The short answer
Yes: you can use GPT, Claude, Gemini and any other model in the same conversation, switching between them per message from the model picker. The whole thread carries across the switch: your messages, the previous model's responses, and your attached files, so Claude can critique what GPT just wrote without you re-pasting anything.
Each turn is charged at the credit cost of the model that answers it, and the model must be included in your plan. The rest of this page is when switching pays off and when it does not.
What happens when you switch
The new model receives the full conversation so far, including your messages, the previous model's responses, and any files you attached. From its perspective it is joining a discussion already in progress, with everything visible.
That is why this is different from copying text between three browser tabs. When you paste into a new tool you bring the output but lose the reasoning, the constraints you established over five messages, and the document you uploaded. Switching in place keeps all of it.
Two things are worth knowing. The new model sees the previous model's output as part of the conversation, which can anchor it toward agreement. When you want a genuinely independent view, say so explicitly: Evaluate the answer above on its merits. I want your assessment, not a synthesis of what has already been said. And the per turn input budget tops out at 40,000 tokens, with a model whose context window is small getting proportionally less, so a very long thread is trimmed to fit whichever model you switch to; past that cap, a bigger context window does not send more of the thread. If a switch produces an answer that has forgotten the early part of the discussion, the context budget is usually why.
The four patterns worth learning
1. Phase routing. The workflow most people converge on. Different stages of one task go to different models.
Structure in GPT, prose in Claude, verification with web search turned on. It applies to writing a post, a proposal, a spec, or a report, and it works because these really are three different skills rather than three difficulty levels of the same one.
2. The second opinion. Get an answer, then switch and ask the new model to attack it.
A colleague proposed the answer above. Find what is wrong with it: factual errors, missed edge cases, a simpler approach, or an assumption that does not hold. If it is actually sound, say so plainly rather than inventing objections.
Two useful outcomes. Either a real flaw surfaces before you act on it, or the second model agrees despite being pushed to disagree, which is meaningful confirmation. Iterating with the same model produces neither, because models tend to agree with themselves.
3. The reframe. When an explanation does not land, switch rather than re-reading.
Claude and GPT explain the same concept with genuinely different framings, and the second framing is frequently the one that clicks. This is the single most underrated reason to have more than one model available, and it costs one click.
4. Capability routing. Switch because only one model can do the thing.
A 300 page document goes to a model with a million-token context window. An image goes to a model that can read images. A question about last week is the one case that needs no switch at all: web search in Whizi is a per-conversation toggle in the composer, it runs on the model you are already using, and turning it on costs nothing in credits. The rest is not preference, it is a hard constraint, and it is the case where a single-model subscription simply stops.
Which model for which phase
A starting point rather than a rule. Your own tasks should override it within a week of paying attention.
| Phase | Model | Why |
|---|---|---|
| Brainstorm, structure, outline | GPT | Fast, follows a brief tightly, clean hierarchies |
| Draft prose, tone, persuasion | Claude | Least filler, sustains a voice, best at difficult tone |
| Research, recent facts, sources | Any model, with the web search toggle on | Search is a per-conversation setting in the composer, not a model property |
| Long documents, large corpora | Gemini | 1M token context window on the current rows |
| Strict formats, tables, JSON | GPT | Most reliable at obeying an exact schema |
| Critique and red teaming | Anything that did not write it | Independence is the whole point |
The table names the three families most people arrive with, and the picker is wider than that. The open weight families each have their own reference page, with the credit cost per message and the plan that opens each row: Llama, Mistral, Kimi and GLM.
The one rule worth holding regardless of task: the model that critiques should never be the model that drafted.
Making the switch actually pay off
Switching helps most when you tell the new model what you want from it. Dropping in with no framing gets you a continuation of the same conversation in a slightly different voice.
Handing off a phase
We now have the outline above. Draft section 2 only. Use the evidence listed under it. Voice: [paste sample]. Do not restate the outline back to me.
Asking for independence
Ignore the framing established above and answer this from scratch: [restate the question]. I want a genuinely independent take, not a refinement of what is already here.
Changing the job entirely
Stop drafting. Switch to reviewing. Here is the criterion: [state it]. Report only problems, do not rewrite.
One further habit: when a switch produces something noticeably better, note which model and which phase. After a couple of weeks you will have a routing map that is specific to your work rather than to a benchmark, and that map is worth more than any published comparison.
When not to switch
Switching is cheap but not free, and there are cases where it makes things worse.
- Mid-draft, for style. Changing model halfway through a long piece produces a visible seam, because voice consistency is exactly what changes between models. Finish the draft, then switch for the critique.
- When you have not defined the task. Switching does not fix a vague prompt. If the first answer was unhelpful because the question was unclear, a different model gives you a different unhelpful answer.
- To find agreement. Cycling until a model tells you what you want is a way of laundering a decision you had already made. If two of three disagree, that is the finding.
- On a very long thread, into a smaller-context model. If the early context matters and the thread is huge, either stay with the large-context model or summarise the thread first and start fresh.
For running two models on the same prompt simultaneously rather than sequentially, see comparing models side by side.
- Assign a model to each phase of the task before you start
- Tell the new model what job it is taking over, rather than just continuing
- Ask explicitly for independence when you want a real second opinion
- Never let the model that drafted something also be the one that critiques it
- Switch when an explanation does not land, instead of re-reading the same one
- Finish a draft before switching, so the voice stays consistent
- Keep notes on which model wins which phase for your own work
Frequently asked questions
Does the next model see the full conversation?
Yes. Whizi passes the full conversation context forward by default, including your messages, previous model responses, and attached files, so the new model joins with everything visible. The one caveat is the per turn input budget, which tops out at 40,000 tokens and is proportionally smaller on models with small context windows: a very long thread is trimmed to fit. An answer that seems to have forgotten the early part of a long discussion is usually hitting that budget, not a property of the model you switched to.
Can I hide part of the thread from a specific model?
You can fork the thread or start a new chat with only the context you want to carry over, which is the right approach for sensitive snippets and also for long threads that have accumulated irrelevant detail. Starting clean with a short summary of what matters often produces better answers than carrying forty messages of history, since older context competes for attention with the actual question.
Will switching models mid-draft change the writing style?
Yes, noticeably, and that is a reason to finish a draft in one model before switching. Voice consistency is precisely what differs between models, so a handover halfway through a long piece leaves a visible seam. The productive pattern is to draft entirely in one model, then switch for the critique, the fact check, or a structured extraction.
Is there automatic model routing?
Yes, as a single picker row called Auto. It reads each message, sorts it into one of six fixed rungs, and answers on the model pinned to that rung, never above what your plan can open. There are no routing rules to write: nothing in Settings edits the ladder, weights a rung, or pins Auto to a model. Picking a model by hand stays a per-message decision, and the manual habit of choosing per phase tends to teach you more about which model actually suits your work.
Does switching count as extra messages?
Changing the model in the picker sends nothing, so nothing is charged until your next message. Each turn is charged at the credit cost of the model that answers it, so the price of the next answer changes with the switch: a message answered by a 1 credit model costs 1 credit and the same message answered by a 20 credit model costs 20; the credit scale lists every model's cost. Asking the same question of a second model is a second charged turn, which is worth it when the answer matters.