Choosing a model
A decision guide for daily chat, long context, reasoning and free use.
New chats start on DeepSeek V4 Flash at 0.3×. Switch when you have a concrete reason — a bigger context window, images, visible reasoning, or a smaller bill. The full lineup with exact numbers is on Models.
Pick by what you need
| You want | Use | Cost | Trade-off |
|---|---|---|---|
| A good default for roleplay | DeepSeek V4 Flash | 0.3× | None — nothing cheaper supports plugins or lore lookup |
| The same price, a bigger window | MiMo V2.5 | 0.3× | 262,144 tokens instead of 163,840 |
| To spend nothing | Llama 3.1 8B | 0× | 15 messages per 3 hours, and no plugins, lore lookup or memory writes. Free models |
| A very long chat, or a large world book | GLM 5.3 Flash | 0.5× | 1,048,576 tokens, but reasoning is always on and can't be turned off |
| A million-token window without forced reasoning | DeepSeek V4.1 Flash | 0.7× | Slightly pricier than GLM 5.3 Flash |
| The model to read images you attach | GLM 5.3 Flash (0.5×), DeepSeek V4.1 Flash (0.7×) or Gemini 3 Flash Preview (1.2×) | 0.5×–1.2× | Gemini 3 Flash Preview needs a subscription |
| To watch the model think | any model marked Optional, with reasoning switched on | unchanged | Thinking tokens are billed like reply tokens |
| Reasoning that never switches off | DeepSeek R1 | 1.0× | No tool support: no plugins, lore lookup or memory writes |
| The advanced tier | Gemini 3 Flash Preview (1.2×) or GLM 5 (1.3×) | 1.2×–1.3× | Requires an active subscription |
A big context window is not a paid feature: two basic-tier models hold a million tokens.
What a turn actually costs
A turn is billed on everything sent plus the reply. A mature chat — character card, persona, world-book entries, history and a full reply — averages about 8,000 tokens:
| Multiplier | Models | Credits for an 8,000-token turn |
|---|---|---|
| 0× | Llama 3.1 8B | 0 |
| 0.3× | DeepSeek V4 Flash, MiMo V2.5 | ~2,400 |
| 0.5× | DeepSeek V3.2, GLM 5.3 Flash | ~4,000 |
| 0.7× | DeepSeek V4.1 Flash, DeepSeek V4 Pro | ~5,600 |
| 1.0× | DeepSeek R1, GLM 4.7 | ~8,000 |
| 1.2× | Gemini 3 Flash Preview | ~9,600 |
| 1.3× | GLM 5 | ~10,400 |
For scale: a Pro plan's 2,600,000 monthly credits is roughly 1,000 turns on the default model, or about 250 on GLM 5. Short early messages cost far less than this; the number grows as history accumulates. Credits
Reading the picker's stats
Choose AI Model shows community data next to each model. It is the only quality signal in the product — Reverie does not publish its own ranking.
| Stat | How it reads | What it means | Appears after |
|---|---|---|---|
| Quality | <n>% preferred | Share of blind side-by-side picks that went to this model | 30 decided comparisons |
| Confidence | 1–5 | How much data stands behind that percentage | — |
| Head-to-head | Head-to-head (<n>) | This model's record against one specific other model | 5 comparisons for that pair |
| Like rate | a percentage | Share of replies given a thumbs up rather than a thumbs down | 3 rated replies |
| Speed | TTFT <time> · <rate> tok/s | Time to the first token, then tokens per second | — |
Below those thresholds the picker says Collecting data, No data yet or Just released instead of a number. A model with few samples isn't worse — it's newer.
The comparison data comes from the card that offers two replies to your first message in a chat; whichever you keep is recorded as a vote. Comparing two replies. Like rates come from the thumbs up and thumbs down on individual replies.
When a different model isn't the answer
Reply length, point of view and pacing are settings, not model traits. Before paying more per turn, try Settings → Replies & Generation in the side panel for a standing change, or a slash command for one turn only. Repetition and drift in a long chat are usually a context problem — a bigger window or tidier memory helps more than a pricier model.
Related
Models
The full table: multipliers, context windows and capabilities.
Free models
The 0× model, its quota and its limitations.
Comparing two replies
Where the picker's preference data comes from.
Reasoning mode
Which models think, and how to see it.
Subscriptions
What a plan includes, including the advanced models.