Best AI for Physics: Which Model Actually Solves It

The best AI for physics in October 2026 is a three-way tie at the top. Claude Opus 5.5, GPT-6 Astra and GPT-6.1 Sol all score 31.7% on CritPt, the research-level physics test that Artificial Analysis runs. Meta's Muse Wonder Full edges ahead at 32.9%, but only coders can reach it. Those scores sound weak, and they are misleading. When physicists re-graded popular physics benchmarks in September 2026, they found that 238 of 250 answers marked wrong were broken questions or broken grading. Only 12 were real model mistakes.

So the useful question is not which chatbot knows physics. All of the frontier models do. The useful questions are which one your plan actually gives you, which thinking setting to use, and how to hand it a problem with a diagram. This guide answers all three with published scores, each stamped with the model version and the test that produced it. It ends with the best free option and a simple rule for checking any answer before you hand it in.

The Key Takeaways

• Top tier is a tie: Claude Opus 5.5, GPT-6 Astra and GPT-6.1 Sol score 31.7% on CritPt, within one challenge of each other. Gemini 4 Argon scores 27.1%.
• The scores undersell the models: after experts fixed flawed questions, GPT-5.6 Sol's physics score on Humanity's Last Exam rose from 47.3% to 78.7%. Frontier models reached 87% to 92% on undergraduate physics.
• The thinking setting matters most: GPT-6 Sol scores 4.0% on CritPt with reasoning off and 30.9% at Max. Always pick the highest thinking level your plan allows.
• Diagrams are not optional: on image-based olympiad problems, scores fell from 72.75% to 49.85% without the diagram. A careful text description got back to 71.73%.
• Best free pick: Claude's free plan includes Sonnet 5.5, which scores 31.4% at its Max setting. ChatGPT's free GPT-6 Luna tops out at 19.4%.

What Is the Best AI for Physics Right Now?

The best AI for physics right now is Claude Opus 5.5 or GPT-6, with Gemini 4 Argon a step behind. The cleanest current ranking is CritPt, a set of 70 unpublished research problems written by more than 50 working physicists across 30 institutions and 11 subfields. Artificial Analysis grades every major model on CritPt independently and counts it as 10% of its Intelligence Index v4.3.2.

The table shows each model at its best published setting, plus how you actually get it. Scores come from the Artificial Analysis CritPt leaderboard and its model pages, read on October 18, 2026.

Why the Top Physics Scores Are a Tie

The top CritPt physics scores are a statistical tie, because the test is small. GPT-6.1 Sol also scores 31.7%, but OpenAI offers it only in ChatGPT Work and Codex, not in the chat window. CritPt has 70 test challenges, so one solved challenge is worth about 1.4 points. Every model from 30.9% to 32.9% sits within two challenges of the leader. Muse Wonder Full, the 32.9% leader, launched on October 12, 2026, for subscribers of Muse Code, Meta's coding tool, so it is not in a regular chat app.

Below the tie, the gaps are real.

Gemini 4 Argon trails by about five points. The cheap, fast models (GPT-6 Luna, Claude Haiku 5.5 at 18.9%, Gemini 3.8 Flash at 18.3%) score a little over half of what the frontier models score.

Best AI for Physics by Level

The best AI for physics depends on the level of the problem, because the evidence at each level comes from a different test. Homework, olympiad and research physics each have their own scoreboard, and the leader is not the same on all three.

Homework and University Physics

For homework and undergraduate physics, any frontier model with thinking turned on is accurate enough. The best evidence is an expert audit of UGPhysics, a benchmark of undergraduate physics problems. After physicists removed or repaired broken questions, GPT-5.6 Sol scored 92.07%, Gemini 3.1 Pro 90.85% and Claude Fable 5 87.80%. Those scores come from the previous generation of models, tested in September 2026.

At this level, the model matters less than how you use it.

Pick the highest thinking setting, give the model the full problem including any diagram, and ask for the working, not just the final number. If you are studying rather than checking homework, a tutoring mode such as ChatGPT Study Mode walks you through the steps instead of handing you the answer.

Olympiad Physics

For olympiad physics, Google's Gemini models had the strongest record in the most recent independent test. The ScienceArena benchmark, published on August 31, 2026, graded 14 models on 13 science competitions with the official marking schemes. Gemini 3.1 Pro scored 29.45 out of 30 on the 2025 International Physics Olympiad (IPhO) theory paper, above the best human theory score of 29.2. Across the whole physics suite, Gemini 3.1 Pro reached 89.5% of full marks, the highest of any model.

OpenAI won the newest exams. On four 2026 contests (three physics olympiads and one chemistry olympiad) held back as a fresher test, GPT-5.5 had the highest score at 93.31%. Across five repeated runs of the IPhO paper, GPT-5.5 averaged 29.75, Gemini 3.5 Flash 28.75 and Claude Opus 4.7 26.75. ScienceArena tested models one generation behind the current ones, and many results came after the 2025 exams were public, which the authors flag as a contamination risk. Treat the Gemini edge as a pattern, not a verdict on Gemini 4 Argon.

Research-Level Physics

For research-level physics, Claude Opus 5.5 and GPT-6 are joint leaders, and no model is reliable on its own. CritPt is the research test, and its maintainers say every problem is currently under expert review. An independent expert audit published in September 2026 found benchmark errors in 21 of the 56 CritPt challenges it checked. A researcher should treat any model as a fast, well-read colleague whose working needs checking line by line.

Why AI Physics Scores Look So Low

AI physics scores look low mainly because the benchmarks contain broken questions and broken grading, not because the models fail at physics. That is the finding of an expert re-grading study submitted on September 11, 2026. Faculty and graduate researchers in each subfield checked the problems that GPT-5.6 Sol had supposedly failed, along with the reference answers and the grader's verdicts.

What the Physics Experts Found

The physics experts found that almost every apparent model error was a fault in the test. The audit sorted every apparent error into one of three types. A benchmark error is a wrong reference answer or an ambiguous question. A grader error is a correct answer marked wrong, often because it was written in an equivalent form. Only the third type is a real model error. Across four benchmarks, experts reviewed 250 answers that had been marked wrong. They traced 238 (95.2%) to faults in the questions or the grading, and only 12 to the model. As the authors put it, "Frontier models are not failing these physics problems. The benchmarks are failing to pose them."

How Much the Physics Scores Changed

Fixing the physics tests raised every model's score, in some cases by more than 60 points. The experts repaired or removed the flawed questions and re-ran three frontier models of the previous generation. Scores are the average of four attempts, and HLE-Physics and CMT-Benchmark were run with tools enabled:

CritPt moved the most. Artificial Analysis lists GPT-5.6 Sol at 32.3% on all 70 challenges. On the 54 challenges the experts kept after repairing 19 and excluding 2, GPT-5.6 Sol averaged 87.5% and solved 94.4% in at least one of four attempts. The authors conclude that current benchmarks "substantially understate" what frontier models can do on well-posed physics problems.

That does not make the leaderboard useless. Every model faces the same flawed questions, so the ranking still says something even when the absolute number does not. For a plain-language guide to what each test measures, see our explainer on AI benchmarks such as GPQA and Humanity's Last Exam.

The Thinking Setting Matters More Than the Model

The thinking setting changes a model's physics score more than switching to a rival model does. Artificial Analysis publishes CritPt scores for each effort level, and the spread inside one model is far wider than the spread between the top models:

GPT-6 Sol with reasoning off solves almost no research physics. At Max it is in the top tier. That is why the plan matters as much as the model. ChatGPT Plus ($20 a month) runs GPT-6 Sol at Medium and High in the chat window, where it scores about 25%. Extra High and the Astra-powered GPT-6 Pro need a Pro plan from $100 a month.

The practical rule is simple. In ChatGPT, pick the highest thinking level your plan shows. In the Gemini app, choose the model's high effort level, which Google lets you set per model. In the Claude app, pick the highest effort setting it offers. For quick unit conversions this barely matters, but for anything with more than two steps, a low setting is the most common reason a good model gets physics wrong.

How to Give an AI a Physics Diagram

Give an AI the diagram itself or a complete written description of it, because leaving the diagram out costs more accuracy than any choice of model. ScienceArena tested this on 12 olympiad questions in physics, chemistry and biology that depend on a figure, using 12 vision-capable models. Each question was run three ways:

A careful description worked almost as well as the picture, and leaving the figure out cost about 23 points. The olympiad medalists who reviewed the answers traced many failures to visual grounding, meaning reading the figure correctly. Others came from losing track of a long problem, not from missing terminology.

A four-step routine fixes most of this:

1. Attach a clean photo or screenshot of the whole problem, diagram included. Crop out other questions on the page.
2. If you can't attach an image, describe it completely. Name every object, every labelled length, mass, angle and charge, and the direction of every arrow. Say what is fixed and what can move.
3. Ask the model to restate the setup before solving. A prompt that starts with First list every given quantity with units and describe the diagram in your own words. Then solve. exposes a misread figure before it ruins the answer.
4. Check the units and the limiting cases. Ask what happens to the answer if a mass goes to zero or an angle goes to 90 degrees. A wrong setup usually fails that test.

The Best Free AI for Physics

The best free AI for physics is Claude's free plan, because it includes Claude Sonnet 5.5, a model that scores 31.4% on CritPt at its Max setting. That puts it level with the paid flagships on paper. At High effort Sonnet 5.5 drops to 24.6%, so use the highest effort setting the free app offers.

ChatGPT's free plan runs GPT-6 Luna, OpenAI's lightweight model, and free users can switch on Think for harder questions. GPT-6 Luna scores 19.4% at its best setting, about two thirds of the frontier level. It handles textbook problems well but is the weakest of the three on multi-step work. Our guide to what ChatGPT's free plan includes covers its limits.

Gemini's free tier changed on October 9, 2026. Without a paid plan, Gemini uses Auto, which sends most prompts to Gemini Flash-Lite and passes harder ones to Flash or Pro. You cannot pin Gemini 4 Argon without Google AI Pro at $19.99 a month. For a hard problem on the free tier, Claude is the safer first stop, and ChatGPT with Think switched on is a good second opinion.

Get a Second Opinion Without a Second Subscription

The fastest way to catch a wrong physics answer is to ask a second model the same question. When two models built by different companies reach the same result on their own, a shared mistake is less likely than with one answer. When they disagree, you have found the step to check by hand.

Doing that with separate apps is slow.

It means two subscriptions, two tabs and pasting the problem and the diagram twice. Fello AI turns it into one click: ask Claude, then send the same problem to GPT or Gemini in the same window, without a second bill. It runs natively on Mac, iPhone and iPad and includes a free tier, so you can test the workflow on tonight's homework before paying for anything.

Which AI Should You Use for Physics?

Use Claude Opus 5.5 if you pay for one plan and want the strongest physics model in a chat window for $20 a month. GPT-6 Astra matches it, but in ChatGPT's chat window Astra needs the Pro plan from $100 a month. If you pay nothing, use Claude Sonnet 5.5 on Claude's free plan at its highest effort setting.

The model is the smaller decision. The bigger ones are turning thinking all the way up, always including the diagram, and checking important answers with a second model. A student who does those three things with a free plan will get more physics right than one who uses the best paid model on its lowest setting. For the same comparison on pure maths, read our guide to the best AI for math, and for study tools beyond physics, see the best AI for students.

FAQ