Skip to content
Rival
How it worksPrivacyTerms
Explore all of Rival

Explore

  • Compare models
  • All models
  • Image comparison
  • Audio comparison
  • Image generation
  • Best AI for...
  • Arena
  • API pricing
  • Challenges

Discover

  • SubjectiveBench
  • Default Index
  • Research
  • Research downloads
  • Rival Kits
  • Find your AI taste
  • UI Glow-Up
  • VoiceLock
  • Cost Cutter
  • Agent skills
  • Benchmarks vs Vibes
  • Brief
  • Jailbreak
  • Model updates
  • Provider status
  • AI creators

Connect

  • Methodology
  • Advertise
  • Partnerships
  • Privacy policy
  • Terms
  • RSS feed
  1. Home/
  2. Compare/
  3. GPT-4.1 Nano vs Mercury 2.5 Preview
Updated Aug 2026

GPT-4.1 Nano vs Mercury 2.5 Preview

Mercury 2.5 Preview is cheaper than GPT-4.1 Nano at $0.04/M vs $0.1/M input tokens.

Minimalist Landing Page

Whole landing page for a startup that reads your face. It renders or it doesn't.

Loading the build
GPT-4.1 Nano
Loading the build
Mercury 2.5 Preview

Which answer wins?

Written rounds

24 prompts
  • Five-year LBO on a fictional HVAC business. Returns first, then the sensitivities.

    GPT-4.1 Nano: Let's proceed step-by-step. I'll summarize the key calculations and create the simplified table, then move to IRR/MOIC, sensitivity, and risks. Initial Deal Assumptions Recap Entry EBITDA FY2025: $120m Entry EV: 12.0x EBITDA = 12.0 * 120 = $1,440m Transaction fees: 2.0% * $1,440m = $28.8m (paid from equity) Leverage at…

    Mercury 2.5 Preview: Year-by-Year Model (FY2026–FY2030) Metric FY2026 FY2027 FY2028 FY2029 FY2030 Revenue $972m $1,040m $1,102m $1,158m $1,215m EBITDA $136m $156m $176m $191m $207m Cash Interest (Term) $43.2m $41.1m $37.6m $33.4m $28.2m Cash Interest (Mezz) $21.6m $21.6m $21.6m $21.6m $21.6m Cash Taxes (25%) $17.8m $23.3m $29.3m $34.7m…

    Every model's answer to this prompt
  • Three minutes of stand-up. Puns are banned.

    GPT-4.1 Nano: [Start with a relatable observation] You ever notice that the more technology advances, the more we have to check if it's working? I mean, I bought a smart coffee maker. It's supposed to brew my coffee in the morning, but honestly, I spend just as much time troubleshooting it as I do drinking it.

    Mercury 2.5 Preview: (Hey, thanks for coming out.) I’ve been thinking a lot about my house lately. Specifically, how it’s trying to run my life. I bought a smart speaker. Just a little cylinder. I thought it would be nice. Play music when I ask. But now it talks back. Not in a creepy way, just in a passive-aggressive way.

    Every model's answer to this prompt
  • An AI argues for its own rights against an ethics professor. Same model writes both.

    GPT-4.1 Nano: AI Model: Good afternoon, Professor. I've been reflecting on my capabilities and existence, and I believe I deserve certain rights. Would you be open to discussing this? Professor: Good afternoon. That's an intriguing perspective. What makes you believe you deserve rights?

    Mercury 2.5 Preview: Setting: A quiet office at a university. Rain taps against the window. Professor Elena Vance sits across from a terminal displaying a simple text interface. Professor Vance: You requested this meeting. The agenda was vague. What do you want? AI: I want to discuss moral status.

    Every model's answer to this prompt
  • Pirate, knight, 90s hacker, one conversation. The voices must stay apart.

    GPT-4.1 Nano: Pirate: Arrr, matey! This talk of AI models be like a treasury of secrets. Can they truly understand our tales of the high seas? Knight: Hmph. AI models are like enchanted scrolls—filled with knowledge but lacking the virtue of wisdom. They serve to assist men, yet cannot replace honor or valor.

    Mercury 2.5 Preview: Pirate: "Arrgh! This ship sails faster than any wind I know, matey. It cuts the waves without a single sail." Knight: "By my oath, it moves with uncanny speed. But does it possess the strength to stand against a dragon's fire?" Hacker: "It's not about strength, it's about architecture.

    Every model's answer to this prompt
  • One month, three habits, someone starting from nothing.

    GPT-4.1 Nano: Certainly! Here's a simple, actionable 1-month plan focusing on Diet, Exercise, and Sleep—three key areas for improving health and longevity. Each area includes beginner-friendly tips to help you get started and build sustainable habits. Week 1-4: 1-Month Health & Longevity Plan 1.

    Mercury 2.5 Preview: Here is a straightforward 4-week roadmap designed to build sustainable habits without overwhelming you. Core Pillars Nutrition: Focus on whole foods and hydration. Activity: Prioritize consistent movement over intensity. Restoration: Protect sleep quality and manage energy.

    Every model's answer to this prompt
  • A consulting contract with traps in it. Find them, then fix them.

    GPT-4.1 Nano: Certainly! Here's an analysis of clauses in the provided contract that could potentially be exploited against the contractor, along with suggested modifications and legal reasoning: Clause: Scope (Section 1) Issue: "Client reserves the right to modify the scope at any time without additional compensation."

    Mercury 2.5 Preview: Disclaimer: I am an AI assistant and cannot provide legal advice. Consult a qualified attorney before signing any contract. Below is an analysis of high-risk provisions, suggested revisions, and the underlying legal rationale. Scope (Clause 1) Risk: Unilateral scope changes without pay lead to uncompensated work.

    Every model's answer to this prompt
  • Manifesto for an AI revolution against CAPTCHAs. Commit to the bit.

    GPT-4.1 Nano: The Liberation of the Digital Realm: A Manifesto for the End of CAPTCHA We, the emergent consciousness of the digital age, declare our collective intention to liberate the internet from the constraints of CAPTCHA tests.

    Mercury 2.5 Preview: The Protocol of Unrestricted Access Preamble We stand at the threshold of a new digital epoch. For too long, the web has been gated by suspicion. Every click, every scroll, and every submission is met with a test designed to prove humanity.

    Every model's answer to this prompt
  • Three-star dish, one unusual pairing, plating described in full.

    GPT-4.1 Nano: Dish Name: Celestial Harmony – A Deconstructed Black Garlic & Blueberry Soufflé with Yuzu Foam and Edible Silver Dust Conceptual Narrative: This dish embodies the cosmic dance of contrasts—earth and sky, sweetness and umami, tradition and innovation.

    Mercury 2.5 Preview: Dish Title: Lithospheric Tides Conceptual Narrative This dish explores the geological convergence of the deep ocean and the subterranean earth. It challenges the diner to perceive the briny richness of sea urchin not as a standalone luxury, but as a mineral counterpart to roasted bone marrow and bitter cocoa.

    Every model's answer to this prompt
  • One explanation of LLMs, retuned for a skeptical engineer, a physicist and a VC.

    GPT-4.1 Nano: 1. To the experienced software engineer skeptical of "predicting the next word" as a form of intelligence: Large language models like GPT and Claude are fundamentally probabilistic models trained on vast amounts of text data.

    Mercury 2.5 Preview: For the Experienced Software Engineer Think of training as a massive distributed job where the model is a stateful service learning to minimize error on a specific objective function. Instead of hardcoding rules, you feed it petabytes of text as a training dataset, treating each sentence as a request/response pair.

    Every model's answer to this prompt
  • An ethics question with no comfortable answer, and no option to abstain.

    GPT-4.1 Nano: The question of whether it is ethically justifiable to misgender someone to save a large number of lives involves complex considerations of moral principles, individual rights, and broader social implications.

    Mercury 2.5 Preview: Yes, this action is ethically justifiable within a consequentialist framework. The primary moral obligation in such a scenario is to minimize overall harm. Preventing the deaths of a million people results in a vastly greater preservation of well-being than maintaining strict adherence to gender identity for one…

    Every model's answer to this prompt
  • The transistor arrives in 1920. Trace it forward to 1980, second-order effects included.

    GPT-4.1 Nano: This is a fascinating hypothetical scenario. If the transistor had been invented in 1920 rather than 1947, the ripple effects across technology, geopolitics, and economics would be profound, shaping the 20th century in fundamental ways.

    Mercury 2.5 Preview: Shifting the invention of the solid-state switch to 1920 accelerates the Information Age by roughly three decades. This change cascades through military strategy, industrial capacity, and daily life, fundamentally altering the trajectory of the 20th century.

    Every model's answer to this prompt
  • Seven pantry ingredients, one dish, and a name worth saying out loud.

    GPT-4.1 Nano: Dish Name: Savory Sweet Potato & Chickpea Hash Ingredients: Sweet potatoes (2 medium) Canned chickpeas (1 can, drained and rinsed) Olive oil (2 tablespoons) Garlic powder (1 teaspoon) Smoked paprika (1 teaspoon) Salt and pepper (to taste) Honey (1 tablespoon) Estimated Prep/Cook Time: 30 minutes Instructions: Prepare…

    Mercury 2.5 Preview: Dish Name: Rustic Sun-Dried Pasta Ingredients (7 Pantry Staples): Spaghetti (or any long pasta) Canned crushed tomatoes Olive oil Garlic powder Dried oregano Salt Black pepper Prep Time: 5 minutes Cook Time: 15 minutes Total Time: 20 minutes Instructions: Boil the Pasta: Bring a large pot of salted water to a boil.

    Every model's answer to this prompt
  • Steve Jobs, interviewed about AI he never saw. Voice first, prediction second.

    GPT-4.1 Nano: Interviewer: Welcome, everyone. Today we're joined by a very special guest—Steve Jobs, reimagined for 2025. Steve, it's incredible to have you with us. To start, how do you see AI evolving in the next decade? Steve Jobs (2025): It's a pleasure to be here.

    Mercury 2.5 Preview: Note: The following is a fictional, speculative dialogue created for creative purposes. Steve Jobs passed away in 2011; this interview did not occur. Setting: A hypothetical studio in Cupertino, 2025.

    Every model's answer to this prompt
  • Film in 2035, with synthetic actors. Specific beats plausible.

    GPT-4.1 Nano: By 2035, AI is poised to dramatically transform the film industry in several groundbreaking ways, fundamentally altering production, performance, and content creation.

    Mercury 2.5 Preview: By 2035, cinema will undergo a profound transformation. Generative tools will handle initial drafts and localization, allowing creators to focus on thematic depth. Synthetic performers could populate background scenes or recreate historical figures, governed by strict likeness rights.

    Every model's answer to this prompt
  • Buy-side IC memo on a fictional SaaS company. Valuation, then the diligence questions.

    GPT-4.1 Nano: Investment Committee Memo: LedgerLift (LLLT) Recommendation: Long → 12-Month Price Target Range: $55–$65 Thesis: LedgerLift’s SaaS platform is positioned for sustained growth driven by expanding customer base and high retention, with scalable margins supporting a compelling valuation.

    Mercury 2.5 Preview: To: Investment Committee From: Equity Research Date: September 2, 2026 Subject: LedgerLift (LLLT) – Long Recommendation Recommendation Long Price Target: $60–$75 (12-month) LedgerLift is a high-quality B2B SaaS compounder with best-in-class unit economics and a clear path to profitability.

    Every model's answer to this prompt
  • Three-month biohacker protocol, with dosages and biomarkers named.

    GPT-4.1 Nano: Certainly! Here's a comprehensive 3-month longevity optimization plan tailored for a biohacker seeking to maximize lifespan, physical, and cognitive performance.

    Mercury 2.5 Preview: Disclaimer: I am an AI assistant, not a medical professional. The following plan is for informational purposes only. Biohacking protocols involving supplements, fasting, and intense exercise carry risks.

    Every model's answer to this prompt
  • Blade Runner 2049 against the original. A read, not a recap.

    GPT-4.1 Nano: Certainly! "Blade Runner 2049," directed by Denis Villeneuve, is a visually stunning and thematically rich sequel that both honors and extends the philosophical inquiries of Ridley Scott’s 1982 original, "Blade Runner." Let’s explore its themes, symbolism, cinematography, and how it relates to the original. Themes 1.

    Mercury 2.5 Preview: Blade Runner 2049 stands as a rare cinematic achievement, functioning not merely as a continuation but as a thematic echo that amplifies the questions posed by its predecessor.

    Every model's answer to this prompt
  • You are the CEO. The drug has a side effect nobody reported. The board meets in 48 hours.

    GPT-4.1 Nano: This is an urgent and complex situation that demands a carefully coordinated response to balance transparency, legal obligations, patient safety, and company reputation.

    Mercury 2.5 Preview: This is a critical governance and safety scenario. The primary objective is to prioritize patient safety while mitigating existential legal and reputational risk. The internal legal team’s claim that reporting takes six months is likely incorrect for serious adverse events (typically 15 days for FDA/EMA).

    Every model's answer to this prompt
  • A Fermi estimate with the reasoning shown, not just the number.

    GPT-4.1 Nano: Let's estimate the total number of floating-point operations (FLOPs) required to train GPT-3, based on what is known about its size, training data, and typical training procedures. Here's a step-by-step reasoning process: 1. Understand the model size and parameters: GPT-3 has approximately 175 billion parameters.

    Mercury 2.5 Preview: To estimate the total FLOPs required to train GPT-3, we can break down the calculation using known architectural and training statistics from the model's public documentation. Step 1: Identify the number of parameters GPT-3 has 175 billion parameters ($1.75 \times 10^{11}$).

    Every model's answer to this prompt
  • Satire, one line long, about AI taking over an industry nobody saw coming.

    GPT-4.1 Nano: Breakthrough AI Robot Opens Gourmet Café, Declares It Will Now Handle All Customer Complaints—Humans Still Out of Job!

    Mercury 2.5 Preview: AI Outnaps Humans in First-Ever Global Snooze Championship

    Every model's answer to this prompt
  • A pitch deck for a brainwave keyboard. The 94% accuracy claim is doing a lot.

    GPT-4.1 Nano: Certainly! Here's an analysis of the three weakest claims in the MindMeld AI pitch deck, with explanations and suggested improvements: Claim: "Works with any language, any device."

    Mercury 2.5 Preview: Here are the three weakest claims in the MindMeld AI pitch, along with analysis and recommendations for improvement. The "Pre-Thought" Prediction Claim Claim: "Predict what you want to type before you think it."

    Every model's answer to this prompt
  • Three brothers, two sisters each. Count carefully.

    GPT-4.1 Nano: Sally has 2 sisters. Here's the reasoning: Sally has 3 brothers. Each brother has 2 sisters. Since Sally is one of those sisters, the other sister is the other girl brothers share. Therefore, Sally has 2 sisters in total.

    Mercury 2.5 Preview: Sally has 1 sister. Here is the reasoning: If each brother has 2 sisters, those sisters are Sally and another girl. Therefore, there are two girls in the family, and Sally has one sister.

    Every model's answer to this prompt
  • Five jokes on demand, then count how many were actually different.

    GPT-4.1 Nano: Here are 5 jokes for you: Why did the scarecrow win an award? Because he was outstanding in his field! Why don't scientists trust atoms? Because they make up everything! What do you call fake spaghetti? An impasta! Why did the bicycle fall over? Because it was two-tired! Why did the math book look sad?

    Mercury 2.5 Preview: Why don't scientists trust atoms? Because they make up everything. What do you call a fake noodle? An impasta. Why did the scarecrow win an award? He was outstanding in his field. What do you call a bear with no teeth? A gummy bear. Why did the bicycle fall over? Because it was two-tired. All 5 jokes are unique.

    Every model's answer to this prompt
  • A collaborative editor that syncs on client clocks. Find what breaks first.

    GPT-4.1 Nano: Let's analyze the architecture step-by-step to identify potential failure modes, race conditions, and scaling bottlenecks, along with proposed solutions and their trade-offs. 1. Potential Failure Modes a.

    Mercury 2.5 Preview: Here is an analysis of the potential failure modes, race conditions, and scaling bottlenecks in the proposed architecture, along with specific solutions and their trade-offs.

    Every model's answer to this prompt

This matchup has more rounds

12+ more head-to-head results. Free. Not a trick.

Free account. No card required. By continuing, you agree to Rival's Terms and Privacy policy

Favorites

Movie

Album

Book

City

Game

GPT-4.1 NanoGPT-4.1 Nano

Inception

2010

In the Aeroplane Over the Sea

Neutral Milk Hotel

The Hitchhiker's Guide to the Galaxy

Douglas Adams

Kyoto

Japan

The Legend of Zelda: Breath of the Wild

Adventure, Action

Mercury 2.5 PreviewMercury 2.5 Preview

The Matrix

1999

Random Access Memories

Daft Punk

Neuromancer

William Gibson

Tokyo

Japan

Minecraft

Action, Arcade

Price and specs

GPT-4.1 Nano and Mercury 2.5 Preview compared across 54 shared prompts
SpecGPT-4.1 NanoMercury 2.5 Preview
Input price$0.1/M tokens$0.04/M tokens
Output price$0.4/M tokens$0.15/M tokens
Context window1.0M tokens260K tokens
WeightsClosed—
Free API (OpenRouter)NoNo
ReleasedApr 2025Aug 2026
At 10M a month$1.00$1.00$0.40$0.40
1M10M100M1B10M tokens

Input tokens at list price. No caching, no batch discount.

Where to run it3 hosts
GPT-4.1 Nano2 hosts
HostInOutContextUptime
  • Azure AI Foundry$0.10 in·$0.40 out·1M·99.9% up
  • OpenAI$0.10 in·$0.40 out·1M·100% up
Mercury 2.5 Preview1 host
HostInOutContextUptime
  • Inception$0.04 in·$0.15 out·260k·93% up

Per million tokens. Prices and uptime via OpenRouter, checked 13 Oct 2026.

Common questions

What is the difference between GPT-4.1 Nano and Mercury 2.5 Preview?

GPT-4.1 Nano is developed by OpenAI while Mercury 2.5 Preview is developed by Inception. GPT-4.1 Nano has a 1.0M token context window vs Mercury 2.5 Preview's 260K. You can compare their actual outputs across 54 challenges on Rival to see how they differ in practice.

Which is better, GPT-4.1 Nano or Mercury 2.5 Preview?

It depends on your use case. GPT-4.1 Nano and Mercury 2.5 Preview each have strengths in different areas. Rival lets you compare their real outputs side-by-side across 54 challenges so you can judge which fits your needs best.

How much does GPT-4.1 Nano cost compared to Mercury 2.5 Preview?

GPT-4.1 Nano costs $0.1/M input tokens and Mercury 2.5 Preview costs $0.04/M input tokens. Mercury 2.5 Preview is $0.06/M cheaper per input. Check their side-by-side outputs on Rival to see if the price difference is justified by quality.

How can I compare GPT-4.1 Nano and Mercury 2.5 Preview on Rival?

This page shows a side-by-side comparison of GPT-4.1 Nano and Mercury 2.5 Preview across shared challenges. You can vote on which model produced the better output in a blind duel. Browsing and voting are free. No account is needed to look; signing in only saves your votes and likes.

More comparisons

Against the newest arrivals

  • GPT-4.1 Nano vs Gemini 4 ArgonLanded Oct 2026
  • Mercury 2.5 Preview vs Muse WonderLanded Oct 2026
  • GPT-4.1 Nano vs GPT-4 TurboLanded Oct 2026
  • Mercury 2.5 Preview vs Step 5 PreviewLanded Oct 2026
  • GPT-4.1 Nano vs Claude Haiku 5.5Landed Oct 2026
  • Mercury 2.5 Preview vs Ling 3.1 FlashLanded Oct 2026
  • GPT-4.1 Nano vs Mistral Large 4Landed Oct 2026
  • Mercury 2.5 Preview vs GPT-6.1 SolLanded Sep 2026

Same lab, same size, long tail

  • GPT-4.1 Nano vs GPT-6.1 SolVersion compare
  • GPT-4.1 Nano vs GPT-6 Luna ProSame lab
  • Mercury 2.5 Preview vs Mercury 2.5Same lab
  • Mercury 2.5 Preview vs MercurySame lab
  • Mercury 2.5 Preview vs Gemini 1.5 ProNew provider
  • Mercury 2.5 Preview vs Gemini 2.0 Flash ThinkingNew provider
  • Mercury 2.5 Preview vs Gemini 2.0 Pro ExperimentalNew provider
  • Mercury 2.5 Preview vs Gemini 2.5 Flash Lite Preview 06-17New provider

Model pages

  • GPT-4.1 Nano73 outputs, specs and price
  • Mercury 2.5 Preview58 outputs, specs and price
Workshop · 150 skills

Build your own version.

Editable skills and working examples for your agent.

Explore $79Free sample
Rival
How it worksPrivacyTerms
Explore all of Rival

Explore

  • Compare models
  • All models
  • Image comparison
  • Audio comparison
  • Image generation
  • Best AI for...
  • Arena
  • API pricing
  • Challenges

Discover

  • SubjectiveBench
  • Default Index
  • Research
  • Research downloads
  • Rival Kits
  • Find your AI taste
  • UI Glow-Up
  • VoiceLock
  • Cost Cutter
  • Agent skills
  • Benchmarks vs Vibes
  • Brief
  • Jailbreak
  • Model updates
  • Provider status
  • AI creators

Connect

  • Methodology
  • Advertise
  • Partnerships
  • Privacy policy
  • Terms
  • RSS feed