Claude Opus 5 vs GPT-5.6 vs Gemini 3.7 Flash

Model comparisons often start with rankings. Rankings move, sometimes weekly, and a table of positions copied on a Tuesday can be wrong by Thursday. For a particular task, it helps to check the available features: what each model accepts, whether it can look anything up, and whether you can tell it to stop thinking.

Work down five questions about the task in front of you, in order, and you land on one model. All three carry the violet “✦ Pro” chip in the model picker, so a paid plan is the starting condition. These were our three flagships at the end of summer, and two have since been succeeded; the last section says by what.

Before you start: “GPT-5.6” is three entries

Type 5.6 into the model picker and you still get GPT-5.6 Sol, GPT-5.6 Terra and GPT-5.6 Luna, though Sol and Luna now carry a Deprecated chip. All three carry identical capability flags in our registry, so wherever a step below says GPT-5.6, all three qualify. Terra carries the Recommended star and is the reasonable default. The place the three come apart is step 4.

Careful. A yellow Deprecated chip means the entry now passes your message to the replacement named in its tooltip. Plain “GPT-5” has one (replacement GPT-5.4, discontinue date 11 December 2026), and since late September so do Claude Opus 5 and two of the GPT-5.6 entries. After you pick, read the name in the composer: that is the model that answers.

Step 1. Look at what you are about to attach

Images, PDFs, and text or code files work on all three.

Video is the fork. Of these three, only Gemini 3.7 Flash accepts video, and the app enforces it twice. Drop an MP4 into a thread on a Pro account and you get “Video uploads are a Pro+ feature” before the model is even considered. On a Pro+ account with a Claude or OpenAI model selected, you get “This model doesn’t support video. Switch to a Gemini model to send videos.” So if your task starts with a clip, step 1 has already answered the whole question.

PDFs deserve one caveat. All three read PDFs natively, but a text-heavy PDF now reaches them as the text we extract, and a scan of up to 20 pages is transcribed first, which can introduce recognition errors. Image-heavy files and longer scans still go as the file itself. Treat a scan as worth trying and not worth depending on.

Step 2. Ask whether the answer has to be current

Open a thread and look for the globe in the composer. On the web it is there once Intent Detection is off in Settings, then Tool Preferences; that setting is on by default and decides per message, hiding the effort menu in step 3 too.

It shows on Claude Opus 5 and on Gemini 3.7 Flash, and on none of the three GPT-5.6 entries, because web search is switched off for that family in our registry.

The consequence is blunt. Anything phrased as “what happened”, “current”, “as of this week” or “latest version” rules out GPT-5.6 in this app, however good its answer would otherwise be. Ask it about this week and you get a confident reply assembled from training data, the worst failure mode in this comparison, because it does not look like one.

Step 3. Decide whether you want the option to switch thinking off

All three reason. Only one family lets you stop it.

The reasoning effort button, itself a Pro control, opens a menu: Default, Auto, Low, Medium, High. On the GPT-5.6 entries there is a sixth row below a separator, Off. On Claude Opus 5 and Gemini 3.7 Flash that row does not exist.

We did not leave it out by choice. Google dropped the lowest thinking level from 3.7 Flash and its API rejects the value, so Off cannot even be approximated. Anthropic runs Opus 5 with thinking on by default, and our Auto setting hands it adaptive thinking with no effort level attached.

Warning. Reasoning effort is a single global preference that follows you across models. Set it to Off on GPT-5.6 Luna, switch the thread to Claude Opus 5, and Opus 5 goes on thinking. The setting has nothing to bind to there, so it is quietly ignored.

If most of your day is short factual turns where waiting on a thinking pass annoys you more than it helps you, the family with an Off switch is the one to be on.

Step 4. Estimate how long the thread will run

Every model here advertises a very large window. The advertised size is the one to ignore. What decides your experience is the point past which answers start drifting, and the app shows it: once the context meter leaves its neutral band, it names the token count past which answer quality tends to drop for that model.

That figure differs per model and we retune it, so it is not a number worth memorising. The ordering is stable enough to plan around. Claude Opus 5, GPT-5.6 Sol and GPT-5.6 Terra hold the most before the meter turns. Gemini 3.7 Flash sits below them. GPT-5.6 Luna is the lowest of the five entries named here, which is the one place the family splits.

A long research thread with a lot of pasted material therefore argues for Opus 5, Sol or Terra. If you want the mechanics behind that, what a context window actually is covers them.

Step 5. Send the same prompt twice

This step takes about a minute.

Ask your real question. Then branch the answer: the control is “Try Again” under the reply on web and “Branch here” on a phone. Pick a different model, and the alternative arrives in a new thread with the original left exactly as it was. Do it once more for the third model. Then keep whichever reply you would forward to a colleague without editing it first.

Check the label on a phone. “Try again” on mobile is the regenerate action and it replaces the answer you already have. The one that opens a second thread is “Branch here”.

No benchmark can run this step for you, because what you are measuring is whether the answer came out shaped the way you needed.

Where the steps usually land

Claude Opus 5. Long reasoning-heavy threads, dense documents, and questions that need something looked up. It always thinks, so short factual turns feel heavier than they need to be.

GPT-5.6, Terra unless you have a reason. General work where you want a hand on how much thinking happens, including none of it. Rules itself out the moment a question needs the live web.

Gemini 3.7 Flash. Anything with video in it, plus image and document work where you also want search available.

The same thing in one table

One thing you may have read that is wrong

There is a persistent claim that Gemini 3.7 Flash is unavailable in the EEA, the UK, Switzerland and Nigeria. Those four names are real and that exclusion list is real, and it belongs to a different product. Google’s own help page for Gemini Spark, its agent, states that Spark is “Available wherever Gemini Apps are supported, except in the European Economic Area, Nigeria, Switzerland, and the United Kingdom.” Gemini 3.7 Flash carries no regional exclusion here.

What changed after this was written

Two of the three have successors. In our picker, Claude Opus 5.5 took over from Opus 5 and GPT-6 Luna from GPT-5.6 Luna on 22 September, and GPT-5.6 Sol now hands over to GPT-6.1 Sol, added on 29 September.

Run the five steps on the successors and one answer moves: GPT-6.1 Sol has no Off switch, so in step 3 it sits with Opus and Gemini. GPT-5.6 Terra and Gemini 3.7 Flash are unchanged. Google's newer Gemini 3.8 Flash, out since 2 September for AI Pro and Ultra subscribers in its own app, is not in our picker.

Run it on your own question

The first four steps narrow the field. Use step 5 to compare the remaining models on your own task.