GuideArena
Comparing models in CSuite.
Every control in the Arena workspace, with a figure for each, followed by recipes for the comparisons people run most. The short version lives on the main guide; this page is for when you want to know exactly what happens when you press Run.
Part 1
Get around
1.1
The workspace
Open Arena from the rail, the last entry of the second group. It sends one prompt to several models at once and shows their results side by side, each with its time, tokens and cost. Every Run is a round, and a comparison keeps all of its rounds, so you can come back to them, tweak the prompt and run again.
- 1 · Comparisons
- Your project folder pill, the + for a new comparison, and every comparison with its round count and when it last changed. Rename or delete from a row’s More actions.
- 2 · Name
- Taken from the first round’s prompt; click it to rename the comparison in place.
- 3 · Rounds
- Round 3 of 3, with previous and next. The newest round shows after every Run.
- 4 · Models
- The lineup: how many models, with each model and its route in the tooltip. Opens the Type and the slots.
- 5 · Result cards
- One per model, in the round’s order: name and route, the result, then time, tokens and cost.
- 6 · Prompt
- Holds the prompt of the round on screen, so a small edit runs as a new round.
- 7 · Settings and Run
- The next round’s shared settings and the Blind eye, then Run, which becomes Stop while a round runs.
1.2
Starting a comparison
1
Create it
Press New comparison on the empty screen, the + beside the folder pill, or
⌘N. A comparison you opened and never ran is reused rather than doubled.2
Let it name itself
It reads New comparison until the first round, then takes the first line of that prompt, up to about 60 characters. Rename it from the header or from the row’s More actions; renaming also renames the folder its files are in.
3
Keep or delete it
Comparisons are listed in every project and stored in the app’s database on your machine. Delete asks first, then removes the comparison and every round in it; the files its rounds saved stay in your project folder.
1.3
Choosing the lineup
Open Models in the header. A new comparison starts with two empty slots; Add model adds one, up to six, and the × beside a slot removes it, down to two. Each slot is the usual model picker, with search and the All, Cloud and Local filter.
- The list holds every model your providers and local runtimes offer for the chosen type, including models you have hidden from the workspace pickers. Auto and RunPod are not offered: each slot is one concrete model on one route.
- One model can run on several routes, so Claude Sonnet 5.5 can sit in one slot on Runware and in another on OpenRouter. The same model on the same route is greyed out in the other slots as Already in the lineup.
- Models that need a file to start from are left out, since a round starts from a prompt alone: image upscalers, Runware’s vectorizer, and video models that edit or upscale a clip.
- The lineup is saved with the comparison and applies to the next round; earlier rounds keep the models they ran. While a round runs, the slots are locked: “The lineup can change once this round finishes.”
1.4
Type and Audio type
Type is the first field of the Models popover: Text, Image, Audio or Video. An audio comparison also picks an Audio type: Text to Speech, Music or Sound Effects, and the slots then list only that kind of model. Transcription models never appear, since they listen rather than generate.
Changing either one empties the lineup back to two slots. Once a comparison has a round, both are locked, and hovering says why: “This comparison already has rounds. Start a new comparison to compare another type.”
Part 2
Run rounds
2.1
How a round runs
1
Run
Press Run or
Enter (Shift+Enter adds a line). Run stays greyed out until every slot holds a model, no two slots repeat a model on one route and a prompt is typed; its tooltip names what is missing.2
Watch the cards fill
Cloud models start together. Local models (Ollama, Hugging Face and GGML) run one after another so each gets your machine to itself, and a card that is queued reads “Waiting for the previous local model”. Text streams in as it is written; image, audio and video cards read “Generating…” until the file is saved.
3
Stop if you need to
Stop or
Esc ends the round at once: text that already arrived stays on its card marked “Stopped”, and models still waiting are skipped.- Every model gets your prompt exactly as typed. Your instructions under Account are not added, and each model runs on its own defaults, so the difference you see comes from the models.
- There is no Auto fallback. If a model can’t finish, its card says so (“This model couldn’t finish.” or the provider’s message) and the rest of the round stands.
- A round keeps running when you open another comparison and saves its results there. Quitting mid-round saves what each card had as stopped.
2.2
Shared settings
The next round’s settings sit along the bottom of the prompt box, each select naming itself. They are saved with the comparison and copied onto every round you run.
- Aspect ratio (image, default 1:1; video, default 16:9). Each model gets the closest shape it supports, so the cards line up; a model with no shape control keeps its own.
- Duration (video, music and sound effects): each model’s default, or a target length that each model meets as closely as it can: 4 to 10 seconds for video, 10 to 120 for music, 3 to 20 for sound effects. Speech models read the whole script on their default voice.
- Any setting every chosen model has, with the values they all accept, such as Format for Nano Banana 2.1, Flux 2 Pro and Seedream 5 Pro. Each opens on “model default”, which sends nothing; a value you pick is sent only while the lineup still shares it.
- While a round runs the settings are disabled: “Available once this round finishes”.
2.3
Stepping through rounds
One round fills the screen at a time, the newest by default and after every Run. Step through the others with the pager in the header (“Round 2 of 3”) or ← / → when no field, menu or player has focus; in Arabic and Hebrew the arrows follow the mirrored pager.
TipStepping to a round puts its prompt and settings back in the prompt box. Edit a word and press Run, and the tweak runs as a new round you can compare with the old one.
Part 3
Read the results
3.1
Reading a card
- 1 · Model and route
- The model and the route it ran on, such as Runware, a vendor key or Ollama.
- 2 · Copy and Open in Text
- Text results only. Open in Text saves the answer as a new document in your project and opens it in the Text workspace.
- 3 · Pick
- Marks this result as the round’s pick.
- 4 · Stats
- Time from Run to the finished result, the first output for text, input and output tokens, and the cost.
The cost reads billed when the provider reported it (Runware, OpenRouter and xAI), estimated from the model’s published rates on other routes, Free, ran locally for local models and No price listed when there is no rate to go by. Image, audio and video cards play or show the file and put Show in folder in the footer’s corner.
3.2
Picking a result
Press Pick on the result you like best. A round has one pick: picking another card moves it, and pressing the check again (“Remove pick”) clears it. Pick waits for the result to finish, and the picks stay with the comparison, so you can step back later and see which model won each round.
3.3
Blind rounds
1
Turn on Blind
Press the eye in the settings row under the prompt: “Hide model names and shuffle the results until you pick one”. It applies to the next round you run.
2
Judge the results
The round stores its cards in a shuffled order and names them Model A, Model B and so on, without routes. Show in folder stays off: “Show the models first: the file name includes the model”.
3
Reveal
Pick a card and the names come back for good, even if you remove the pick later. Show models beside the pager reveals them without a pick, for as long as you stay on the comparison.
3.4
Where the files go
Image, audio and video results are saved in the project folder that was open when the round ran, in a folder named after the comparison, as round-1-flux-2-pro-product-still-life-of-a-brushed.webp: the round, the model and the first six words of the prompt. Text answers are stored with the comparison instead.
- The workspaces list your project’s top level only, so these folders do not show in the Image, Audio or Video sidebar. Use Show in folder on a card to get to a file.
- Renaming the comparison renames its folder once no round is running. Deleting the comparison leaves the files where they are.
- A file is recorded on its card as soon as it is saved, so it survives a closed window. A round that saved nothing leaves no empty folder behind.
- A comparison started in another project still shows its results: each card remembers the full path of its file.
Part 4
Recipes
4.1
Choose a writing model
1
Line up the candidates
Set the Type to Text and fill three slots: the model you use today, a cheaper one, and a local one if you have it installed.
2
Use a real prompt
Paste a brief from your actual work rather than a toy question, and run it. Read the answers, then the time and cost lines.
3
Tweak, rerun, decide
Change one thing in the prompt and run again;
← takes you back to compare. Make the winner your default under Models → Defaults.4.2
One model, two routes
1
Add the same model twice
Put Claude Sonnet 5.5 on Runware in one slot and on OpenRouter in another, and a third on its vendor key if you have one.
2
Run a few short rounds
Answers vary from run to run, so compare time and first output over three rounds rather than one.
3
Read the cost line
Both gateways report what they billed, so the cards show the real price of the same answer on each route.
4.3
Is a local model enough?
1
Pair local with cloud
Put one or two local models (Gemma 4 12B on Ollama, say) next to the cloud model you pay for. Local models take turns, so more of them make the round longer.
2
Turn on Blind
With the names hidden you judge the answers, not the price tag.
3
Run your everyday prompts
Pick a winner each round, then step back through the picks. If the local model keeps winning, it is free and never leaves your machine.
4.4
An image shoot-out
1
Set the shape first
Set the Type to Image, fill three or four slots, and pick the aspect ratio your final image needs, so every model renders it.
2
Run and compare
Settings the lineup shares, such as Format, appear beside the aspect ratio; leave them on model default unless you need one.
3
Pick and take it
Pick the winner and press Show in folder to get the file, already saved in your project.
4.5
Judge voices blind
1
Choose speech models
Set the Type to Audio and the Audio type to Text to Speech, then add three voices’ worth of models. Each reads on its default voice.
2
Hide the names
Turn on Blind, paste your script and run it.
3
Listen, then pick
Play Model A, B and C in turn and pick the one you would ship; the names come back with the pick.
4.6
Try video models for less
1
Keep it short
Set the Type to Video, the Duration to 4 s and the aspect ratio of your final cut. Video is priced by the second, so short test clips cost less.
2
Run once and read the costs
Video takes a while; open another comparison while it runs. Each card shows what its clip cost.
3
Make the real clip
Generate the full-length clip with the winner in the Video workspace.
Part 5
Reference
5.1
Keyboard shortcuts
⌘ is Ctrl on Windows and Linux. The main guide lists every workspace’s shortcuts together.
Prompt
Enter / Shift+Enter- Run the prompt on every model / add a line
Esc- Stop the running round
Comparisons
⌘N- New comparison
← / →- Previous or next round
Delete / Backspace- Delete the selected comparison (asks first)
⌘+ / ⌘- / ⌘0- Zoom the interface in, out, or back to 100%
5.2
Troubleshooting
- Run is greyed out. Hover it: “Choose a model for every slot”, “Choose a different model for each slot” (one model twice on the same route) or “Type a prompt first”. A license that needs attention is named there too.
- A model is missing from the slot list. Add its provider’s key or install it locally first. Models that need a source file (upscalers, the vectorizer, video edit and upscale models) and transcription models are never offered.
- The Type is greyed out. The comparison already has rounds; start a new comparison to compare another type.
- “This model couldn’t finish.” Arena never swaps in another model. Check that provider’s key under Models → Providers, or put another model in the slot for the next round.
- “Waiting for the previous local model” for a long time. Local models run one at a time; the card starts when the one before it finishes.
- A card says “Stopped”. The round was stopped, or the app quit while it ran. The text that had arrived is kept; run the round again for a full answer.
- Show in folder is greyed out. The round is blind and still hidden; pick a card or press Show models.
- “This file was moved or deleted.” The result file is no longer where the round saved it. Move it back into the comparison’s folder, or run the round again.
- I can’t find the results in the Image workspace. The workspaces list the project’s top level only; the comparison’s folder sits one level down. Use Show in folder.
- An image came back in another shape. The model doesn’t offer that aspect ratio, so it got the closest one it has, or it has no shape setting at all.
- “This comparison type needs a newer version of CSuite.” A newer build made this comparison. Its rounds still show; update the app to run new ones.
- “Couldn’t load this comparison.” Press Retry; if it keeps failing, restart the app.