Compare in promptflow: Send One Prompt to Several AI Models and Line Up Text or Images

We asked three image models for the same picture, two images each. One column cost 8 PFC. Another cost 24. Same subject, same number of images, three times the price, and the results look clearly different.

That side-by-side view is what Compare is for. promptflow has four modes: Chat, Compare, Debate, and Co-Draft. This post covers Compare only.

• Compare sends one prompt to several AI models at once and shows the answers in columns. It works for text and for images.
• For text, the default list has a "Recommended" section and a "Choose by maker" section with 15 models from 9 makers. The advanced filter opens up 303 models. Images get the same two sections, with 147 models in total.
• Every column shows the PFC it used, so quality and cost sit on the same screen.

What the Compare screen does

Open Compare in promptflow and you land on "Comparison Mode." A toggle at the top switches between "Text" and "Image."

Below it are the "Comparison slots (model x Persona)." A slot is one row for one candidate. In each row you pick a model and, optionally, a Persona, which gives the model a role or point of view. Leave it at "No persona (default)" to get the model's own answer. "Add slot" adds another row. You can put the same model in two slots with different Personas and compare the viewpoints.

Type your prompt and press "Send." A "Cost confirmation" dialog shows the "Estimated total cost." Nothing is sent until you press "Send" in that dialog. Then every slot gets the prompt at the same time. If your balance is too low, it stops before sending with "Insufficient PFC balance," and "View balance & plans" takes you to your balance.

In text mode there is also a "🌐 Web Search" checkbox. Models that support it can search the web when they answer. Each model decides from the question whether to search. Slots with a model that does not support it show "🌐 Unsupported."

Each answer column also tells you whether a search happened. A column that searched shows something like "🌐 Web search used (2)" with links to the pages it used. A model without search shows "🌐 Web search not supported — model knowledge only," and a column that did not search shows "🌐 No search was performed." When you ask about recent events, you can see which answers came from the web and which came only from what the model learned.

Same model, different Persona

A slot can hold any model, including one that is already in another slot. Put GPT-5.6 Terra in two slots, leave one at "No persona (default)," give the other a Persona with a specific role, and send. Now the comparison shows how the role changes the answer, not how the model does.

Results are keyed by slot, so two columns with the same model name stay separate. Each column shows "Persona: {name}" and the PFC it used. The same screen works for comparing models and for comparing roles.

Comparing text

Here, three models got the same prompt: "List three tips for writing clear work emails for new employees." The models are GPT-5.6 Terra (OpenAI), Claude Sonnet 5.5 (Anthropic), and Gemini 3.8 Flash (Google).

All three open with a clear subject line, and all three put "lead with the main point" second. That is the part of the answer you can lean on first, after checking it against your own style guide. Where they differ is the wording, the level of detail, and the example subject lines each one suggests. Laid out like this, you can see at a glance what every model said, where they split, and what only one model raised.

Cost is shown per column too. The "Estimated total cost" before sending was 8 PFC. After the answers came back, the columns showed 1, 2, and 1, and the "Total PFC cost" was 4. The estimate is a preview before sending. You are charged for what the models used.

Each column has pin and copy buttons, so you can keep the answer you want or paste it elsewhere. Under the answers, "Continue chat with {model}" lets you keep talking to the model whose answer you liked, without starting over.

Letting the models review each other

"Run cross-review between AIs" sends the answers back to the models that wrote them. Each AI reads the other AIs' answers and points out what is missing or wrong. The button works once at least two models have answered.

While it runs, the panel shows "Running cross-review...", "Reviewing {count} models," and the elapsed time, so you can see it is still moving on long answers. When it finishes, "Cross-Review Results" lists each model's review as "Completed" or "Failed," along with a count such as "3/3 models completed" and the "Total PFC cost" of the review. If a review fails, "Try again" runs it again with the same answers and the same models.

From the review results, "Continue with {model}" keeps talking to a reviewer and "Save as artifact" keeps the reviews as a document. Having the models flag each other's weak spots first tells you where to look when you read the answers yourself.

The models you can pick

Click the model name in a slot and the "Choose a model" dialog opens. At the top are the "Search by model or maker" box and a "Sort" menu. Below them, model cards are grouped under "⭐ Favorites," "Recommended," and "Choose by maker." Click the star on a card to add it to Favorites, and it shows at the top of the list next time. "Sort" offers "Recommended," "Price: low to high," "Price: high to low," and "Name."

The count line in the default view reads "15 models · 9 makers · Recommended." That count covers only the "Choose by maker" section, not the models under "Recommended." These are the 15 models under "Choose by maker":

"Recommended" shows the newer models of each maker's main series. Claude Sonnet 5.5 and Gemini 3.8 Flash in the example above came from there. Others include GPT-6.1 Sol, Claude Opus 5.5, Gemini 4 Argon, Grok 4.7, Qwen3.8 Max (0902), Mistral Medium 3.5, and Command A.

To search everything, open "Browse all models (advanced)" and set "Featured" to "All." The count line changes to "303 models · 50 makers." Models are added and replaced over time, so treat the on-screen count as the current number.

Each card shows an estimated PFC price for input and output per million tokens, a short description, what the model is good at, and the inputs it accepts (text, image, audio).

Those estimates vary a lot. On screen, GPT-5.6 Luna shows 35 for input and 208 for output, o3 shows 346 and 1381, Claude Opus 4.7 shows 863 and 4315, and GPT-6.1 Sol shows 4391 and 16465. On output alone, GPT-6.1 Sol is about 80 times GPT-5.6 Luna. A comparison made only of top-end models costs more per run, so one approach is to start with cheaper models and add an expensive one only for the prompts where answers split.

Which models you can pick depends on the plan. The pricing page says the Free plan includes Compare but only lightweight models, and Personal Pro and above can choose every model. The screenshots in this post come from an account that can choose every model.

Comparing images

Switch the toggle to "Image" and the model list changes to image models only. "Images per model" sets how many pictures each model makes, from 1 to 4.

The next screen shows three models given the same subject, "Concept art of a park and skyscrapers in the morning light," two images each.

The models are Nano Banana 2 (Gemini 3.1 Flash Image) from Google, Seedream V4 from ByteDance, and Qwen Image 2512 from Alibaba. Column headers show each model's internal ID. The "imagen-4-fast" header on the Nano Banana 2 column is promptflow's internal ID for that model.

The pictures differ clearly from model to model:

• Nano Banana 2: wide images in a colorful illustrated style. Both fit a wide skyline behind the park. The first has a pond in front, and the second puts a path with people walking toward the towers.
• Seedream V4: tall images in an illustrated style. The first has a pink and orange sunset-colored sky with the towers in the middle distance. The second shows glass towers lit by the sun behind the trees.
• Qwen Image 2512: tall images in an illustrated style. A path, benches, and people are in front, with the towers in the background.

The aspect ratio differs by model too, so if you already know where the image will go, a wide header or a tall post, that is part of the comparison. If one model fails, the failure shows only in its column and the other columns keep their images.

Cost differs as well: 24, 12, and 8 PFC for two images, 44 PFC in total. The "Estimated total cost" before sending is the model's price per image times the number of images. Images that were not made are not charged.

The default image list is also split into "Recommended" and "Choose by maker." Its count line reads "9 models · 7 makers," and these are the 9 models under "Choose by maker":

Setting the advanced filter to "All" shows 147 image models from 34 makers. "Recommended" also lists models such as Recraft V4 Pro, Ideogram V2, FLUX 2 Flex, and Krea 2 Large.

How to read a comparison

Sort the answers into three groups: points every model made, points where they split, and points only one model raised. That tells you where to spend your checking time.

Agreement is not proof. Models can lean on the same sources and repeat the same mistake. Check even the shared points against the original source before you use them.

Where answers split, you can dig in without leaving Compare. Run a cross-review to have each model name the weak spots in the others. If you want one model's reasoning, use "Continue chat with {model}" and ask about its assumptions. Sometimes the split turns out to come from a vague prompt. Fix the prompt and send it again to the same slots.

Keep the setup identical across models, not just the question. If the conditions differ, you cannot tell a model difference from an input difference. Compare sends one input to every slot, so the prompt itself is always the same.

For images, keep the subject and the number of images the same. You then see which model fits the use you have in mind, next to what each picture costs.

The short version

• Compare sends one prompt to several AI models at once and lines up text or images side by side.
• You can pick from 303 text models and 147 image models. The default lists show a "Recommended" section and a "Choose by maker" section.
• Every column shows the PFC it used, so you weigh quality and cost on one screen.
• Agreement between models is not proof. Use the comparison to decide what to check.

Terms

• PFC: promptflow's usage credit. It puts different provider costs on one scale.
• Slot: one row in Compare for one candidate, with a model and an optional Persona.
• Persona: a setting that gives a model a role or point of view. The same model can answer from a different angle with a different Persona.
• Cross-review: an action where each AI reviews the other AIs' answers.
• Artifact: a saved document in promptflow, such as a draft or a table, that you can reuse outside the conversation.

This series shares practical knowledge from building and operating our flagship SaaS, Prompt Flow Studio, and other AI products. If you are interested, please like and follow.

Prompt Flow Studio is a multi-AI workspace where you use, compare, and debate multiple AIs on one screen. Learn more at https://prompt-flow.studio
Our other products are listed at https://www.eln.ne.jp/products

EarthLink Network builds its own products to run the whole company on AI — each one born from a need along the way. See the full index of what we build:

→ The 18 products EarthLink Network builds in-house

For more about the company and each product, visit www.eln.ne.jp.