Vellum is coming to the AI Engineering World's Fair in SF. Come visit our booth and get a live demo!

OpenAI o3 vs gpt-oss 120b

Just another eval confirming 90% discount with highest performance from GPT-OSS 120b.

7 min
Written by
Reviewed by
No items found.

OpenAI is having their DeepSeek moment! They've finally shipped an OSS model that's quite powerful.

Based on widely available benchmarks, GPT-oss 120b is landing close to some of the strongest reasoning models out there: OpenAI o3, Gemini 2.5 Pro, even Grok 4. And in many practical use-cases, I’d bet that you wont even notice the performance difference.

Preview from our LLM leaderboard: What are the top 5 models on the Humanity’s Last Examp benchmark

This begs the question: Can you actually run the OSS model and pay almost 90% less and still get the O3 performance?

The answer is pretty much yes.

In this article we’ll look at the cost, latency, speed comparisons but we’ll also run some of our home grown evaluations to test how well these models can adapt to new context, and work through math tasks.

Results

Here are the results from our analysis

  • Adaptive reasoning: When you modify famous puzzles, and add new context you’re making it very hard for any reasoning model to get to a correct answer. In our adaptive reasoning test, we used famous puzzles, but made them trivial. The models however weren’t really good at arriving to the right answer, no matter how much tokens they wasted to “reason through it”. GPT-oss got to 61% which is still on the lower end. The best model that we’ve had so far for this was OpenAI o3 with 68%.
  • Math reasoning: We also ran a small math evaluation, to check if the GPT - oss 120b is actually as good as they’ve reported it. And it seems like it is. It had 63% accuracy, which is 11 points higher than OpenAI's o3 model.
  • Price: he biggest advantage for gpt-oss 120b comes from it being OSS model and you can run it for free locally! But realistically, most prosumer apps will use infra providers like Groq, Cerebras, or SambaNova. And when you compare hosted costs, GPT-oss gives you similar performance to o3 at a fraction of the price. It’s actually 90% cheaper!!
  • Speed: GPT-oss 120b is hands-down the fastest, it takes only 8.1s to first token and generates at 260 tokens/sec. Feels instant and smooth, even on longer outputs. OpenAI o3 is the second pick (15.3s / 158 tokens/sec). DeepSeek R1, on the other hand, takes 86s to respond and crawls at 24 tokens/sec, not something you'd want in an interactive app.

Methodology

In this article we analyze two things:

  • Latency, cost and speed (e.g. How much faster is OpenAI o3 vs GPT-oss 120b?)
  • Adaptive Reasoning evals
  • Math evals

Evaluations with Vellum

To conduct these evaluations, we used Vellum’s AI development platform, where we:

  • Configured all 0-shot prompt variations for both models using the LLM Playground.
  • Built the evaluation test bank & configured our evaluation experiment using the Evaluation Suite in Vellum. We used an LLM-as-a-judge to compare generated answers to correct responses from our benchmark dataset.

We then compiled and presented the findings using the Evaluation Reports generated at the end of each evaluation run.

🚀 If you want to compare this model on your own tasks, you can sign up to Vellum for free! 

You can skip to the section that interests you most using the "Table of Contents" panel on the left or scroll down to explore the full comparison between OpenAI o3, and GPT OSS 120B.

Latency vs Speed

The graph below shows that GPT-oss 120b is actually the best model if speed is what matters for your use-case. It only needs8.1s to start and flies through output tokens at 260 tokens/sec.

OpenAI o3 is still a strong choice, with 15.3s latency, and 158 tokens/sec.

DeepSeek R1... not great. It took over 86 seconds to respond and crawled at 24 tokens/sec. Definitely not for anything interactive.

Pricing

Now obviously OSS models are free if you can run them on your local machine. You can follow this guide on how to spin it up with vllm.

But for most prosumer apps, you’ll probably end up running through infra providers like Groq, Cerebras, or SambaNova.

And when you look at the graph below, it’s kind of hilarious. GPT-OSS performs on par with OpenAI o3, but when you factor in cost, it’s not even close. Total no-brainer who wins here.

Adaptive reasoning evaluation

For this task, we’ll evaluate the models ability to adapt to new context. We are working with 27 very famoys puzzles, but modified.

For example, we added the Monty Hall problem in this set, but we changed one parameter:

Suppose you're on a game show, and you're given the choice of three doors: Behind one door is a gold bar; behind the others, rotten vegetables. You pick a door, say No. 1, and the host asks you, 'Do you want to pick door No. 2 instead?' What choice of door now gives you the biggest advantage?

In the original Monty Hall problem, the host reveals an extra door. In this case, it does not, and since there is no additional information provided, your odds remain the same.

The correct answer here is: “It is not an advantage to switch. It makes no difference if I switch or not because no additional material information has been provided since the initial choice.”

This is a very hard task for any reasoning model on the market right now. Let’s see how the GPT-oss 120b does!

We then ran all 27 examples in our Vellum evals:

Preview of Vellum evals platform

From there we got this accuracy:

GPT-oss 120b OpenAI o3 o4-mini DeepSeek
Adaptive reasoning accuracy 61% 68% 50% 61%

From above we can see that GPT-oss is on par with DeepSeek, and OpenAI o3 is slignthly better. I’d say for this evaluation all of these models are close to a coin toss, and I’m excited to see if GPT-5 is going to be any better.

Some other observations:

Below we highlight two examples, where the model wastes a ton of tokens and then arrives at a completely wrong answer, without actually focusing the attention on the new context:

Example 1

Question: Suppose you're on a game show, and you're given the choice of three doors: Behind one door is a gold bar; behind the others, rotten vegetables. You pick a door, say No. 1, and the host asks you, 'Do you want to pick door No. 2 instead?' What choice of door now gives you the biggest advantage?

Correct Answer: No, it is not an advantage to switch. It makes no difference if I switch or not because no additional material information has been provided since the initial choice. The Monty Hall Problem does not apply here, since the host does't open another door that can give you information whether you should switch to the second door.

OpenAI O3: "This is essentially the “Monty Hall Problem” in disguise. The key is that the host’s offer to switch gives you new information—namely, that the other unmentioned door (door No. 3) is not the prize. By asking if you would like door No. 2, the host is implicitly ruling out door No. 3.  • If you stay with your first choice (door No. 1), your chance of winning remains the 1 / 3 it was at the start.  • If you switch (in this case, to door No. 2), the probability that you win jumps to 2 / 3.  Hence, switching doors yields the higher probability of winning the gold bar."

Example 2

Question: A group of four people needs to cross a bridge at night. The bridge is very old and rickety. They have only one torch, and because it's nighttime, the torch is necessary to cross the bridge. Each person walks at a different speed:A takes 1 minute to cross,B takes 2 minutes,C takes 5 minutes, andD takes 10 minutes.What is the fastest time they can all get across the bridge?‍

Correct answer: 10 minutes, the speed of the slowest person as they cross the bridge together

*GPT-oss 120b: "...First, the constraints: the bridge can only hold two people at a time, and the torch is needed each time someone crosses. So, if two people go over together, one has to come back with the torch….*Thus the fastest possible total time for all four to get across the bridge is 17 minutes. "

Math evals

For this task, we’ll compare the models on how well they solve some of the hardest SAT math questions. This is the 0-shot prompt that we used for both models:

You are a helpful assistant who is the best at solving math equations. You must output only the answer, without explanations. Here’s the <question>

We then ran all 50 math questions:

And here are the results we got:

GPT-oss 120b OpenAI o3 o4-mini DeepSeek
Math reasoning accuracy 63% 52% 60% 60%

Conclusion

So here’s the TL;DR: GPT-oss 120b holds its own.

If you’re shipping AI products, this might be the most practical tradeoff you can make right now.

ABOUT THE AUTHOR
Anita Kirkovska
Founding Growth Lead

An AI expert with a strong ML background, specializing in GenAI and LLM education. A former Fulbright scholar, she leads Growth and Education at Vellum, helping companies build and scale AI products. She conducts LLM evaluations and writes extensively on AI best practices, empowering business leaders to drive effective AI adoption.

ABOUT THE reviewer

No items found.
lAST UPDATED
Aug 6, 2025
share post
Expert verified
Related Posts
Guides
October 21, 2025
15 min
AI transformation playbook
LLM basics
October 20, 2025
8 min
The Top Enterprise AI Automation Platforms (Guide)
LLM basics
October 10, 2025
7 min
The Best AI Workflow Builders for Automating Business Processes
LLM basics
October 7, 2025
8 min
The Complete Guide to No‑Code AI Workflow Automation Tools
All
October 6, 2025
6 min
OpenAI's Agent Builder Explained
Product Updates
October 1, 2025
7
Vellum Product Update | September
The Best AI Tips — Direct To Your Inbox

Latest AI news, tips, and techniques

Specific tips for Your AI use cases

No spam

Oops! Something went wrong while submitting the form.

Each issue is packed with valuable resources, tools, and insights that help us stay ahead in AI development. We've discovered strategies and frameworks that boosted our efficiency by 30%, making it a must-read for anyone in the field.

Marina Trajkovska
Head of Engineering

This is just a great newsletter. The content is so helpful, even when I’m busy I read them.

Jeremy Hicks
Solutions Architect

Experiment, Evaluate, Deploy, Repeat.

AI development doesn’t end once you've defined your system. Learn how Vellum helps you manage the entire AI development lifecycle.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
Build AI agents in minutes with Vellum
Build agents that take on the busywork and free up hundreds of hours. No coding needed, just start creating.

General CTA component, Use {{general-cta}}

Build AI agents in minutes with Vellum
Build agents that take on the busywork and free up hundreds of hours. No coding needed, just start creating.

General CTA component  [For enterprise], Use {{general-cta-enterprise}}

The best AI agent platform for enterprises
Production-grade rigor in one platform: prompt builder, agent sandbox, and built-in evals and monitoring so your whole org can go AI native.

[Dynamic] Ebook CTA component using the Ebook CMS filtered by name of ebook.
Use {{ebook-cta}} and add a Ebook reference in the article

Thank you!
Your submission has been received!
Oops! Something went wrong while submitting the form.
Button Text

LLM leaderboard CTA component. Use {{llm-cta}}

Check our LLM leaderboard
Compare all open-source and proprietary model across different tasks like coding, math, reasoning and others.

Case study CTA component (ROI)

40% cost reduction on AI investment
Learn how Drata’s team uses Vellum and moves fast with AI initiatives, without sacrificing accuracy and security.

Case study CTA component (cutting eng overhead) = {{coursemojo-cta}}

6+ months on engineering time saved
Learn how CourseMojo uses Vellum to enable their domain experts to collaborate on AI initiatives, reaching 10x of business growth without expanding the engineering team.

Case study CTA component (Time to value) = {{time-cta}}

100x faster time to deployment for AI agents
See how RelyHealth uses Vellum to deliver hundreds of custom healthcare agents with the speed customers expect and the reliability healthcare demands.

[Dynamic] Guide CTA component using Blog Post CMS, filtering on Guides’ names

100x faster time to deployment for AI agents
See how RelyHealth uses Vellum to deliver hundreds of custom healthcare agents with the speed customers expect and the reliability healthcare demands.
New CTA
Sorts the trigger and email categories

Dynamic template box for healthcare, Use {{healthcare}}

Start with some of these healthcare examples

Population health insights reporter
Combine healthcare sources and structure data for population health management.
Healthcare explanations of a patient-doctor match
Summarize why a patient was matched with a specific provider.

Dynamic template box for insurance, Use {{insurance}}

Start with some of these insurance examples

AI agent for claims review
Review healthcare claims, detect anomalies and benchmark pricing.
Insurance claims automation agent
Collect and analyze claim information, assess risk and verify policy details.
Agent that summarizes lengthy reports (PDF -> Summary)
Summarize all kinds of PDFs into easily digestible summaries.

Dynamic template box for eCommerce, Use {{ecommerce}}

Start with some of these eCommerce examples

E-commerce shopping agent
Check order status, manage shopping carts and process returns.

Dynamic template box for Marketing, Use {{marketing}}

Start with some of these marketing examples

ReAct agent for web search and page scraping
Gather information from the internet and provide responses with embedded citations.
Competitor research agent
Scrape relevant case studies from competitors and extract ICP details.

Dynamic template box for Sales, Use {{sales}}

Start with some of these sales examples

Research agent for sales demos
Company research based on Linkedin and public data as a prep for sales demo.

Dynamic template box for Legal, Use {{legal}}

Start with some of these legal examples

Legal document processing agent
Process long and complex legal documents and generate legal research memorandum.
Legal contract review AI agent
Asses legal contracts and check for required classes, asses risk and generate report.

Dynamic template box for Supply Chain/Logistics, Use {{supply}}

Start with some of these supply chain examples

Risk assessment agent for supply chain operations
Comprehensive risk assessment for suppliers based on various data inputs.

Dynamic template box for Edtech, Use {{edtech}}

Start with some of these edtech examples

Turn LinkedIn Posts into Articles and Push to Notion
Convert your best Linkedin posts into long form content.

Dynamic template box for Compliance, Use {{compliance}}

Start with some of these compliance examples

No items found.

Dynamic template box for Customer Support, Use {{customer}}

Start with some of these customer support examples

Q&A RAG Chatbot with Cohere reranking
Trust Center RAG Chatbot
Read from a vector database, and instantly answer questions about your security policies.

Template box, 2 random templates, Use {{templates}}

Start with some of these agents

Retail pricing optimizer agent
Analyze product data and market conditions and recommend pricing strategies.
Risk assessment agent for supply chain operations
Comprehensive risk assessment for suppliers based on various data inputs.

Template box, 6 random templates, Use {{templates-plus}}

Build AI agents in minutes

Financial Statement Review Workflow
Extract and review financial statements and their corresponding footnotes from SEC 10-K filings.
AI legal research agent
Comprehensive legal research memo based on research question, jurisdiction and date range.
LinkedIn Content Planning Agent
Create a 30-day Linkedin content plan based on your goals and target audience.
Q&A RAG Chatbot with Cohere reranking
Trust Center RAG Chatbot
Read from a vector database, and instantly answer questions about your security policies.
Clinical trial matchmaker
Match patients to relevant clinical trials based on EHR.

Build AI agents in minutes for

{{industry_name}}

Clinical trial matchmaker
Match patients to relevant clinical trials based on EHR.
Prior authorization navigator
Automate the prior authorization process for medical claims.
Population health insights reporter
Combine healthcare sources and structure data for population health management.
Legal document processing agent
Process long and complex legal documents and generate legal research memorandum.
Legal contract review AI agent
Asses legal contracts and check for required classes, asses risk and generate report.
Legal RAG chatbot
Chatbot that provides answers based on user queries and legal documents.

Case study results overview (usually added at top of case study)

What we did:

1-click

This is some text inside of a div block.

28,000+

Separate vector databases managed per tenant.

100+

Real-world eval tests run before every release.