About and FAQ
What this is
Chat with All sends one question to up to four AI models at the same time and streams their answers side by side. Every lane runs the same system prompt, so the only thing that differs between the answers is the model itself - a controlled comparison rather than a comparison of products. Each lane is its own server-kept conversation, so a follow-up continues where that model left off.
What a comparison costs
One question across four lanes is four metered turns. The meter above the composer shows the total worst-case hold before you send, checked against your balance, and each answer is stamped with its settled cost once the run finishes. The publisher takes no markup on this app, so turns bill at platform cost. One US dollar is 10,000 credits. Creating a session, listing models and estimating a cost are all free.
Which models you can compare
The picker is built from the platform's live model list and shows only models the deployment currently reports as available, each labelled with its real billed rate per million output tokens. Anthropic (Claude) models are unavailable on this deployment at present, so they are filtered out of the picker rather than offered as a lane that would fail after taking your hold. You do not need an OpenAI or Anthropic API key: every lane runs through SkillSafe on your credit balance, with no per-provider keys, no separate billing accounts and no rate limits to manage yourself.
The readout: where the answers agree and differ
Under the lanes, every finished comparison gets a readout built entirely in your browser from replies you have already paid for. It names the words every answer shares, the words only one answer uses, and the fastest, cheapest and longest lane of the run. It is a word-overlap reading aid that points at where to look, not a verdict from a model - no extra turn is run to produce it, and it works on restored history when you are signed out. The same readout is included in the copied and downloaded Markdown.
Follow-ups do not have to bill every lane
Each lane header has an Ask checkbox, and every answer has a "Follow up here only" button that unticks the rest. Drilling into the answer you liked costs one turn instead of four, and "Ask all" re-ticks every lane when you want the full comparison back. Because a follow-up bills the conversation the model is holding as well as your new message, the real hold on a second turn in the same lane is higher than the meter's figure; the meter says so once a lane has history.
Your comparisons are saved
Each comparison is written to your SkillSafe account's key-value store and mirrored into this browser, so it survives a reload and follows your account to another device. You can reopen, copy or download any past comparison as Markdown from the History panel without spending anything. A guest identity's history is separate from your signed-in account's, because every guest session is a new subject.
When a model fails
Whatever already streamed is kept on screen and marked as a partial answer. A worst-case hold is placed when a turn starts and the unused part is refunded at settlement, so you are billed for the output the model actually produced and not for the reserve. A failed lane can be retried on its own, which bills one turn rather than re-asking all four.
Signing in
You can open the app, choose a lineup and see the live cost estimate as a guest, but sending a question needs a signed-in SkillSafe account with credits, because each turn is metered. Sign-in opens in a popup so your half-typed question is not lost.
Driving this app over HTTP is documented in the API tutorial; the token this browser uses is managed on the tokens page.