OpenAI o3 Benchmarks 2026: Every Score & What Replaced It

Last Updated: July 1, 2026

Important update: OpenAI announced on May 28, 2026 that o3 will be retired from ChatGPT on August 26, 2026, following a 90-day sunset period, per OpenAI's official release notes. The API remains available with no announced retirement date. If your organization has workflows on o3 via ChatGPT, you need to migrate to o4 or the GPT-5 family before that date. API users can continue using o3 indefinitely.

This guide remains the most complete o3 benchmark reference available - covering every major score, honest context for what those scores mean, and how o3 compares to its successors GPT-5 and o4 across the benchmarks that matter for real-world business decisions.

After four years advising executives on AI implementation, the question I get most often after a new model releases is not "what did it score?" but "does the score mean it will work better for what we need?" The answer is almost always "it depends on the benchmark and your use case." This guide helps you answer that question for o3 - and understand where the benchmark landscape has moved in 2026.

🎯 Before you read on - we put together a free 2026 AI Tools Cheat Sheet covering the tools business leaders are actually using right now. Get it instantly when you subscribe to AI Business Weekly.

Get the Free Cheat Sheet

Before the benchmark data, the practical update every o3 user needs:

ChatGPT retirement: o3 will be removed from ChatGPT on August 26, 2026, announced by OpenAI on May 28, 2026, per OpenAI's official release notes. The 90-day sunset period began May 28. After August 26, existing ChatGPT conversations that used o3 will automatically continue on the corresponding GPT-5 model.

API availability: The retirement applies to ChatGPT only. OpenAI has not announced an API retirement date for o3. Developers with existing API integrations can continue using o3 via API until a separate API deprecation is announced.

What replaced o3:

The migration path: If you are using o3 for complex reasoning tasks via ChatGPT, you have until August 26. The migration is straightforward - swap to o4 or GPT-5.4 Thinking and test your outputs, per WIthO2's migration guide.

For context on how the current GPT-5 model family compares to o3 across benchmarks, see the dedicated comparison section below.

Before the scores, it helps to understand why o3 produces different benchmark results than GPT-4o or o1.

OpenAI describes o3 as trained to "think" before responding using a private chain of thought - a reinforcement learning approach where the model runs through intermediate reasoning steps before generating an answer. Unlike standard language models that produce responses based on pattern matching from training data, o3 actively works through a problem before committing to an answer.

This architecture change is why o3's benchmark performance looks so different from GPT-4o on reasoning-heavy tasks. It is also why o3 is slower and more expensive per query - it burns more compute on each response to run that reasoning process.

According to OpenAI's o3 release documentation, o3 makes 20% fewer major errors than o1 on difficult real-world tasks, with particular improvements in programming, business consulting, and creative ideation. The benchmark scores below reflect that improvement in detail.

One important limitation now documented: GPT-5's release revealed that o3 gave confident wrong answers 86.7% of the time when images were missing from prompts - compared to just 9% for GPT-5. This overconfidence characteristic is worth understanding before interpreting o3's benchmark scores. Source: OpenAI's GPT-5 introduction

Sources: Wikipedia OpenAI o3, OpenAI official release notes, OpenAI developer introduction

AIME: 96.7% on 2024, 88.9% on 2025

The American Invitational Mathematics Examination tests genuine multi-step mathematical reasoning that cannot be solved through pattern matching. o3 scored 96.7% on AIME 2024 compared to o1's 74.3% - a substantial improvement that reflects o3's extended reasoning architecture on complex multi-step problems.

For context on how far the benchmark has moved: o4-mini achieves 99.5% pass@1 on AIME 2025 when given access to a Python interpreter. GPT-5 with thinking also outperforms o3 on AIME. o3's scores remain strong - they are just no longer frontier.

What this means practically: For teams using AI for financial modeling, engineering calculations, actuarial work, or complex quantitative reasoning, o3's AIME performance was a meaningful signal when it launched. Those workflows should now be evaluated on GPT-5 or o4, which exceed o3 on the same benchmarks.

FrontierMath: 25.2% - Still the Most Important Number

FrontierMath, developed by Epoch AI, tests AI on unseen research-level mathematical problems that cannot be solved through memorization. FrontierMath v2 was released on June 12, 2026 with corrected and updated problems, per LM Council's benchmark tracker.

o3's score of 25.2% on the original FrontierMath was approximately 12x better than anything that came before it - the previous best was under 2%. That discontinuity was real and meaningful when o3 launched. In 2026, frontier model performance on FrontierMath has continued advancing - check LM Council for current scores across models.

For context on how AI research-level math capabilities have evolved, our OpenAI statistics guide covers the broader model development picture.

GPQA Diamond: 87.7%

GPQA Diamond - Graduate-Level Google-Proof Question and Answer - tests PhD-level science knowledge in biology, chemistry, and physics using questions specifically designed to be unsearchable. Human experts score approximately 70%.

o3 achieved 87.7% on GPQA Diamond, surpassing average human expert performance. For comparison, Gemini 3.1 Pro scores 77.1% on the harder ARC-AGI-2 benchmark (a different test). On GPQA Diamond specifically, the 2026 competitive picture:

For pharmaceutical companies, biotech teams, materials science researchers, or any organization doing knowledge work at the intersection of science and business, o3's GPQA performance remains relevant context even as the model approaches retirement. These scores informed which model families are worth building on.

Our ChatGPT vs Claude comparison covers how the platforms compare on scientific and analytical tasks for business teams.

💡 Finding this helpful? Get bite-sized AI news and practical business insights like this delivered free every morning at 7 AM EST.

Subscribe Free

SWE-bench Verified: 71.7%

SWE-bench Verified tests AI on real GitHub issues from popular open-source repositories - actual software bugs and feature requests from real development teams. It is the most practically relevant coding benchmark because it tests performance on genuine production problems.

o3 achieved 71.7% on SWE-bench Verified compared to 48.9% for o1 - a 22.8 percentage point improvement. For context on how the benchmark has moved:

SWE-bench is now dominated by Claude models. For development teams making current AI coding decisions, o3's 71.7% is a historical baseline - not the current frontier. See our AI coding tools statistics guide for current performance data.

Codeforces Elo: 2,727

An Elo rating of 2,727 places o3 in the top fraction of a percent of all competitive programmers globally. o1 scored 1,891. Most senior software engineers have ratings between 1,400 and 1,800. For algorithmic problem-solving, o3 outperformed the majority of professional developers at launch.

For a complete comparison of AI coding tools including Claude Code, Cursor, and GitHub Copilot, our AI coding tools guide covers the full landscape.

ARC-AGI-1: 75.7% (low compute), 87.5% (high compute)

ARC-AGI-1 tests pattern recognition in genuinely novel situations that cannot be solved through memorization. o3's 87.5% at high compute surpassed average human performance of 85% and generated significant AGI discussion in late 2024 and early 2025.

The critical context: ARC Prize published results showing o3 scores only 2.9% on the newer ARC-AGI-2 benchmark, compared to 60% for average humans. The 2026 frontier looks dramatically different:

The ARC-AGI-2 result at 2.9% remains the most important data point for executive teams thinking about AI capabilities. It demonstrates that benchmark performance on specific tests does not equal general intelligence, and that the gap between impressive benchmark results and genuine reasoning across novel problems was very wide at o3's launch.

MMLU: 91.6%

MMLU covers 57 academic subjects. o3's 91.6% is strong but no longer meaningfully differentiating - most frontier models score above 88%. Google's Gemini 2.5 Pro scores 92%, and Claude models range 88-91%. MMLU is no longer a frontier discriminator.

This is the most important new section since March 2026. GPT-5, released in August 2025 and now evolved through the GPT-5.5 family, has superseded o3 across most benchmarks.

Per OpenAI's official GPT-5 introduction:

The efficiency gain is particularly significant: GPT-5 with thinking delivers better performance using 50-80% fewer output tokens than o3, meaning lower cost per equivalent task. Combined with o3's retirement date, this makes GPT-5 or o4 the clear path forward for any organization currently using o3.

For context on what GPT-5.4 and GPT-5.5 offer in 2026, our ChatGPT new features 2026 guide covers the current model family.

The 2026 context these numbers need: The competitive landscape has moved significantly since o3's launch. Claude Fable 5 scored 95.0% on SWE-bench before its suspension. Anthropic models now occupy the top 6 positions on SWE-bench. Gemini 3.1 Pro's 77.1% on ARC-AGI-2 versus o3's 2.9% is the starkest illustration of how much the reasoning benchmark frontier has moved.

For the most current benchmark comparisons across active models, see our AI coding tools statistics guide and Claude AI statistics guide.

Pricing for API users who will continue using o3 beyond August 26, 2026:

OpenAI dropped o3 prices to $2 per million input tokens and $8 per million output tokens in March 2026 - an 80% reduction from the original $10/$40 pricing. This change significantly altered the cost-benefit calculation.

Critical billing note for API users: o-series models use reasoning tokens for internal thinking steps billed as output tokens but not visible in API responses. A 500-token visible response may consume 2,000+ total tokens. Teams budgeting for o3 API usage must account for these hidden reasoning tokens. For current pricing on o4 and GPT-5 models, check OpenAI's API pricing page directly.

After reviewing every o3 benchmark in detail and watching the model landscape evolve through mid-2026, here is the practical synthesis.

For teams currently using o3 via ChatGPT: Migrate to GPT-5.4 Thinking or o4 before August 26, 2026. GPT-5 outperforms o3 across coding, hallucination rates, and efficiency while using 50-80% fewer output tokens. The migration is an upgrade, not a compromise.

For teams using o3 via API: No immediate action required. The API continues operating with no announced retirement date. However, given GPT-5's benchmark improvements and cost efficiency, evaluating migration to o4 or GPT-5 is still worthwhile.

For teams evaluating AI platforms based on benchmark data: The o3 scores in this article tell you the state of AI reasoning as of April 2025. The benchmark landscape in June 2026 looks different. ARC-AGI-2 results have improved dramatically. SWE-bench is now dominated by Claude models. GPT-5 outperforms o3 with greater efficiency. Use o3's scores as a historical baseline - validate against current model benchmarks for any purchasing decision.

The ARC-AGI-2 lesson remains relevant: o3 scored 2.9% where humans score 60%. GPT-5 has improved on this. But the fundamental point - that impressive performance on specific benchmarks does not equal general reasoning capability - applies to every model evaluation you will do in 2026.

For a practical framework on which current AI platform suits which business use case, our best AI chatbots for business guide covers the current landscape.

ChatGPT New Features 2026
Full breakdown of GPT-5.5 and every update through June 2026 - the models that replaced o3.

OpenAI Statistics 2026
OpenAI's revenue, valuation, and IPO status - the business context behind o3's development.

Claude AI Statistics 2026
How Anthropic's models now dominate the SWE-bench leaderboard that o3 once led.

AI Coding Tools Statistics 2026
Current AI coding tool benchmark performance beyond o3 - Cursor, GitHub Copilot, Claude Code.

ChatGPT vs Claude: Which AI Is Better for Business?
How the GPT-5 family and Claude compare across business workflows in 2026.

What is ChatGPT? Complete Guide 2026
How o3 fit into ChatGPT's model family and what replaced it.

When is OpenAI o3 being retired?
OpenAI announced on May 28, 2026 that o3 will be retired from ChatGPT on August 26, 2026, following a 90-day sunset period, per OpenAI's official release notes. The retirement applies to ChatGPT only - no API retirement date has been announced. Existing ChatGPT conversations using o3 will automatically continue on the corresponding GPT-5 model after retirement. API users with existing integrations can continue using o3 via API until a separate deprecation is announced.

What is OpenAI o3's score on the AIME math benchmark?
o3 scored 96.7% on AIME 2024 and 88.9% on AIME 2025, compared to o1's scores of 74.3% and 79.2% respectively, per OpenAI's official benchmarks. AIME tests advanced multi-step mathematical reasoning requiring genuine problem-solving rather than memorized answers. These remain strong scores. For context, o4-mini achieves 99.5% pass@1 on AIME 2025 with Python interpreter access, and GPT-5 with thinking also outperforms o3 on AIME.

How does o3 score on GPQA Diamond?
o3 scored 87.7% on GPQA Diamond, testing PhD-level science questions in biology, chemistry, and physics designed to be unsearchable. This surpasses average human expert performance of approximately 70% and compares to o1's 78%. Gemini 2.5 Pro scores 86.4% and Claude Opus 4.6 scores 84.2% on the same benchmark. These are all close - GPQA Diamond is no longer a clear differentiator between frontier models.

What did o3 score on ARC-AGI and what does it mean?
o3 scored 75.7% on ARC-AGI-1 at low compute and 87.5% at high compute, surpassing average human performance of 85%. On the newer ARC-AGI-2 benchmark, o3 scores only 2.9% versus 60% for average humans. This gap is the most honest data point in the o3 story - impressive performance on specific benchmarks does not equal general reasoning capability. In 2026, Gemini 3.1 Pro has reached 77.1% on ARC-AGI-2, showing how quickly the frontier moves.

What is o3's SWE-bench score?
o3 scored 71.7% on SWE-bench Verified, which tests AI on real GitHub issues from production software repositories. This compared favorably to o1's 48.9% when o3 launched. However, GPT-5 scores 74.9% on the same benchmark per OpenAI's developer introduction, surpassing o3. Claude Fable 5 scored 95.0% before its June 12 suspension. Anthropic models currently occupy the top 6 positions on SWE-bench - the benchmark frontier has moved significantly beyond o3.

How does o3 compare to GPT-5?
GPT-5 outperforms o3 across multiple dimensions per OpenAI's official introduction. GPT-5 scores 74.9% on SWE-bench versus o3's 71.7%. GPT-5's responses are approximately 80% less likely to contain factual errors than o3 when thinking. GPT-5 uses 50-80% fewer output tokens than o3 for equivalent performance. GPT-5 was preferred over o3 by testers 70% of the time on frontend web development tasks. For teams migrating from o3, GPT-5 and o4 represent genuine upgrades across most use cases.

What replaced o3 in ChatGPT?
In ChatGPT's model picker, GPT-5.4 Thinking is now the primary reasoning model for paid users following the simplified model picker update in June 2026. The picker now shows three primary options: Instant, Thinking, and Pro. For agentic coding specifically, GPT-5.3 Codex is the successor. For complex reasoning API tasks, o4 is o3's direct API successor. o3 remains available in ChatGPT until August 26, 2026 and via API with no announced retirement date.

How much does the o3 API cost in 2026?
Following an 80% price reduction in March 2026, o3 costs $2 per million input tokens and $8 per million output tokens at standard pricing. o3-pro costs $20/$80 per million tokens. Note that o-series models generate hidden reasoning tokens billed as output but not visible in API responses - actual costs per visible response are typically 2-4x higher than the visible token count suggests. For current pricing on o4 and GPT-5 models, check OpenAI's API pricing page directly.

What are OpenAI o3's benchmark scores?
OpenAI o3's key benchmark scores are: AIME 2024 math - 96.7%; GPQA Diamond PhD science - 87.7%; SWE-bench Verified software engineering - 71.7%; ARC-AGI-1 general reasoning - 87.5% at high compute; ARC-AGI-2 - 2.9% (humans score 60%); Codeforces competitive programming - 2,727 Elo; FrontierMath research math - 25.2%; MMLU general knowledge - 91.6%. o3 is being retired from ChatGPT on August 26, 2026, per OpenAI's May 28 announcement. GPT-5 and o4 are its successors.

When is OpenAI o3 being retired?
OpenAI o3 will be retired from ChatGPT on August 26, 2026, following a 90-day sunset period announced May 28, 2026, per OpenAI's official release notes. The retirement applies to ChatGPT only. No API retirement date has been announced. GPT-5.4 Thinking and o4 are the recommended replacements for complex reasoning tasks. API users can continue using o3 indefinitely until a separate deprecation announcement.

How does o3 compare to GPT-5 on benchmarks?
GPT-5 outperforms o3 across multiple dimensions per OpenAI's official announcements. GPT-5 scores 74.9% on SWE-bench Verified versus o3's 71.7%. GPT-5 with thinking is approximately 80% less likely to hallucinate than o3. GPT-5 uses 50-80% fewer output tokens for equivalent performance. GPT-5 was preferred over o3 by testers 70% of the time on frontend coding. GPT-5 also shows significantly lower deception rates (2.1% vs o3's 4.8%) and improved honesty when tasks cannot be completed.

What is the o3 ARC-AGI-2 score and why does it matter?
o3 scores 2.9% on ARC-AGI-2 compared to 60% for average humans, per the ARC Prize organization. This followed o3's impressive 87.5% on the original ARC-AGI-1 benchmark. The gap demonstrates that strong performance on a specific benchmark version does not represent general reasoning capability. In 2026, the frontier has moved - Gemini 3.1 Pro achieves 77.1% on ARC-AGI-2. The o3 ARC-AGI-2 result remains the most important data point for understanding the gap between AI benchmark performance and genuine general intelligence.

o3's benchmark scores tell a story that is both genuinely impressive in its context and now primarily historical. The AIME, GPQA, SWE-bench, and Codeforces results represented real capability improvements over o1 when o3 launched in April 2025. The 80% price reduction made those improvements broadly accessible.

In June 2026, the practical picture is clear. ChatGPT users should migrate to GPT-5.4 Thinking or o4 before August 26. GPT-5 outperforms o3 on coding, hallucination rates, and efficiency while using 50-80% fewer output tokens. API users have more time but should evaluate migration given GPT-5's improvements.

The ARC-AGI-2 result - 2.9% where humans score 60% - remains the most honest data point in the entire o3 story regardless of what model you use today. Benchmark performance on specific tests does not equal general intelligence. That lesson applies to every model evaluation you will do in 2026 and beyond.ng else. The benchmark scores point precisely to where the performance premium is real.

📨 Don't miss tomorrow's edition. Subscribe free to AI Business Weekly and get our 2026 AI Tools Cheat Sheet instantly - bite-sized AI news every morning, zero hype.

Subscribe Free