Artificial General Intelligence (AGI) is usually defined as systems capable of solving a broad class of intellectual tasks at the level of a competent human and of transferring skills across domains without task-specific retraining. AGI differs from “narrow” AI in its universality, and from hypothetical superintelligence (ASI) in the absence of systematic superiority over humans. In practice, no strict, operationalizable, and generally accepted criterion for reaching AGI exists, which makes any claim of its “achievement” a matter for methodological caution.

The period from 2024 through the first half of 2026 proved pivotal: the center of gravity of development shifted definitively from academic teams to industry, while capital, access to compute, and the quality of internal evaluation systems became decisive competitive factors alongside model architecture.

The methodological basis of this review is a structured analysis of primary and authoritative secondary sources. Primary sources were prioritized for model capabilities, benchmark results, system characteristics, regulatory developments, and official corporate announcements; these included developer system cards, technical reports, peer-reviewed publications, regulatory documents, and official disclosures. Reputable secondary sources, including major financial and news organizations, were used where primary documentation was unavailable or where independent reporting was required to establish transaction values, company valuations, financing events, and other market developments.

Throughout the analysis, developer-reported performance was distinguished from independently verified results. Benchmark figures were treated as independently verified only where the evaluation methodology, model version, and source of the reported result could be identified. Where evidence was incomplete or materially inconsistent across sources, the relevant claim was qualified rather than treated as established fact.

Material discrepancies between developer-reported and independently reproduced results were examined individually, taking into account differences in model version, evaluation harness, benchmark configuration, sampling procedure, and reporting date. No developer-reported benchmark result was treated as independently verified solely on the basis of numerical proximity.

Taken together, the data support several generalizations. First, industry's lead over academia has become definitive. Second, there is a paradox: alongside rapid capability growth, a persistent failure on long-horizon agentic tasks remains, and transparency is declining. Third, the regulatory factor (the AI Act, the Grok case) is becoming an independent variable.

Conclusions about approaching AGI are best drawn in stages. A trigger for revision should be the sustained crossing of thresholds — nominally, above 50% on the pass^k metric on τ2-bench and above 30% on ARC-AGI-2 — by several independently verified systems. Marketing claims about a specific “probability of AGI” should be excluded from the evidence base.

By mid-2026 the race to AGI is a multipolar parity among a small group of leaders (OpenAI, Google DeepMind, Anthropic), with open Chinese models closing the gap rapidly and the influence of capital, infrastructure, and regulation growing.

The main methodological conclusion is the need for a strict distinction between reported and independently verified results, and for a cautious, staged interpretation of any claims of reaching AGI. Further research should focus on the reliability of agentic systems, the validity of benchmarks, and restoring transparency at the frontier.