OpenAI O3 Mini vs DeepSeek for Agentic Use

As of 31st January, OpenAI launched their next generation of reasoning models with O3 Mini Model. Overall, it seems OpenAI O3 Mini model is 15 times cheaper with approximately the same accuracy as O1 Model, OpenAI’s previous generation of reasoning Models. The new Model, OpenAI O3 Mini model is available now in FlowHunt less than a day passed from OpenAI’s launch. In this blog, we’ll deep dive into the model capabilities and compare it with Deepseek.

Evaluation Metrics for OpenAI O3 Mini vs DeepSeek

In this evaluation, I utilized a straightforward yet effective workflow involving chat input, chat output, and an AI generator to deliver a clear and concise explanation of the main line of the Ruy Lopez opening in chess. By comparing the performance of DeepSeek against O3 Mini, I was able to assess how each system handled the task of outlining key moves—such as 1. e4 e5, 2. Nf3 Nc6, 3. Bb5 a6, 4. Ba4 Nf6, and 5. O-O Be7—and their subsequent developments. This comparative analysis not only demonstrates the capabilities of these AI tools in processing and relaying complex chess strategies, but also highlights the value of a simple conversational flow in achieving readability and practical evaluation outcomes.

This is the flow:

DeepSeek Results

I have asked the model the following:
“Give me first 10 moves of Ruy Lopez.”
The results is as follows. This is the thinking process of DeepSeek:

o3 mini vs deepseek Restarting session Please wait give me first 10 moves of ruy lopez
Okay, so I’m trying to figure out the first 10 moves of the Ruy Lopez opening in chess. I remember that the Ruy Lopez is one of the oldest and most popular openings, starting with the moves 1.e4 e5. Then, White plays 2.Nf3, developing the knight and preparing to control the center. Next, 2…Nc6, where Black develops their knight and supports the e5 pawn. Now, the Ruy Lopez proper starts with 3.Bb5, which attacks the knight on c6. I think this is the key move that defines the opening. Then, Black usually responds with 3…a6 to challenge the bishop. After that, White plays 4.Ba4, retreating the bishop to a safe square. So far, I have six moves: 1.e4 e5 2.Nf3 Nc6 3.Bb5 a6 4.Ba4. Now, Black’s next move is often 4…Nf6, developing the knight and attacking the pawn on e4. Then, White castles kingside with 5.0-0 to get the king to safety and bring the rook into the game. That’s eight moves. Next, Black might play 5…b5 to challenge White’s bishop on a4 and gain space on the queenside. Then, White responds with 6.c3, supporting the d4 push and preparing to develop other pieces. So, the tenth move would be 6…d6, solidifying the e5 pawn and preparing to develop the dark-squared bishop. Wait, but I’m not entirely sure if Black always plays 5…b5. Sometimes, they might play 5…Be7 instead, but I think 5…b5 is more common in the Closed Ruy Lopez. Also, after 6.c3, Black can choose between 6…d6 or 6…b5 again, but 6…d6 seems more standard to support e5. I should double-check the order to make sure I didn’t mix up any moves. Let me go through it again: e4 e5 Nf3 Nc6 Bb5 a6 Ba4 Nf6 0-0 b5 c3 d6 Yes, that seems correct. So the first ten moves are as I listed. I think that’s a solid foundation, but I’m a bit unsure if there are alternative lines Black might take earlier, like after 3.Bb5, maybe 3…g6 instead of 3…a6, but I think 3…a6 is the main line leading to the Morphy Defense. So, I’ll stick with that for the first ten moves. Foundation for both White and Black, with White aiming for a strong center and piece development, while Black challenges White’s bishop and prepares for further development.

Final result:

The first ten moves of the Ruy Lopez opening, following the Morphy Defense, are as follows:
e4 e5 Nf3 Nc6 Bb5 a6 Ba4 Nf6 0-0 b5 c3 d6
This sequence establishes a solid foundation for both White and Black, with White aiming for a strong center and piece development, while Black challenges White’s bishop and prepares for further development.

OpenAI O3 Mini Results

O3 wins. O3 Mini’s answer is more understandable and correct.

The Ruy Lopez (or Spanish Game) is a rich opening with many variations. One of the most common ways to reach a classical Closed Ruy Lopez is with the following moves:
e4 e5 Nf3 Nc6 Bb5 a6 Ba4 Nf6 O-O Be7 Re1 b5 Bb3 d6 c3 O-O h3 Nb8 d4

Here’s a brief explanation of the ideas behind these moves:

• 1. e4 e5 sets up the classical battle for the center.
• 2. Nf3 develops a knight and attacks the e5 pawn; Black defends with Nc6.
• 3. Bb5 pins the knight on c6 (which is defending e5) and increases central pressure; Black challenges with a6.
• 4. Ba4 retreats while still keeping the pin; Black develops another piece with Nf6.
• 5. O-O secures White’s king; Black plays Be7 to prepare castling.
• 6. Re1 reinforces the central e4 pawn; Black’s b5 kicks the bishop and gains space on the queenside.
• 7. Bb3 preserves the bishop’s diagonal; Black shores up the center with d6.
• 8. c3 prepares to support a later d4 push while also preparing for a possible retreat of the bishop if needed, and Black castles with O-O.
• 9. h3 prevents any piece (like a knight or bishop) from coming to g4, where it could harass White’s center or pieces; Black retreats the knight with Nb8 to re-route it.
• 10. d4 is White’s break in the center, intending to challenge Black’s pawn structure.

Keep in mind this is just one of many valid move orders in the Ruy Lopez. Variations such as the Open Ruy Lopez or other lines in the Closed Ruy Lopez might differ after move 4 or later. But the sequence above is a well‐known path that many players follow in classical chess theory.

Evaluating DeepSeek and OpenAI O3 Mini on Agentic Behaviour

We have evaluated the reasoning models on a task needing 2-3 tools for AI Agents. Here is the task:

Given the company name, search and find the pricing page of product and suggest enhancements.
Also take screenshot of pricing page.
Then send an email to me with the results.

The task requires connecting to Google search to fetch the pricing page, crawling the pricing page and sending a detailed report to email.

The task was intentionally not described in details to check the reasoning capabilities of the mentioned models.

Using DeepSeek as AI Agent

Deepseek was not able to send email. Overall, our experience reveals that Deepseek is not as accurate in tool calling comparing to OpenAI GPT4o or OpenAI O3 Mini Models. This is the report generated from DeepSeek, analyzing pricing page of FlowHunt. Along the way this is the thought process of DeepSeek:

1. Call to Google Tool:

2. Scrape URL:

The reasoning:

Subject: FlowHunt Pricing Page Review and Suggestions

Dear [Your Name],

I hope this message finds you well. Below is a summary of the FlowHunt pricing page, along with some suggestions for enhancements.

Pricing Overview:

FlowHunt offers four pricing tiers:

Starter Plan (€10/month):

• 10 Credits
• 5 Workflows per Workspace
• 1 Workspace
• 1 Team Member per Workspace
• 5 Chatbots per Workspace
• 2 Websites per Workspace
• 5 Documents per Workspace
• 50 FAQs per Workspace
• Access to All Latest OpenAI LLMs, Including Custom Fine-Tuned Models
• All Integrations
• Email Support

Pro Plan (€50/month):

• 55 Credits (Includes 10% Discount on Credits)
• 20 Workflows per Workspace
• 5 Workspaces
• 10 Team Members per Workspace
• 20 Chatbots per Workspace
• 10 Websites per Workspace
• 50 Documents per Workspace
• 100 FAQs per Workspace
• Access to All Latest OpenAI LLMs, Including Custom Fine-Tuned Models
• Pro Support
• API Key Access
• Basic Analytics (Coming Soon)
• All Integrations

Premium Plan (€500/month):

• 600 Credits (Includes 20% Discount on Credits)
• 50 Workflows per Workspace
• 10 Workspaces
• 100 Team Members per Workspace
• 50 Chatbots per Workspace
• 30 Websites per Workspace
• 100 Documents per Workspace
• 200 FAQs per Workspace
• Access to All Latest OpenAI LLMs, Including Custom Fine-Tuned Models
• Priority Support
• API Key Access
• Advanced Analytics (Coming Soon)
• Initial Workflow Consultation
• Initial Workflow Training Support
• All Integrations

Enterprise Plan (Custom Pricing starting from €1500/month):

• Contracted Tokens (Bring Your Own Key – BYOK)
• Unlimited Workflows
• Unlimited Workspaces
• Unlimited Users
• Unlimited Chatbots
• Unlimited Schedules
• Unlimited Documents
• Unlimited Questions and Answers
• Unlimited Chat History
• Access to All Latest OpenAI LLMs, Including Custom Fine-Tuned Models
• Dedicated Priority Support
• API Key Access
• Advanced Analytics (Coming Soon)
• Customizable Workflow Modules
• Optional Training Support
• Advanced Security and on-premises Deployment

Suggested Enhancements:

• Feature Comparison Table: Create a side-by-side comparison table highlighting the features of each plan. This would make it easier for users to quickly identify the differences between the plans.
• Customer Testimonials/Case Studies: Adding testimonials or case studies from existing customers can build trust and credibility.
• Pricing Calculator: Implement an interactive pricing calculator that allows users to estimate their costs based on their specific needs, especially for the Enterprise plan.
• Free Trial Highlight: Make the free trial option more prominent on the page to encourage users to try the service before committing to a paid plan.
• Detailed Documentation/Setup Guides: Provide links to detailed documentation or setup guides to help users understand how to make the most of the features included in each plan.

Using OpenAI O3 Mini as AI Agent

On the other hand, with this less amount of description OpenAI O3 Mini model was able to decode the exact process:

1. Google Search

2. Scrape URL

3. Email Sending

This is the email I received:

Hello,

Below are the results of the review of FlowHunt’s pricing page and some suggested enhancements:

Pricing Page URL: https://www.flowhunt.io/pricing/

Screenshot: A screenshot of the pricing page was taken. (Please refer to the attached screenshot in your system.)

Summary of the Pricing Page:

The pricing page clearly displays four plans: Starter (€10/month), Pro (€50/month), Premium (€500/month), and Enterprise (from €1500/month). Each plan lists the credits, workflow and team limits, core features, and additional support options. The page includes both an overview and a detailed comparison using a table format, which helps potential customers compare the features across plans.

Suggested Enhancements:

• Interactive Pricing Calculator: Consider adding an interactive pricing calculator that allows users to adjust parameters (e.g., number of workflows, team members) to see a dynamic estimate of their monthly cost.
• Visual Enhancements and UX Improvements: Introduce more engaging visuals (e.g., icons or infographics) to further differentiate the features of each plan. Add subtle animations when hovering over different plan details to improve user interactivity.
• Currency Converter: Since pricing is in Euros, include a currency converter option to cater to international users for a better localized experience.
• Clearer Call-to-Action (CTA): Enhance the CTA buttons by making them more prominent and possibly adding a secondary CTA (e.g., “Learn More” about plan benefits) for users who need more clarity before signing up.
• Enhanced FAQ Section: Expand the FAQ section with more detailed answers regarding the credits system, potential overages, and common customer scenarios. Consider using collapsible sections to keep the page clean while providing additional information when needed.
• Social Proof and Testimonials: Incorporate customer testimonials or success stories near the pricing section to build credibility.
• Mobile Optimization: Ensure that the pricing table and all interactive elements render optimally on mobile devices.
• Chat Support Integration: Add an integrated live chat or chatbot support option on the pricing page to instantly address any queries related to pricing or plan details.

By implementing these enhancements, FlowHunt can provide a more engaging and user-friendly experience, potentially increasing conversion rates.

Best regards,
[Your Name]

(P.S. This email includes the results and suggestions for FlowHunt’s pricing page as well as information about the captured screenshot.)