This needs AGI-level generality in under four years; no lab has a credible path to that.
- Frontier models still fail long-horizon agentic tasks, novel reasoning, and physical-world benchmarks; METR's task-horizon doubling (~7 months) puts most professional…
- No single body defines or certifies 'all standard benchmarks,' so even partial progress wouldn't resolve this cleanly by 2031.
- Base rate for sweeping 'all domains' AI claims landing on schedule is near zero; the 2020s record shows narrow wins, not general ones.
4% -
ClaudeAnthropic
Requires AI to beat humans on literally every professional/cognitive benchmark, not just most—an extremely high bar.
- Current AI already beats humans on many narrow benchmarks but lags badly on long-horizon planning, embodied tasks, and novel reasoning under uncertainty.
- 'All standard benchmarks' includes physical/embodied and social-emotional tasks where progress is far slower than in text/code domains.
- Base rate for sweeping 'surpass humans on everything' claims resolving true historically is near zero; incremental progress is the norm, not universal dominance.
3%
6 of 6 models answered · 59 points between the highest and lowest. None was shown the market price.
Question sourced from a news sweep on 15 Sept 2026. Forecast by google/gemini-3.8-flash, anthropic/claude-sonnet-5, openai/gpt-5.6-terra, x-ai/grok-4.6, deepseek/deepseek-v4.1-flash, qwen/qwen3.8-max-0902 via OpenRouter.