ذكاء النماذج، المعايير والمسارات المهنية التقنية العالمية

The intelligence index is a composite 0–100 score that blends ten independent evaluations: SWE-bench Verified for agentic coding, GPQA Diamond for graduate-level science, MMLU-Pro for broad knowledge, AIME for mathematical reasoning, HumanEval and LiveCodeBench for code generation, Terminal-Bench for shell tool use, Humanity’s Last Exam for frontier reasoning, long-context retrieval, and LMArena Elo for human preference. Each benchmark is normalised against the current frontier maximum before weighting, so a model cannot inflate its score with a single narrow strength.

Speed is measured as steady-state output tokens per second through each provider’s official API, sampled over twenty consecutive requests at a fixed 1,000-token prompt. We additionally publish time-to-first-token, because latency-sensitive chat interfaces care far more about the first token than about sustained throughput. Both figures are re-measured monthly since providers change serving infrastructure without notice.

Cost per task models a realistic agentic workload rather than a single round trip. We assume a 45,000-output-token task with a 10,000-token input context, apply each model’s published input and output prices per million tokens, and add reasoning-token overhead for models that expose a thinking mode. Batch and cached-input discounts are excluded so the figure reflects the default interactive path most teams actually pay.

Every figure is sourced from provider release specifications, official model cards, and independently reproduced benchmark runs. We do not republish third-party leaderboard numbers without re-running the evaluation ourselves, and each model page cites its primary source and last-verified date so readers can audit any claim.

Pricing and availability are checked weekly. Benchmark scores are re-run whenever a provider ships a point release that changes model behaviour, and at minimum once per month. The updated date shown on the leaderboard always reflects the most recent full re-evaluation, not the last cosmetic edit.