splitscreen rooms can ask Claude, GPT or Gemini. They are not the same, and the differences are the kind you only find out about halfway through a job. Here they are in one place, written out rather than ticked off.
| Feature | ClaudeAnthropic key | GPTOpenAI key | GeminiGoogle key |
|---|---|---|---|
| Search the webIt looks things up while it answers, and shows what it read. | YesThe oldest Claude models search with an earlier, plainer version of it. | Yes | Yes, on most modelsNot on Gemini 3.1 Flash Image or Gemini 3 Pro Image. Those two make pictures and take no tools. |
| Read links you pastePaste a link and it opens that page instead of guessing at it. | Yes, on most modelsNot on Haiku 4.5. It rides the same switch as web search - there is no second chip. | YesGPT reads a pasted link through the same web search tool. | Yes, on most modelsNot on Gemini 3.1 Flash Image or Gemini 3 Pro Image. Those two make pictures and take no tools. |
| Run code in a sandboxIt runs commands and edits files on the provider's own machines - never on yours. | Yes, on most modelsNot on Haiku 4.5. With the sandbox on, files you attach go into it too. | Yes | Yes, on most modelsNot on Gemini 3.1 Flash Image or Gemini 3 Pro Image. Those two make pictures and take no tools. |
| Read files you attachDrop in an image or a document and it reads it. | YesImages and PDFs alike. Code and text files are read as text. A .zip goes into the code sandbox, where it is unzipped and explored. | YesImages are read. A PDF is passed as a note naming the file, not its contents. Code and text files are read as text. A .zip goes into the code sandbox, where it is unzipped and explored. | YesImages and PDFs alike. Code and text files are read as text. A .zip cannot be opened here - its sandbox takes no files. |
| Make imagesIt draws a picture and the room can open it. | NoClaude has no image generation. | YesIt is a tool on the ask, and it carries a per-image fee. | YesImage on a Gemini ask runs on Gemini 3.1 Flash Image, and turns search and the sandbox off for that ask. |
| Choose how hard it thinksA switch in the composer: think lightly, normally, or as hard as it can. | Yes, on most modelsNot on Haiku 4.5. The composer calls it Light, Standard and Max. | Yes | Yes, on most modelsNot on Gemini 3.1 Flash Image or Gemini 3 Pro Image. Those two make pictures and take no tools. |
| How much it can read at onceHow much conversation, or how long a document, it can hold in one go. | YesSizes differ by model - each one is in the table below. | YesSizes differ by model - each one is in the table below. | YesSizes differ by model - each one is in the table below. |
Every answer above is read from the same table the app itself uses to build a request, so this page cannot claim something the room will not do.
The same list the picker shows, in the same order, with the same tags. Older models stay on the list and are marked, because a room already pinned to one should still be able to see what it is.
The default. Strongest all-round reasoning and long code.
Anthropic's strongest — deepest reasoning, cheapest cache reads.
Fable 5.1 replaces it. Same price, a step behind.
The previous Opus. Same price, a step behind.
Quick and capable for everyday work, 40% of Opus.
Cheapest and fastest. No sandbox, basic web search.
OpenAI's strongest — search, sandbox and images.
Everyday GPT at half the price of Sol.
Cheapest GPT — quick lookups and drafts.
The default Gemini. Search, sandbox, huge context.
Deepest Gemini. Preview, and dearer past 200k tokens.
Previous Flash. Same price, a step behind.
Older Flash, kept for rooms that pinned it.
Older and dearer than 3.8.
The previous Lite. A little dearer than 3.1.
Cheapest and fastest Gemini for quick lookups.
Makes images (Nano Banana 2). No search or sandbox.
Best images (Nano Banana Pro), up to 4K. No tools.
"Last checked" is a real ask we paid for and watched work, most recently on 3 Sept 2026.
Providers do not charge by the question, they charge by the piece of text. A token is about three quarters of an English word - so roughly 750 words is 1,000 tokens. You pay for what you send (everything in the room the model has to read, every time) and for what it writes back, at different rates. That is why a long conversation costs more per reply than a short one, and it is why splitscreen shows you the price of one typical reply instead of the rates: nobody can do this arithmetic in their head, and nobody should have to.
| Model | What you send | What it writes | One reply |
|---|---|---|---|
| Opus 5 · Claude | $6.50 / million | $32.50 / million | about 2.3¢ a reply |
| Fable 5.1 · Claude | $13 / million | $65 / million | about 4.6¢ a reply |
| Fable 5 · Claude | $13 / million | $65 / million | about 4.6¢ a reply |
| Opus 4.8 · Claude | $6.50 / million | $32.50 / million | about 2.3¢ a reply |
| Sonnet 5 · Claude | $2.60 / million | $13 / million | about 0.9¢ a reply |
| Haiku 4.5 · Claude | $1.30 / million | $6.50 / million | about 0.5¢ a reply |
| GPT-5.6 Sol · GPT | $5.20 / million | $26 / million | about 1.8¢ a reply |
| GPT-5.6 Terra · GPT | $2.60 / million | $15.60 / million | about 1¢ a reply |
| GPT-5.6 Luna · GPT | $0.26 / million | $1.56 / million | about 0.1¢ a reply |
| Gemini 3.8 Flash · Gemini | $0.97 / million | $4.88 / million | about 0.3¢ a reply |
| Gemini 3.1 Pro · Gemini | $2.60 / million | $15.60 / million | about 1¢ a reply |
| Gemini 3.7 Flash · Gemini | $0.97 / million | $4.88 / million | about 0.3¢ a reply |
| Gemini 3.6 Flash · Gemini | $0.97 / million | $4.88 / million | about 0.3¢ a reply |
| Gemini 3.5 Flash · Gemini | $1.95 / million | $11.70 / million | about 0.8¢ a reply |
| Gemini 3.5 Flash-Lite · Gemini | $0.39 / million | $3.25 / million | about 0.2¢ a reply |
| Gemini 3.1 Flash-Lite · Gemini | $0.33 / million | $1.95 / million | about 0.1¢ a reply |
| Gemini 3.1 Flash Image · Gemini | $0.65 / million | $78 / million | about 4¢ a reply |
| Gemini 3 Pro Image · Gemini | $2.60 / million | $156 / million | about 8.1¢ a reply |
Prices shown are our credit prices. On your own API key you pay your provider’s rate, with no markup from us.
“One reply” is one fixed comparison - about 750 words in and 375 words back - priced on every model so the models can be compared. It is not a promise about your replies, which will be longer or shorter than that.