Artificial Intelligence trend analysis is framed as a way to measure how models change over time across capability, cost, speed, openness, and geography. The tracking covered here draws on public benchmarks including GPQA, HumanEval, MMLU, AIME, and SWE-Bench, along with provider pricing, latency and throughput from proxy traffic, release timelines, and human-preference ratings. The dataset spans 500+ models and 50+ benchmarks and is updated continuously, with the goal of helping users compare models, estimate workloads, and follow the direction of the industry.
The competitive picture is shifting quickly. Models that led benchmarks six months ago are now mid-pack, while the clearest advances are in reasoning depth, multimodal understanding, and parameter efficiency. A 7B model today can hit scores that took 70B+ parameters last year. OpenAI, Anthropic, Google, xAI, and Meta are raising the bar in proprietary systems, while Mistral, DeepSeek, Qwen, and Alibaba are pushing open-weight models closer to frontier performance. US labs still lead most benchmarks, but Chinese labs including DeepSeek, Alibaba, and ByteDance are closing the gap, especially in reasoning and coding.
Open and closed model competition is tightening. Llama, Mistral, and Qwen now match or beat GPT-4 on several benchmarks, and capable local deployment is increasingly practical. Open-weight releases typically lag proprietary models by 6 to 18 months, and that window keeps shrinking. Multimodal capability is also becoming standard at the frontier, while reasoning-focused models are often trading speed for accuracy. Human preference remains a separate lens, highlighting where benchmark performance aligns with user judgments and where it does not.
Costs are falling just as fast as capability is rising. GPT-4-level performance cost ?/M tokens in 2023. Today you can get it for under ?/M. The current trend is described as roughly 10x per year for the same level of performance, driven by competition, better infrastructure, and model efficiency. On benchmarks, GPQA scores went from around 50% to 75%+ in just 18 months. Some evaluations may be starting to saturate, but the broader pattern remains one of rapid improvement in model quality, lower inference costs, and denser competition across both proprietary and open-weight systems.
