DeepSeek and roleplay
Using DeepSeek for roleplay on a phone: setup, off-peak pricing, and cache hits
DeepSeek works for roleplay through its API, though DeepSeek’s own mobile app is described as a general AI assistant, with no mention of character cards or lorebooks. On a phone, you paste a DeepSeek API key into a dedicated roleplay app like Foreverse (Android and iPhone). As of 2026-11-15, deepseek-flash lists at $0.30 per million input tokens at peak hours ($0.15 off-peak) and bills cache hits at 2% of the full input price—roughly $0.46 per 100 replies uncached, or $0.09 at 90% cache hits.
Available for iPhone and iPad in selected regions outside mainland China. Not available on the mainland China App Store. Download help
Can DeepSeek do roleplay?
The model handles roleplay constantly; DeepSeek’s official app just isn’t built for it. On Google Play, DeepSeek sits under Productivity as a general-purpose assistant, and its listing mentions no character cards, personas, or lorebooks. If you want a character to hold a persona across a long scene with greetings, sample dialogue, and triggered world lore, you bring DeepSeek’s API into a tavern-style client instead.
Connecting it is straightforward because DeepSeek speaks both the OpenAI and Anthropic API formats. Any client that accepts a custom OpenAI endpoint works, and apps with a built-in DeepSeek preset—Foreverse included—only ask for your API key. You top up directly on DeepSeek’s platform, and your phone sends prompts straight to DeepSeek. One honest gap: DeepSeek’s terms of use wouldn’t load when we checked today, so we won’t guess what they restrict. Read their terms directly, and see our provider policy breakdown for how model providers govern adult fiction.
Is DeepSeek Flash good for roleplay?
By sheer volume, roleplayers use it heavily. In OpenRouter’s Roleplay collection (ranked by rolling seven-day tokens, November 2026), DeepSeek V4.1 Flash 1017 held the #1 spot at roughly 525B tokens—23.4% of the entire category. It also ranked second on OpenRouter’s SillyTavern monthly chart, while DeepSeek V3.2 topped Janitor AI’s chart at 144B tokens. Keep OpenRouter’s own disclaimer in mind: token charts measure adoption, not prose quality.
For blind quality testing, our closest data point is our July 2026 Fiction Bench, where nine models wrote 20-round Chinese novel continuations for blind judges. DeepSeek V4 Flash took first place on traditional xuanhuan fantasy, dropped to 6th–7th out of nine on formal court romance, and logged zero repetition loops across 20 rounds. That tested long-form Chinese fiction rather than interactive chat, and DeepSeek has since pointed deepseek-flash to the newer DeepSeek-V4.1-Flash-1017—so treat it as a clue on style, then run a few turns on your own cards.
Quick latency tip: DeepSeek’s docs note that deepseek-flash ships with thinking mode enabled by default, alongside a non-thinking toggle. If every reply pauses to reason before speaking, check that setting first.
What does DeepSeek cost for roleplay at list price?
Checked against DeepSeek’s pricing docs on 2026-11-15 (per million tokens). Off-peak hours run at half price—every hour except 01:00–04:00 UTC and 06:00–10:00 UTC Monday through Friday, with weekends and Chinese public holidays billed off-peak all day. Both models carry a 1M-token context window. Our 100-reply columns assume a 2,000-token card + system prompt, 120 user tokens and 350 reply tokens per turn, and history capped at 16,000 tokens (~1.40M input tokens and 35,000 output tokens total).
| Model & window | Input (cache hit / miss) | Output | 100 replies, uncached | 100 replies, 90% cached |
|---|---|---|---|---|
| deepseek-flash, peak | $0.006 / $0.30 | $1.20 | $0.46 | $0.09 |
| deepseek-flash, off-peak | $0.003 / $0.15 | $0.60 | $0.23 | $0.05 |
| deepseek-v4-pro, peak | $0.044 / $1.32 | $3.96 | $1.98 | $0.38 |
| deepseek-v4-pro, off-peak | $0.022 / $0.66 | $1.98 | $0.99 | $0.19 |
DeepSeek notes that rates are subject to change. Curious how these totals compare with monthly subscriptions? See roleplay apps without a subscription.
Why do cache hits matter so much on DeepSeek?
Roleplay apps resend your card and conversation history on every single turn, so most of what you pay for is text the model has already seen. DeepSeek turns prefix caching on automatically: whenever the start of a prompt matches a cached prefix, those tokens bill at $0.006 per million instead of $0.30 on peak deepseek-flash—just 2% of full price. Caching is best-effort with no guaranteed hit rate, but every API response returns prompt_cache_hit_tokens and prompt_cache_miss_tokens so you can verify your hit rate directly.
Mutating the top of your prompt wrecks that discount. In our July 2026 test using identical scripts, models, and keys over 15 rounds, stock SillyTavern hit 91.5% cache reuse ($0.0241 total) while an earlier build of Foreverse hit only 46.5% ($0.0664—2.8x pricier) because triggered lorebook entries kept shifting the start of the prompt. We published that bug against ourselves and fixed it; read the breakdown in Prefix caching and what your chats really cost (note that July’s report used July token rates, whereas the table above reflects 2026-11-15 pricing).
When DeepSeek is not the right pick
- You want character cards inside DeepSeek’s official app. That app is a general assistant without card or lorebook support.
- You mainly read ornate, etiquette-heavy period romance. In our July blind test, V4 Flash topped plain, fast-paced fantasy but landed 6th–7th of nine on court romance.
- Your client scrambles the start of the prompt every turn. DeepSeek’s cache is best-effort, and unstable prefixes forfeit the 98% input discount.
- You want locked-in flat pricing. DeepSeek can revise token rates at any time, and peak hours cost double off-peak hours.
FAQ
Can DeepSeek do roleplay?
Yes—in fact, DeepSeek V4.1 Flash led OpenRouter’s Roleplay collection by token volume in November 2026. Because DeepSeek’s own mobile app doesn’t read character cards, you’ll want to paste a DeepSeek API key into a dedicated roleplay client.
Is DeepSeek Flash good for roleplay?
We haven’t run a dedicated chat-RP benchmark. In our July 2026 Fiction Bench (20-round novel continuation), DeepSeek V4 Flash placed 1st on xuanhuan fantasy and 6th–7th of nine on court romance; deepseek-flash now maps to V4.1 Flash, so test a few turns on your own cards.
How much does DeepSeek cost for roleplay?
As of 2026-11-15, peak rates for deepseek-flash are $0.30/1M uncached input tokens and $1.20/1M output tokens, dropping by half during off-peak hours. In our 100-reply scenario, that works out to roughly $0.46 uncached or $0.09 at a 90% cache hit rate.
What are DeepSeek’s peak hours?
Peak pricing runs Monday through Friday from 01:00–04:00 UTC and 06:00–10:00 UTC, excluding Chinese public holidays. Every other hour—plus all weekend—bills at half-price off-peak rates.
Why is DeepSeek cheaper with a long chat?
Prefix cache hits on deepseek-flash cost just 2% of the standard input rate. Because long chats resend the same card and history every turn, most of that input can bill at the cache price, but the saving only appears if your app keeps the top of the prompt unchanged.
Sources
- DeepSeek API docs: Models & Pricing (checked 2026-11-15)
- DeepSeek API docs: Context Caching (checked 2026-11-15)
- DeepSeek API docs: Your First API Call (checked 2026-11-15)
- Google Play: DeepSeek - AI Assistant (checked 2026-11-15)
- OpenRouter: Roleplay model collection (checked 2026-11-15)
- OpenRouter: SillyTavern usage page (checked 2026-11-15)
- OpenRouter: Janitor AI usage page (checked 2026-11-15)