Frontier systems now meet or exceed established human baselines on long-running benchmarks like ImageNet, SuperGLUE, and MMLU, and have reached or approached human levels on harder reasoning tests — PhD-level science (GPQA Diamond), multimodal reasoning (MMMU), and competition math (AIME). On SWE-bench Verified, performance rose from roughly 60% in 2024 to close to 100% of the human baseline in 2025.

The frontier is converging — even across countries

• The top four Arena models are now separated by fewer than 25 Elo points, down from roughly 97 a year earlier. Anthropic (1,503), xAI (1,495), Google (1,494), and OpenAI (1,481) cluster at the top, with Alibaba (1,449) and DeepSeek (1,424) trailing only modestly.
• The U.S.–China gap has effectively closed. In February 2025 DeepSeek-R1 trailed the top U.S. model by just 5 Arena points (0.4%); as of March 2026 the top U.S. model leads the top Chinese model by 2.7%, fluctuating in the single digits all year.
• The open-weight gap reopened after briefly closing. The top closed model now leads the top open model by about 3.3%–3.4%, up from 0.5% in August 2024 — though open models are far more competitive than a few years ago.

Benchmarks are saturating — and facing reliability concerns

Evaluations meant to be hard for years are now saturated in months. A Stanford review of nine widely used benchmarks found invalid-question rates ranging from 2% on MMLU Math to 42% on GSM8K, and separate research suggests Arena leaderboard standing may partly reflect adaptation to the platform rather than general capability. With capability no longer a clear differentiator, competition is shifting toward cost, reliability, and real-world usefulness.

Frontier models gained 30 percentage points in a single year on Humanity's Last Exam — a benchmark built to be hard for AI and favorable to human experts — climbing from under 10% to 38.3%. On GPQA Diamond, mean accuracy reached 93%, 12 points above the expert validator baseline of 81.2%. On MMMU, Gemini 3.1 Pro Preview scored 88.2%, within 0.4 points of the best human expert.

Jagged intelligence: IMO gold vs. analog clocks

• Gemini Deep Think scored 35 points to win gold at the 2025 International Mathematical Olympiad, working end to end in natural language within the 4.5-hour limit — up from the 28-point silver achieved in 2024.
• Yet on ClockBench, the top model read analog clocks correctly only 50.6% of the time, versus 90.1% for humans. When models read the time wrong, their median error was about one to three hours, compared with three minutes for people.
• On FrontierMath Tier 4, accuracy rose from near 0% to 31.3% since 2024 — but the best models still fail roughly two of every three problems at the hardest tier. On MathArena, answer-based accuracy hit 97% while rigorous proof-writing lagged far behind.

AI agents advanced from answering questions to completing multistep tasks in 2025, though they still fail roughly one in three attempts on structured benchmarks. On OSWorld, accuracy rose from about 12% to 66.3%, within 6 points of human performance; on WebArena it reached 74.3%, within 4 points of the 78.2% human baseline; on Cybench the unguided solve rate hit 93%, up from 15% in 2024.

Lab triumph, household failure

• In controlled simulation, robotic manipulation on RLBench reached an 89.4% success rate (EquAct), up from about 48% in 2022. But in the 2025 BEHAVIOR-1K household challenge, the top team completed just 12.4% of real tasks — a stark measure of how far AI is from the unstructured physical world.
• On ResponsibleRobotBench, even the best model (GPT-4o, safe score 0.64) completed under a third of tasks safely when real hazards were present.
• Humanoid hardware proliferated: Figure AI's Figure 02 logged over 1,250 hours at a BMW plant, loading 90,000+ parts across 30,000+ vehicles. Yet most milestones remain framed as future ambitions rather than verified operations.

Autonomous vehicles: the deployment exception

Self-driving cars reached mass-scale deployment in 2025. Waymo operated roughly 2,500 robotaxis across five U.S. cities, recording around 450,000 weekly trips. In China, Baidu's Apollo Go provided approximately 11 million fully driverless rides — a 175% year-over-year increase, up from 1.5 million in 2022. Deployments so far concentrate in favorable-weather areas with off-site humans available to take over.