OpenAI o3 Benchmarks 2026: Every Score & What Replaced It

Last Updated: November 1, 2026

OpenAI o3 scored 88.9% on AIME 2025, 83.3% on GPQA Diamond, 69.1% on SWE-bench Verified and a 2,706 Codeforces Elo in OpenAI's April 16, 2025 launch post. o3 was retired from ChatGPT on August 26, 2026, and OpenAI's o3 API snapshot shuts down on December 11, 2026, with gpt-5.6-sol named as its replacement.

This page keeps every official o3 benchmark score, each labelled with the date OpenAI or the ARC Prize Foundation published it, and adds how OpenAI's GPT-6 family (Astra, Sol and Luna) compares using only benchmark figures OpenAI published in September 2026.

One warning before the numbers: OpenAI scored GPT-6 on mostly different benchmarks than o3. The GPT-6 launch posts lean on newer agent and computer-use tests such as DeepSWE v1.1 and OSWorld 2.0, so a straight o3 versus GPT-6 comparison only works on a few shared tests. This guide flags where the comparison is direct and where it is not.

🎯 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

No in ChatGPT, and only for a few more weeks in the API. OpenAI o3 was retired from ChatGPT on August 26, 2026, and the o3 API snapshot is scheduled to shut down on December 11, 2026.

ChatGPT (retired August 26, 2026): OpenAI announced the o3 retirement in its ChatGPT release notes in late May 2026, with a 90-day transition ending August 26, 2026, as reported at the time. OpenAI's original o3 launch post now carries a note that o3 and o4-mini "are no longer available in ChatGPT and are deprecated in the API."

API (shutdown December 11, 2026): On June 11, 2026, OpenAI notified developers that the o3-2025-04-16 and o3-pro-2025-06-10 snapshots will shut down on December 11, 2026, according to OpenAI's API deprecations page. OpenAI lists gpt-5.6-sol as the replacement for o3, and gpt-5.6-sol in pro reasoning mode as the replacement for o3-pro.

Source: OpenAI API deprecations, checked November 1, 2026.

How much time is left: Between November 1 and December 11, 2026, any team still calling o3 through the API has 40 days to finish moving. Test real production prompts on the replacement before switching, because a new model can change output length, format and cost per request.

OpenAI's current flagship API lineup is GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna, per OpenAI's models page. Per-token prices for each are broken down in our OpenAI API pricing guide.

OpenAI o3 was a reasoning model trained with large-scale reinforcement learning to work through a problem before answering, which is why its scores jumped on math, science and coding tests compared with o1.

OpenAI said in its April 16, 2025 launch post that o3 was the first of its reasoning models able to use and combine every tool inside ChatGPT, including web search, Python and image analysis, as part of its reasoning. OpenAI also said o3 could reason over photos, charts and sketches directly inside its chain of thought.

The trade-off with o3's longer reasoning was compute per answer. OpenAI's launch post says o3's performance kept climbing the longer it was allowed to think, which means slower and more expensive responses on hard problems.

In evaluations by external experts, o3 made 20% fewer major errors than o1 on difficult real-world tasks, especially in programming, business and consulting, and creative ideation, according to OpenAI's o3 announcement (April 16, 2025).

In its April 16, 2025 launch post, OpenAI reported o3 at 91.6% on AIME 2024, 88.9% on AIME 2025, 83.3% on GPQA Diamond, 69.1% on SWE-bench Verified, 82.9% on MMMU and 2,706 Elo on Codeforces. Every figure in the table below comes from that post unless marked otherwise.

Source: OpenAI, Introducing OpenAI o3 and o4-mini (April 16, 2025, with a July 28, 2025 SWE-Lancer update). Some sites still quote higher o3 numbers from OpenAI's December 2024 preview of a pre-release configuration. The table above uses the figures OpenAI published for the o3 model it actually shipped.

AIME: 91.6% on 2024 and 88.9% on 2025 (no tools, April 2025)

OpenAI o3 scored 91.6% on AIME 2024 and 88.9% on AIME 2025 without tools, up from o1's 74.3% and 79.2%, per OpenAI's April 16, 2025 launch post. The American Invitational Mathematics Examination rewards multi-step reasoning, which is where o-series models made their biggest gains.

With a Python interpreter, o3 reached 98.4% pass@1 on AIME 2025, and its sibling o4-mini hit 99.5%, according to the same OpenAI post. OpenAI itself warned that tool-assisted AIME scores should not be compared with no-tool results, because access to a computer makes the exam much easier.

Our read: by late 2026, AIME no longer separates frontier models, which is why OpenAI's GPT-6 launch reported FrontierMath instead. For quantitative work such as financial modelling, o3's AIME score is best treated as a floor, not a buying signal.

FrontierMath: where GPT-6 Astra now sets the bar

FrontierMath, run by Epoch AI, tests unpublished research-level math problems that cannot be memorised. The benchmark charts in OpenAI's April 2025 o3 launch post did not include FrontierMath.

OpenAI reported GPT-6 Astra at 97.6% on FrontierMath Tier 4 (v2) at its September 3, 2026 launch, as reported by DataNorth and Vellum. Vellum's analysis also points out that OpenAI funded FrontierMath and has access to part of the problem set, so vendor-reported FrontierMath scores deserve some caution.

For the business side of OpenAI's model releases, including revenue and user numbers, see our OpenAI statistics guide.

GPQA Diamond: 83.3% (no tools, April 2025)

GPQA Diamond is a set of graduate-level biology, chemistry and physics questions written to be hard to look up online. OpenAI o3 scored 83.3% on GPQA Diamond without tools, per OpenAI's April 16, 2025 launch post.

Here is how o3 compared with OpenAI's other reasoning models on GPQA Diamond in that same April 2025 post, all measured without tools:

For pharma, biotech and materials science teams, o3's GPQA Diamond result showed a clear step up from o1 on expert science questions. OpenAI's GPT-6 launch coverage focuses on agent and computer-use tests rather than GPQA Diamond, so this page does not list a GPT-6 GPQA figure we could confirm across sources.

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: 69.1% (April 2025)

SWE-bench Verified tests AI on real GitHub issues from popular open-source repositories. OpenAI o3 resolved 69.1% of a fixed 477-task subset, up from 48.9% for o1, per OpenAI's April 16, 2025 launch post.

On the SWE-Lancer IC SWE Diamond set of freelance coding tasks, o3 earned $86,100 in task payouts versus $43,958 for o1, roughly double, according to OpenAI's launch post, which carries a July 28, 2025 SWE-Lancer update.

OpenAI's GPT-6 launch materials, as reported by DataNorth and Vellum, use DeepSWE v1.1 rather than SWE-bench Verified. On DeepSWE v1.1, OpenAI reported GPT-6 Astra at 74.1%, GPT-6 Sol at 68.8% and GPT-6 Luna at 66.6% (September 2026). DeepSWE and SWE-bench Verified are different tests, so those GPT-6 numbers should not be lined up directly against o3's 69.1%.

For development teams, o3's 69.1% on SWE-bench Verified is now a historical baseline. Current adoption and performance data for coding tools is in our AI coding tools statistics guide.

Codeforces Elo: 2,706 (April 2025)

OpenAI o3 reached a 2,706 Codeforces Elo with terminal access, up from 1,891 for o1 and 2,073 for o3-mini, per OpenAI's April 16, 2025 launch post. OpenAI said o3 set a new state of the art on Codeforces at launch.

For a side-by-side look at AI coding tools including Claude Code, Cursor and GitHub Copilot, see our AI coding tools guide.

ARC-AGI-1: 75.7% and 87.5% for the December 2024 preview, 53% for the released o3 at medium compute

The ARC Prize Foundation reported on December 20, 2024 that a preview version of o3 scored 75.7% on the ARC-AGI-1 semi-private set within its $10,000 compute limit, and 87.5% with a high-compute configuration using 172 times more compute (ARC Prize).

The o3 model OpenAI actually released in April 2025 scored lower when ARC Prize retested it: 41% (low) and 53% (medium) on ARC-AGI-1, and 1.9% (low) and 2.9% (medium) on the harder ARC-AGI-2, as reported by The Decoder on April 27, 2025.

OpenAI reported GPT-6 Astra at 99.9% on ARC-AGI-3 at its September 3, 2026 launch, against 7.8% for its predecessor, as reported by DataNorth and Vellum. ARC-AGI-3 is a newer test than the versions o3 took, so the GPT-6 and o3 ARC numbers are not directly comparable.

Our read: the drop from o3's preview score to its released-model score is the most useful lesson on this page. ARC results depend on the exact model version and compute budget, so always check which configuration a headline number came from.

GPT-6 beats o3 on the one major benchmark both were scored on with tools: OpenAI reported GPT-6 Astra at 57.2% on Humanity's Last Exam with tools (September 3, 2026), against 24.9% for o3 with Python and browsing (April 16, 2025). That is a 32.3-point gain, or about 2.3 times o3's score.

OpenAI launched GPT-6 Astra on September 3, 2026, and GPT-6 Sol and GPT-6 Luna on September 22, 2026, per DataNorth and TechCrunch. The table below lists OpenAI-published GPT-6 scores, with o3's result shown only where the same benchmark exists.

Sources: GPT-6 Astra figures from OpenAI's September 3, 2026 launch and GPT-6 Sol and Luna figures from OpenAI's September 22, 2026 announcement, as reported by DataNorth (Astra, Sol and Luna) and Vellum. The o3 figure is from OpenAI's April 16, 2025 o3 post. n/a means we found no OpenAI-published figure confirmed across sources. OpenAI ran most rival-model comparisons itself, so treat cross-vendor gaps as directional.

The Sol model now listed in OpenAI's API is GPT-6.1 Sol (gpt-6.1-sol), which OpenAI describes as near-Astra performance at a lower cost. For the model and feature history inside ChatGPT, see our ChatGPT new features 2026 guide, and for the GPT-5 generation that came between o3 and GPT-6, see our GPT-5 statistics page.

For most teams leaving o3, GPT-6.1 Sol is the sensible default, GPT-6 Luna suits high-volume jobs, and GPT-6 Astra fits only the hardest agent work. OpenAI's own one-for-one API replacement for o3 is gpt-5.6-sol.

Our verdict: GPT-6 Astra makes sense only where extra accuracy on long agent runs pays for a token price five times higher than GPT-6.1 Sol. For most former o3 workloads, start testing on GPT-6.1 Sol and drop to GPT-6 Luna wherever quality holds.

For wider benchmark and adoption context across active models, see our AI coding tools statistics guide and Claude AI statistics guide.

GPT-6.1 Sol, the most likely o3 replacement, costs $2 per million input tokens and $10 per million output tokens at standard rates, according to OpenAI's API pricing page checked November 1, 2026. OpenAI no longer lists o3 on that pricing page.

Sources: OpenAI API pricing and models pages, standard tier, checked November 1, 2026.

Worked example: a workload of 10 million input tokens and 2 million output tokens a month costs $40 on GPT-6.1 Sol ($20 input plus $20 output), $2 on GPT-6 Luna ($1 plus $1) and $200 on GPT-6 Astra ($100 plus $100). That 100x spread between Luna and Astra is why model choice matters more than the move off o3 itself.

Reasoning-heavy jobs produce more output tokens, so output price drives most API bills for o3-style work. Batch, priority and cached-input rates for every OpenAI model are in our OpenAI API pricing guide.

OpenAI o3's benchmark scores now work as a historical baseline: they show what a top reasoning model could do in April 2025, not what a team should buy in November 2026.

If you used o3 in ChatGPT: o3 has been gone from ChatGPT since August 26, 2026. Check that saved workflows, projects and custom GPT instructions built around o3 still behave as expected on the newer models.

If you call o3 through the API: plan the move before December 11, 2026. OpenAI's named replacement is gpt-5.6-sol, and the GPT-6 family is the current lineup worth evaluating.

If you compare models using benchmarks: check the date and model version behind every score. OpenAI o3's own numbers changed between its December 2024 preview and April 2025 release, and GPT-6 is mostly measured on newer tests that o3 never took.

OpenAI o3's ARC-AGI-2 result, 2.9% for the released model at medium compute in ARC Prize's April 2025 testing, remains a useful reminder that a strong score on one test does not prove general reasoning ability.

For a practical framework on which AI platform suits which business use case, see our best AI chatbots for business guide.

ChatGPT New Features 2026
Every major ChatGPT update in 2026, including the models that followed o3.

OpenAI Statistics 2026
OpenAI's revenue, valuation and user numbers, the business context behind o3 and GPT-6.

GPT-5 Statistics
Key numbers on the GPT-5 generation that sat between o3 and GPT-6.

AI Coding Tools Statistics 2026
Adoption and performance data for AI coding tools such as Cursor, GitHub Copilot and Claude Code.

ChatGPT vs Claude: Which AI Is Better for Business?
How ChatGPT and Claude compare across everyday business workflows.

OpenAI API Pricing
Current per-token prices for GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna.

When was OpenAI o3 retired?
OpenAI o3 was retired from ChatGPT on August 26, 2026, after a 90-day transition announced in late May 2026. In the API, OpenAI's o3 and o3-pro snapshots shut down on December 11, 2026, with gpt-5.6-sol named as the replacement.

What are OpenAI o3's official benchmark scores?
In OpenAI's April 16, 2025 launch post, o3 scored 91.6% on AIME 2024, 88.9% on AIME 2025, 83.3% on GPQA Diamond, 69.1% on SWE-bench Verified, 82.9% on MMMU and 2,706 Elo on Codeforces. The AIME and GPQA figures are without tools, and the Codeforces rating used terminal access.

How does GPT-6 compare to o3?
On Humanity's Last Exam with tools, OpenAI reported GPT-6 Astra at 57.2% in September 2026 versus 24.9% for o3 in April 2025, about 2.3 times higher. Most other GPT-6 results use newer tests such as DeepSWE v1.1 and OSWorld 2.0 that o3 was never scored on.

What did o3 score on ARC-AGI?
A December 2024 preview of o3 scored 75.7% on ARC-AGI-1 within ARC Prize's compute limit and 87.5% at high compute. The o3 model OpenAI released scored 53% on ARC-AGI-1 and 2.9% on ARC-AGI-2 at medium compute in ARC Prize's April 2025 retest.

What is o3's SWE-bench score?
OpenAI o3 scored 69.1% on SWE-bench Verified (a 477-task subset) in OpenAI's April 16, 2025 launch post, up from 48.9% for o1. OpenAI's GPT-6 launches report DeepSWE v1.1 instead, where GPT-6 Astra scored 74.1%.

Can I still use o3 in ChatGPT?
No. OpenAI retired o3 from ChatGPT on August 26, 2026.

Which GPT-6 model should replace o3?
GPT-6.1 Sol for most work, GPT-6 Luna for high-volume tasks and GPT-6 Astra for the hardest agent work. OpenAI's direct API replacement for o3 is gpt-5.6-sol.

How much does GPT-6.1 Sol cost in the API?
GPT-6.1 Sol costs $2 per million input tokens and $10 per million output tokens at standard rates, per OpenAI's pricing page checked November 1, 2026.

Why do o3 benchmark numbers differ between sites?
Some sites quote figures from OpenAI's December 2024 o3 preview. This page uses OpenAI's April 16, 2025 release figures and ARC Prize's retests of the model OpenAI shipped.

OpenAI o3's official April 2025 scores, including 88.9% on AIME 2025, 83.3% on GPQA Diamond and 69.1% on SWE-bench Verified, marked a real step up from o1 across math, science and coding.

As of November 1, 2026, o3 is retired from ChatGPT and its API snapshot shuts down on December 11, 2026. GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are OpenAI's current lineup, and on Humanity's Last Exam with tools, the one shared benchmark, GPT-6 Astra more than doubles o3's score.

The lasting lesson from o3 is about reading benchmarks: check the model version, the date and whether tools were allowed before trusting any headline number, whether it belongs to o3, GPT-6 or whatever OpenAI ships next.

📨 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

Written by the AI Business Weekly Editorial Team. Reviewed, edited and published by Sameer Khan. This article was AI-assisted, then reviewed by Sameer Khan before publishing.