AI Reasoning Models Are Closing the Gap With Human Experts

An 8-billion-parameter open-source model now outperforms Google’s Gemini 2.5 Flash on advanced math competitions. GPT-5.2 scores 96.1% on Mock AIME problems that stump most undergraduate math students. And Google’s Gemini 3.1 Pro hits 94.3% on PhD-level scientific reasoning questions – a score that exceeds what many domain experts achieve. The race toward human-level AI reasoning has produced results that would have seemed implausible even two years ago.

But the picture is far more complicated than any headline suggests. The same models that dominate math and science benchmarks score just 11.5% on frontier physics problems. On Humanity’s Last Exam – a benchmark specifically designed to test reasoning on questions that AI cannot memorize – the best model without tools manages only 37.5%. These numbers reveal a fundamental tension: AI reasoning models have reached or surpassed human performance on some novel benchmarks while remaining far from expert-level on others. Understanding where these models excel and where they fail is now one of the most consequential questions in technology.

This article breaks down the current state of AI reasoning as of March 2026 – the models leading the field, the benchmarks that matter, the architectural innovations driving progress, and the hard limits that remain.

What Makes a Reasoning Model Different

Reasoning models represent a distinct category within large language models. Rather than generating text through pure pattern matching, they employ structured, multi-step logical inference – explicitly working through problems before producing answers. The defining technique is Chain-of-Thought (CoT) prompting, where models outline step-by-step reasoning, often within dedicated <think> blocks that separate the deliberation process from the final output.

This matters because many real-world tasks – mathematical proofs, scientific analysis, multi-turn coding, strategic planning – require more than retrieving memorized patterns. They demand that a system decompose problems, verify intermediate steps, and revise its approach when something doesn’t hold up. Modern reasoning models are trained specifically for this through reinforcement learning with math and programming rewards, supervised fine-tuning on reasoning data, and architectural choices like Mixture of Experts (MoE) that balance massive parameter counts with computational efficiency.

The result is a new class of AI that can autonomously plan, debug code, synthesize information across million-token contexts, and produce professional deliverables. But the key question – and the one benchmarks attempt to answer – is how reliably they do so on problems they have never seen before.

The Benchmarks That Actually Matter in 2026

Not all benchmarks are created equal. Earlier tests like GLUE and SuperGLUE have been effectively solved – top models score above the human baseline of roughly 89.8 on SuperGLUE, making these tests useless for distinguishing frontier capabilities. MMLU, once considered challenging, now sees scores above 90% from leading models. The field has moved to harder, more targeted evaluations.

The Artificial Analysis Intelligence Index v4.0 aggregates 10 of these evaluations, deliberately designed so that no model can score highly by excelling at just one task type. It demands broad, reliable capability across agents, coding, knowledge, and reasoning simultaneously – run independently on dedicated hardware to ensure consistent comparisons.

Where AI Now Matches or Exceeds Human Performance

The numbers on advanced mathematics are striking. GPT-5.2 at its highest reasoning effort scores 96.1% on Mock AIME problems – harder than MATH Level 5 competition problems, where GPT-5 (high) reaches 98.1%. On the actual AIME 2025 exam, the open-source Qwen3-235B-A22B hits 89.2%, while DeepSeek-R1 achieved 91.4% on AIME 2024. These scores exceed what most human math competition participants achieve.

Scientific reasoning tells a similar story at the PhD level. On GPQA Diamond, Gemini 3 Pro Preview leads at 92.6%, followed by GPT-5.2 (xhigh) at 91.4% and Claude Opus 4.6 at 90.5%. The human expert baseline on these questions sits around 50-60%, meaning top AI models now roughly double what PhD-level specialists score on novel scientific problems in their own fields.

Agentic capabilities – where models autonomously use tools, write and execute code, and complete multi-step workflows – have also surged. Claude Sonnet 4.5 tops agentic benchmarks at 70.6% on TAU-Bench, outpacing GPT-5 medium at 65.0%. On SWE-Bench Verified, Claude Sonnet 4.5 leads at 64.8%, demonstrating that these models can fix real-world software bugs with meaningful reliability.

Where AI Still Falls Short

The CritPt benchmark delivers a reality check. Designed to test frontier physics reasoning at the research level, it exposes how far even the best models remain from genuine scientific capability. GPT-5.2 with extended reasoning leads – at just 11.5%. No other model comes close to even that modest figure.

Humanity’s Last Exam tells an equally sobering story. Without tool use, the top model – Gemini 3 Pro Preview – scores 37.52%. The next best, GPT-5 from August 2025, manages 25.32%. These questions were specifically vetted to ensure they cannot be found through web searches, and any question that top AI models answered correctly during development was discarded. The benchmark keeps only questions that stumped the AIs, uses zero-shot evaluation with no fine-tuning, and applies strict grading with no partial credit.

With tool use enabled, scores improve significantly – GLM-4.7 reaches approximately 42.8% on HLE – but this still represents a substantial gap from human expert performance. The benchmark’s creators have been explicit: models unable to answer HLE-type questions should not be trusted unsupervised in domains requiring expert judgment, such as medical diagnosis, scientific research, or legal advice.

There is also the hallucination problem. On the AA-Omniscience benchmark, Google’s models lead in raw accuracy (54% and 51%) but also demonstrate hallucination rates of 88% and 85%. Anthropic’s Claude models show lower hallucination rates (48-58%), creating a critical tradeoff: a model with high accuracy but high hallucination rates may confidently produce wrong answers – arguably a more dangerous failure mode than simply refusing to answer.

The Proprietary vs. Open-Source Race

The gap between proprietary and open-source reasoning models has narrowed dramatically. Perhaps the most surprising development is that DeepSeek-R1-Distill-Qwen3-8B – a compact 8-billion-parameter model – outperforms Google’s Gemini 2.5 Flash on AIME 2025 and nearly matches the 235-billion-parameter Qwen3-235B on select tasks. This defies the assumption that bigger always means better, demonstrating that distillation techniques can compress reasoning capability into models small enough to run on modest hardware.

Open-source models like GLM-4.7 (355 billion total parameters, roughly 32 billion active) now lead in agentic coding with 84.9% on LiveCodeBench v6 and 73.8% on SWE-Bench Verified. Kimi K2, with approximately 1 trillion total parameters but only 32 billion active per token via 384 experts, handles million-token contexts through Yarn extension. These models are not just research curiosities – they power production coding assistants and strategic planning agents.

Architectural Innovations Driving Progress

Three technical shifts underpin the 2026 reasoning leap.

Mixture of Experts at Scale

MoE architectures allow models to pack trillions of parameters while activating only a fraction per token. Kimi K2 uses roughly 1 trillion total parameters but activates just 32 billion per token across 384 experts. GLM-4.7 runs 355 billion total with approximately 32 billion active. This approach delivers massive model capacity for complex reasoning without proportional compute costs – Gemini 3 Flash Reasoning outputs 198.7 tokens per second while scoring 46 on the Intelligence Index, well above the median of 26 for reasoning models in its price tier.

Extended Context and Long-Document Reasoning

Production models now routinely handle 1 million or more tokens in a single context window. Gemini 2.5 Pro demonstrated this practically by analyzing multiple research papers simultaneously, producing detailed summaries, mechanism comparisons, and comprehensive tables – outperforming OpenAI O3 on long-context synthesis tasks. Claude Sonnet 4.6 ships with a 1-million-token context window in beta.

Chain-of-Thought Training and Tool Integration

Models are now explicitly trained to generate reasoning steps, not just fine-tuned with prompts. DeepSeek-R1 and R1-Zero were trained with large-scale reinforcement learning to directly optimize reasoning capabilities. Combining CoT with tool use boosts HLE scores by 20-30%, as models can chain step-by-step logic with code execution, web search, and file analysis. Explicit CoT prompting alone improves AIME scores by 10-15%.

Practical Guidance for Choosing a Model

The right model depends entirely on the task. For research-level mathematics and science requiring multi-stage deductions, GPT-5.2 at high or extra-high reasoning effort delivers the strongest results – 96.1% on Mock AIME, 91.4% on GPQA Diamond. For multimodal tasks and long-context synthesis exceeding a million tokens, Gemini 3.1 Pro leads across the broadest range of benchmarks at competitive pricing ($2.00 per million input tokens).

Enterprise deployments prioritizing controlled, safe outputs should consider Claude Sonnet 4.6, which tops the GDPval-AA Elo leaderboard at 1,633 points for real expert-level office work and delivers near-Opus quality at Sonnet pricing. For cost-sensitive coding and reasoning workloads, GLM-4.7 or Qwen3-235B offer open-source alternatives with benchmark scores that rival proprietary systems.

• Prompting best practice: Use explicit Chain-of-Thought instructions (“Think step-by-step”) combined with tool access – this improves AIME scores by 10-15%
• Deployment sizing: Models in the 32B-235B parameter range offer the best balance of accuracy and efficiency; an 8B distilled model can match 235B giants on select reasoning tasks
• Validation: Do not rely on self-reported benchmarks from AI labs – prioritize independently run evaluations measured on uniform hardware
• Human oversight: For high-stakes domains like legal analysis, medical diagnosis, or scientific research, combine AI reasoning with human review – models still falter on ultra-novel problems (CritPt scores below 12%)

Key Takeaways

AI reasoning models in early 2026 have genuinely reached or exceeded human expert performance on several rigorous benchmarks – particularly in advanced mathematics (up to 96.1% on competition problems) and PhD-level scientific reasoning (92.6% on GPQA Diamond). Open-source models have closed the gap with proprietary systems far faster than expected, with an 8-billion-parameter model matching giants hundreds of times its size through distillation.

But the frontier remains real. On Humanity’s Last Exam, the best model without tools scores 37.5%. On research-scale physics, the ceiling is 11.5%. Hallucination rates above 80% persist even in top-performing models. The honest assessment is that AI currently performs at or above expert level on well-structured problems with clear solution paths, while remaining far from reliable on genuinely novel challenges requiring deep, open-ended reasoning. For organizations deploying these systems, the implication is clear: match the model to the task, validate with independent benchmarks, and keep humans in the loop where the stakes are high.

Sources

• AI Model Leaderboard 2026: Complete Ranking Guide
• Top 10 Open-Source Reasoning Models in 2026
• AI Benchmarks Ranking and Strategy Guide
• Definitive Guide to AI Benchmarks
• The Best AI Models So Far in 2026
• LLM Benchmarking Framework Guide
• AI Model Benchmarks – LM Council (Mar 2026)
• 25 AI Benchmarks: Model Evaluation Examples
• The Ultimate Guide to AI Benchmarks (2026)
• 5 Best AI Reasoning Models of 2026: Ranked