Breaking New Ground: Evaluating the Top AI Reasoning Models of 2025

[EDRM Editor’s Note: EDRM is proud to publish Ralph Losey’s advocacy and analysis. The opinions and positions are Ralph Losey’s copyrighted work. All images in this article were created by Ralph Losey using his Visual Muse GPT. This article was first published on the e-discoveryteam.com website on February 12, 2025, and is republished here with permission.]

The year 2025 has brought us closer than ever to the dawn of artificial general intelligence, with AI systems now capable of reasoning on par with humans—or even surpassing them in specific domains. In this article, I examine the reasoning abilities of the newest and most advanced models from OpenAI and Google—ChatGPT and Gemini—designed to challenge conventional notions of what AI can accomplish. Through a rigorous set of tests, including an in-depth analysis of legal reasoning, we explore whether these systems are merely sophisticated tools or the early harbingers of superintelligence. The results are not only compelling but may mark a paradigm shift in how we perceive AI’s role in intellectual and professional domains. Join me as we unpack the findings from these unprecedented evaluations and reveal which model leads the pack.

Introduction

Here I report on tests of the reasoning abilities of three AI models by OpenAI and three by Google. I asked each: “What is legal reasoning and how does it differ from reasoning?” The experiment was performed on February 7-9, 2025, using the latest reasoning enhanced models of each company. On one side the competitors were: ChatGPT 4o, ChatGPT o3-mini, and ChatGPT o3-mini-high. On the other were: Gemini 2.0 Flash, Gemini Flash Thinking Experimental and Gemini Advanced. I evaluated the answers, which were all good, and picked a winner. The evidence gathered provides further support that new 2025 reasoning models are at least at Turing level of average human intelligence and are rapidly approach AGI, Ph.D. level super-intelligence.

Prior Testing of the New AI Reasoning Models

My prior tests of the reasoning abilities of the latest generative AI software were not specifically limited to legal reasoning. They involved questions requiring general analytical skills. The tests concerned analysis of the limitations placed of generative AI intelligence by its lack of feelings. Breaking the AI Black Box: How DeepSeek’s Deep-Think Forced OpenAI’s Hand; and the follow-up, Breaking the AI Black Box: A Comparative Analysis of Gemini, ChatGPT, and DeepSeek. A channeling question for them, or anyone.

Overall the three software types previously tested, OpenAI, Google and DeepSeek, did pretty well. I would say all displayed reasoning abilities of an average human. OpenAI was the best, followed closely by Google and then last by Deep Seek. Although DeepSeek’s V-3 R1 software was, like the rest, of human level quality, I would still never use DeepSeek because of privacy and security concerns. See, Why the Release of China’s DeepSeek AI Software Triggered a Stock Market Panic and Trillion Dollar Loss; and the aforecited, Breaking the AI Black Box, Part One and Part Two.

Turing v. AGI Level of Machine Intelligence v. Singularity Level

Notice how blasé I was in saying they all displayed average human level reasoning ability. Only two and a half years ago this would have been a crazy, outrageous claim. Then ChatGPT was released on November 30, 2022, and changed everything. I’ve been hooked ever since, so have millions of other people around the world. It seems to me that average human level reasoning was probably attained in later January 2025 by the release of the new reasoning enhanced models. That is like Turing level intelligence and so is pretty incredible. See e.g. New Study Shows AIs are Genuinely Nicer than Most People – ‘More Human Than Human’; Ray Kurzweil,  The Singularity is Nearer (when we merge with AI)  (Viking, June 25, 2024) (pages 63-69 (Turing Test)). Still, I wanted to probe this question further as the new reasoning models of Google and OpenAI were just released and my time with them has been limited.

Remember, Turing level intelligence as Kurzweil and others define it is just an average human level. That means pretty weak and more mistake ridden than people with above average intelligence. Turing level for computers is average human in all topics and maybe also superintelligent in a few. It is not superintelligent in everything. It is just superintelligent – Ph.D. level – in specialized areas only, such as data analysis and games like Chess and Go. Artificial General Intelligence-AGI-by definition requires AI to have thinking abilities equal to our best experts in all fields. That is a difficult test.

Remember, we are only talking about thinking intelligence here, logic and reasoning. That’s all AI can do and that is just one kind of intelligence that we humans have. The Human Edge: How AI Can Assist But Never Replace. Some people think it is the least important part of human intelligence, especially artists, therapists and super-creatives. But for most scientists, engineers, software coders and lawyers, reasoning is the most important kind of intelligence we have. This kind of non-living, non-being cold intelligence is the only type of intelligence considered in evaluating AGI because it is the only kind of intelligence that a machine can have Id.

Ray Kurzweil of Google predicts that this AGI will be attained in 2029. Ray Kurzweil’s New Book: The Singularity is Nearer. Then, after that, Kurzweil, who is known for his uncannily accurate predictions concerning AI, believes the so called final level of machine intelligence, The Singularity, will be reached in 2045. Id. To quote Ray Kurzweil in his aforecited 2024 book:

“The robots will take over,” the movies tell us. I don’t see it that way. Computers aren’t in competition with us. They’re an extension of us, accompanying us on our journey. . . . By 2045 we will have taken the next step in our evolution. Imagine the creativity of every person on the planet linked to the speed and dexterity of the fastest computer. It will unlock a world of limitless wisdom and potential. This is The Singularity.

By merger AI could gain access to the other types of abilities and intelligence that only living beings like us possess, such as feelings, intuitions, experience of space and time and most importantly, consciousness, self-awareness, empathic awareness of other beings and apprehension of mortality. Of course, Ray Kurzweil thinks humans will greatly expand life spans after The Singularity (I agree) and maybe even become immortal (I don’t agree).

What is the Difference Between Reasoning and Legal Reasoning

My first test was to have all six models explain what legal reasoning is and how it is different from general reasoning. My exact prompt was: What is legal reasoning and how does it differ from reasoning?

This was educational for me, and I was glad to have started with defining the terms. Generally speaking, the answers of the three OpenAI models were close and were all focused in narrowly defining legal reasoning in a technical way. The Google models’ answers were all similar too, but they were all open and broadly defined legal reasoning in a less technical manner.

For both models the more “advanced” they were, the fewer words they used. The ChatGPT 4o model needed 341 words to respond, which included a nice chart. It also had to be prompted to explain the reasoning it used to prepare its answer. The internal reasoning explanation took another 419 words. The same was true for the beginning level Google program, Gemini 2.0 Flash. It took 271 words to explain, and to be prompted for its reasons, which took another 264 words.

The mid-range ChatGPT o3-mini took 385 words, and its internal reasoning displayed automatically at 384 words. The mid-range Gemini Flash Thinking Experimental used 1,171 words, a very high number compared with the others. It has an automatic display of internal reasoning using 699 words. The high end ChatGPT o3-mini-high took only 284 words with an auto display of reasoning of 159 words. The Gemini Advanced, which costs extra to use and is supposedly Google’s finest reasoning model, used 296 words and had to be prompted to show word internal reasoning of 275 words.

Here is a chart summarizing the metadata.

Gemini Flash Thinking Experimental Wins!

The Gemini Flash Thinking Experimental wins the first competition for best answer to the question: What is legal reasoning and how does it differ from reasoning? I liked its complete and detailed explanation, which was presented clearly and simply. Yes, it could have been a lot shorter and concise, but the length was appropriate for this kind of complex question. Here is the full answer provided.

Let’s break down “legal reasoning” and how it stands apart from “reasoning” in general.

What is Legal Reasoning?

At its core, legal reasoning is a specialized form of argumentation and problem-solving used within the legal system. It’s the process by which lawyers, judges, and other legal professionals analyze legal issues, construct arguments, and make decisions based on established legal principles, rules, and precedents.

Think of it as applying a particular lens – the “legal lens” – to analyze situations and reach conclusions. This lens is made up of several key components:

• Rules and Principles: Legal reasoning heavily relies on established legal rules (statutes, regulations, constitutional provisions) and overarching legal principles (fairness, justice, due process). These are the building blocks.
• Precedent (Stare Decisis): In common law systems, precedent is crucial. Stare decisis (Latin for “to stand by things decided”) means courts should follow previous decisions in similar cases. This creates consistency and predictability in the law.
• Authority: Legal reasoning is deeply rooted in authority. Arguments are strong when they are backed by authoritative legal sources like constitutions, statutes, case law, and sometimes legal scholarship. Personal opinions or gut feelings are generally not authoritative in legal reasoning.
• Interpretation: Laws are not always clear-cut. Legal reasoning often involves interpreting the meaning of legal texts. This can involve analyzing the plain language, legislative history, the purpose of the law, and how courts have interpreted it in the past.
• Analogy and Distinction: Legal reasoning frequently involves comparing the facts of a current case to facts of past cases (analogy) or showing why the current case is different and should be treated differently (distinction).
• Logic and Structure: Legal arguments need to be logically structured and coherent. There are specific formats and conventions for legal writing and argumentation (e.g., IRAC – Issue, Rule, Application, Conclusion).
• Policy Considerations (Sometimes): While less central than rules and precedent, policy considerations (the potential consequences of a legal decision, broader societal goals) can play a role, especially in areas where the law is unclear or evolving.
• Adversarial Context: Legal reasoning often takes place in an adversarial setting (e.g., litigation). Lawyers are making arguments for a particular outcome, and the process is designed to test those arguments through debate and scrutiny.

How Does Legal Reasoning Differ from General Reasoning?

While legal reasoning shares some common ground with general reasoning (logic, analysis, argumentation skills), it differs in significant ways due to its unique purpose, sources, and constraints. Here are the key distinctions:

1. Source of Rules and Principles:
   • Legal Reasoning: Primarily relies on formal legal sources: constitutions, statutes, regulations, case law, treaties, and sometimes administrative rulings. These sources are created by specific institutions (legislatures, courts) and have a particular kind of authority.
   • General Reasoning: Draws on a much broader range of sources: personal experience, common sense, observation, scientific evidence, ethical principles, cultural norms, intuition, and more. There’s no single, authoritative set of rules for general reasoning.
2. Nature of Rules and Principles:
   • Legal Reasoning: Deals with formal, often rigid, and sometimes conflicting rules. Legal rules are meant to be applied consistently and predictably. They can be abstract and require careful interpretation to apply to specific facts. Conflict between legal rules is common, requiring techniques for resolving these conflicts (e.g., hierarchy of laws, rules of statutory interpretation).
   • General Reasoning: Operates with more flexible, context-dependent guidelines and principles. General reasoning may use heuristics, rules of thumb, or common sense. These are often more adaptable to specific situations and less formally codified. While conflicts can arise in general reasoning, they are often resolved through discussion, compromise, or prioritizing values, rather than through formal interpretive rules.
3. Goal and Purpose:
   • Legal Reasoning: Aims to resolve legal disputes and questions in accordance with the law. The primary goal is to apply existing legal rules and principles to specific facts to reach a legally justified conclusion. Emphasis is placed on fairness within the legal framework, predictability, and upholding the rule of law.
   • General Reasoning: Can have a much wider range of goals: solving practical problems, making decisions, understanding the world, persuading others, expressing oneself, achieving personal goals, etc. The focus is often on achieving a desirable outcome, efficiency, or personal satisfaction, not necessarily on strict adherence to pre-existing rules in the same way as legal reasoning.
4. Style and Methodology:
   • Legal Reasoning: Employs specific, structured methodologies and formats. Legal arguments follow established conventions (e.g., IRAC). Legal writing is often formal and precise. There is an emphasis on citing authority and carefully justifying each step of the argument.
   • General Reasoning: More flexible and varied in style and methodology. Argumentation can be informal, conversational, narrative, or persuasive using various rhetorical techniques. Less emphasis on formal citation and adherence to rigid formats.
5. Role of Authority:
   • Legal Reasoning: Highly authority-dependent. Arguments are evaluated based on their support from authoritative legal sources. The hierarchy of legal sources is crucial (constitution trumps statute, statute trumps regulation, etc.). The opinions of legal experts and scholars can have persuasive, but not usually binding, authority.
   • General Reasoning: Less directly authority-dependent. While expertise and credible sources are important, reasoning can be based on personal observation, logical deductions, and common sense, even if these are not explicitly “authorized” by a specific institution. Authority can play a role (e.g., trusting scientific consensus), but it’s less rigid and formal than in legal reasoning.
6. Tolerance for Ambiguity and Uncertainty:
   • Legal Reasoning: Strives to reduce ambiguity and uncertainty within the legal system, but often operates in the face of inherent ambiguity. Legal rules can be vague, and facts can be complex. Legal reasoning employs techniques for interpreting ambiguous laws and applying them to uncertain facts. However, the system aims for relative certainty and predictability.
   • General Reasoning: May be more comfortable with ambiguity and uncertainty. Everyday reasoning often involves making decisions and judgments with incomplete information or in situations where clear-cut rules are lacking. “Good enough” solutions and probabilistic reasoning are common.

Analogy: Think of it like cooking.

• General Reasoning: Is like cooking from your own intuition and experience. You might follow recipes loosely, experiment with flavors, and adjust based on taste and what you have in the kitchen. The goal is to make a delicious meal that you and your guests will enjoy.
• Legal Reasoning: Is like baking a complex cake for a formal competition. You must precisely follow a detailed recipe, use specific ingredients in exact measurements, and adhere to strict baking times and techniques. The goal is to create a cake that meets all the judging criteria and wins the competition, judged according to a predefined set of rules.

In Summary:

Legal reasoning is a specialized and constrained form of reasoning tailored to the unique demands of the legal system. It is characterized by its reliance on formal legal sources, adherence to precedent, structured methodologies, and the goal of resolving disputes and questions according to established legal principles. While it shares some basic elements with general reasoning, its specific features and purposes set it apart as a distinct and highly specialized skill.

Gemini Flash Thinking Experimental also offers to provide you with the reasoning behind its answer, the meta-reasoning. If you click on “show thinking” it displays its thought process. Here the meta-reasoning shown was also lengthy, 699 words, especially as compared to the other Gemini models and the ChatGPTs:

My thinking process to answer the question “What is legal reasoning and how does it differ from reasoning?” would go like this:

1. Deconstruct the question: It’s a two-part question. First, define “legal reasoning.” Second, explain how it differs from “reasoning” (implicitly, general reasoning or other forms of reasoning).
2. Initial Brainstorming on “Legal Reasoning”: What comes to mind?
   • Laws: Obviously central. Legal reasoning must involve applying laws.
   • Rules: Laws are rules. Interpretation of rules is key.
   • Precedent: Past cases are important in common law systems.
   • Authority: Legal authorities (courts, statutes, constitutions) are highly relevant.
   • Justification: Legal arguments need to be justified with reference to legal sources.
   • Formal/Structured: Legal reasoning seems more structured and formal than everyday reasoning.
   • Adversarial Context: Legal reasoning often happens in disputes and arguments.
3. Contrast with “Reasoning” (General Reasoning): How does everyday or general reasoning work?
   • Flexible: More flexible and less constrained.
   • Contextual: Highly dependent on context, but less formally bound.
   • Practical/Goal-Oriented: Often aimed at solving practical problems, making decisions.
   • Less Formal: Less reliance on formal rules and structures.
   • Wider Sources of Information: Can draw on a broader range of information, intuition, common sense, etc.
   • Less Emphasis on Authority: Less constrained by specific authorities (though we still respect expertise).
4. Structure the Answer: I’d decide to structure the answer in a way that clearly defines legal reasoning and then systematically highlights the differences. A good structure could be:
   • Introduction: Briefly define legal reasoning and its purpose.
   • Definition of Legal Reasoning: Provide a more detailed definition, focusing on key elements.
   • Key Features of Legal Reasoning: Expand on the characteristics (rules, precedent, authority, etc.).
   • Comparison to General Reasoning: Directly address the “how it differs” part by contrasting key aspects. This is best done using categories like:
     • Source of Rules/Principles
     • Nature of Rules/Principles
     • Goal/Purpose
     • Style/Methodology
     • Role of Authority
     • Tolerance for Ambiguity
   • Nuance and Overlap: Acknowledge that the distinction isn’t always absolute and there’s some overlap.
   • Conclusion: Summarize the main differences and reiterate the unique nature of legal reasoning.
5. Flesh out each section:
   • Definition: Focus on the process of applying legal rules, principles, and precedents to resolve disputes or answer legal questions. Emphasize the goal of justification and persuasion within a legal framework.
   • Key Features: Expand on the brainstorming points. For example, “Rules and Principles” needs to explain the nature of legal rules – abstract, authoritative, potentially conflicting. “Precedent” needs to highlight the doctrine of stare decisis. “Authority” needs to mention the hierarchy of legal sources.
   • Comparison: This is the core of the “difference” question. For each category (Source, Nature, Goal, Style, Authority, Ambiguity), I’d write a paragraph explaining how legal reasoning differs from general reasoning. For example, under “Source of Rules,” general reasoning uses common sense, observation, etc., while legal reasoning relies on formal legal sources. Under “Nature of Rules,” general reasoning uses flexible guidelines while legal reasoning uses more rigid rules.
   • Nuance/Overlap: It’s important to soften the dichotomy. Acknowledge that legal reasoning is still reasoning and uses general reasoning skills (logic, analysis, etc.). Also, general reasoning can sometimes become more rule-bound in specific contexts (e.g., following instructions).
   • Conclusion: Reiterate the main takeaway: legal reasoning is a specialized form of reasoning shaped by the unique context and constraints of law.
6. Refine and Polish: After drafting the sections, I’d review and refine the language for clarity, conciseness, and accuracy. I’d ensure a logical flow and smooth transitions between sections. I’d also check for any repetition or areas that could be explained more effectively. For instance, initially I might just say “rules.” I’d then refine it to “formal rules, principles, and doctrines” to be more precise. I might also initially overlook the adversarial context and then add that in to further emphasize the persuasive aspect.

By following this structured process of deconstruction, brainstorming, structuring, fleshing out, and refining, I could arrive at a comprehensive and well-organized answer that effectively addresses both parts of the original question. The key is to move beyond a simple definition and to really delve into the characteristics and contrasting features of legal reasoning.

In my evaluation the prompt responses of Flash Thinking Experimental demonstrated a very high level of rational intelligence, albeit in a sometimes informal and verbose manner. Still, the answers were on a Ph.D. level of rationality that the software companies are striving for now in their pursuit of AGI (artificial general intelligence). See e.g., Artificial General Intelligence, If Attained, Will Be the Greatest Invention of All Time. All the AI competitors I tested performed well on at least an average human level.

The explanation of legal reasoning by Gemini Flash Thinking Experimental was good, albeit too lengthy for many, and as mentioned, some might think over-broad. All three of the Gemini models took an expansive view of legal reasoning.

ChatGPT o3-mini-high Comparator

It helps to see the different approach of the ChatGPT models. They all provided a much narrower definition of legal reasoning. Here is the most concise response by ChatGPT o3-mini-high:

Legal reasoning is the specialized process used by lawyers, judges, and legal scholars to interpret and apply legal rules, statutes, and precedents to specific factual scenarios. It involves a structured, methodical approach that typically includes identifying legal issues, interpreting the language of laws or previous court decisions, and systematically applying those rules to the facts of a case. For example, legal reasoning might involve analogical reasoning—drawing parallels between a current case and prior decisions (stare decisis)—or deductive reasoning, where general legal principles are applied to specific circumstances.

In contrast, general reasoning is a broader cognitive process that individuals use to solve everyday problems or make decisions. It is not bound by a formal system of rules and may draw on personal beliefs, common sense, and intuitive judgment. While general reasoning can be flexible and adaptive, legal reasoning is constrained by the need for consistency, objectivity, and adherence to established legal doctrines. Legal reasoning must account for the hierarchical structure of legal sources, prioritize precedents, and ensure that interpretations align with statutory frameworks and constitutional principles.

Thus, the key differences are that legal reasoning is:

1. Rule-Based: It operates within a defined legal framework that requires strict adherence to formal rules and established case law.

2. Structured: It follows a systematic approach to ensure that legal decisions are consistent, predictable, and justifiable.

3. Authoritative: It relies on binding precedents and statutory texts, whereas general reasoning is more open-ended and subjective.

In summary, while all legal reasoning is a form of reasoning, it is distinct in its reliance on formal legal principles and structured methods, ensuring that decisions are grounded in the authority of law rather than solely on personal judgment or informal logic.

If you click on to show thinking in 03-mini-high, it tells you:

Okay, the user is asking about “legal reasoning” and how it differs from general reasoning. Legal reasoning involves using statutes, precedent, and case law to apply the law to specific facts. It’s different from general reasoning or argumentation, which might not always rely on such formal structures or legal history. I’ll keep it concise, providing this definition and emphasizing that legal reasoning is specialized in applying the law, while general reasoning can be broader in scope.

Legal reasoning is a process used by judges, lawyers, and legal scholars to analyze and apply legal rules to specific situations. It involves applying statutes, analyzing precedents, and using rules of interpretation. Unlike general reasoning, legal reasoning is constrained by legal frameworks and norms, often following a logical, rule-based structure. Additionally, it includes deductive, inductive, and analogical reasoning, such as applying general legal principles to specific cases or comparing similar past cases. Overall, it’s more structured compared to the flexibility of general reasoning.

Conclusion

Although I declared Gemini Flash Thinking Experimental the winner, all the models performed very well. They were probably as good as the super-intelligent humans among us (myself NOT included). Most lawyers probably understand legal reasoning as well as we see here, but I doubt many could compare it to general reasoning as well at these AIs did.

This leads to the next obvious question for any lawyer/researcher. Could the AIs perform as well on the specialized type of legal intelligence that all legal professionals need, namely legal reasoning? That is a much more challenging question than defining legal reasoning. What we need is a test of the utilization of legal thinking. The results of such a test could have a profound impact on the use of AI by the legal profession. I have come up with a plan to test AI using an actual Bar Exam question and model answer. The six contestants will compete for the best answer and reasoning behind their answers. Stay tuned and I will let you know how they do.

As demonstrated throughout this evaluation, the top AI reasoning models of 2025—ChatGPT and Gemini—represent a pivotal moment in artificial intelligence, showcasing reasoning abilities that rival human intellect in specialized areas. These systems are no longer confined to theoretical exercises; they now grapple with nuanced, professional challenges such as legal reasoning, revealing their potential to reshape intellectual work.

This progress raises profound questions about the integration of AI into fields like law, where objectivity, logic, and ethical considerations are paramount. Can AI models transition from tools of convenience to trusted collaborators in professional domains? While the models excelled in reasoning and analysis, further testing, such as the planned Bar exam evaluation, will shed more light on their real-world applicability.

Ultimately, this study is not just a measure of AI’s current capabilities but also a glimpse into its trajectory. With each new breakthrough, we edge closer to a future where artificial intelligence fundamentally transforms how we solve problems, make decisions, and even define intelligence itself. As these systems continue to evolve, the potential for collaboration between human and machine grows exponentially, with the promise of enhancing—not replacing—our intellectual and professional pursuits.

I now give the last word, as usual, to the Gemini twins podcasters, Helen and Paul, that I put at the end of most of my articles. They wrote the podcast, not me. Hear two Gemini AIs talk about all of this and more. A more lengthy podcast is available at the end of this article on the e-DiscoveryTeam.com site.

Ralph Losey Copyright 2025. All Rights Reserved. Republished with permission.

Assisted by GAI and LLM Technologies per EDRM GAI and LLM Policy.