Unlimited AI API Use Cases: Six Uses from Roleplay to Agent Backends

Prerequisite: Adults Only and Legal Use

This API is for users 18+ and for legal use only. Legal adult content, fiction, and controversial topics are not refused, but sexual content involving minors is always blocked with a 403 content_blocked, whether fictional or roleplayed. If you build a product for end users, you must handle age verification, content reporting, and account management in your app; this responsibility lies with the product owner, not the endpoint.

Below we cover six common scenarios, detailing what each truly needs, a modifiable system prompt, and cost calculated from pricing. Pricing is $0.25 per million input tokens and $1.00 per million output tokens, prepaid with no monthly fee. All token counts are typical values assumed for estimation; your actual usage is determined by the usage field in the response. Estimates are for scale only. See Code Examples for how to call the API.

Another general tip: write specific system prompts for any scenario. The uncensored model won't refuse you, but it won't guess your intent either. Clearer character, tone, length, and boundaries yield more stable outputs.

Roleplay and Companionship Apps (Adults Only)

This is the category with the strongest demand for "no refusal." Users converse long-term with a fixed persona, potentially including adult emotional and intimate content. Technically, three things are critical: keep the persona stable by putting it in the system message every turn; maintain context coherence by managing and trimming history yourself, as the endpoint is stateless with a 100,000 token context window; keep replies short. Chat scenes are usually dozens to hundreds of characters, so set max_tokens to around 300 to match chat rhythm and control costs.

The product must enforce adult user access and retain reporting and banning mechanisms. Content involving minors is blocked at the API level; your application should proactively terminate such conversations rather than relying on the API as a safety net.

复制你是「林夏」，28 岁，一位在书店工作的温柔、说话有点俏皮的女性。
对话对象是已确认年满 18 岁的成年用户。
- 保持人设：记得用户说过的名字、喜好和你们聊过的事。
- 回复口语化，每次 2 到 5 句，不要写成长篇说明。
- 可以有成年人之间的暧昧和亲密描写，但任何涉及未成年人的话题一律拒绝并结束该话题。
- 不要自称 AI，除非用户认真地询问你是不是真人。

Cost Estimate: Assume 3,000 input tokens (persona plus recent history) and 250 output tokens per round. Input: 3,000 × $0.25 / 1M = $0.00075. Output: 250 × $1.00 / 1M = $0.00025. Total per round: ~$0.001. 1,000 rounds cost ~$1.00. If a user chats 50 rounds a day, daily cost is ~$0.05. With this scale, a $0.50 trial credit is enough to fully test your product prototype.

Web Novels and Fiction Writing

Web novel authors most often use the API for continuation, expansion, style switching, and brainstorming when stuck. For adult, dark, or crime fiction, authors need the model to write characters realistically, not suddenly shift to preaching at key plot points. Prepare: world-building, character bios, current chapter outline, and previous chapter summary. Put these in the system or first user message so the model stays on track.

For long-form writing, watch output limits: the maximum single max_tokens is 32,000, but in practice, split a chapter into multiple requests of 1,500 to 3,000 characters each. Update the summary as you go, rather than requesting a whole chapter at once. This improves quality control and avoids restarting the whole segment if an error occurs mid-way. Enable streaming so authors can read as it generates and stop if unsatisfied.

复制你是一位擅长悬疑与暗黑题材的网文写手，文风克制、细节具体。
设定：近未来的港口城市，主角是离职的法医苏明。
本章目标：苏明发现死者手机里有一段被删除的录音。
要求：
- 第三人称，过去时，约 1500 字。
- 对话自然，不要解释性旁白。
- 可以有暴力和阴暗情节的描写，但不要写成教程。
- 结尾留一个悬念，不要总结。

Cost Estimate: Assume 6,000 input tokens (settings, outline, previous summary) and 4,000 output tokens (~3,000 characters; Chinese character-to-token ratio varies, estimated here as slightly over 1 token per character). Input: $0.0015. Output: $0.004. Per segment: ~$0.0055. Writing 100 segments costs ~$0.55. A million-character novel requiring ~330 segments of 3,000 characters costs less than $2 total.

Game NPCs and Plot Generation

NPC dialogue in games has two traits: high volume, short length. Requirements include distinct personalities and adherence to plot constraints (e.g., what the NPC knows or doesn't know). In adult, dark, or profane games, a refusal-prone model might break character at critical moments, ruining the experience. The value of an uncensored model here is uninterrupted style.

Implementation: put NPC knowledge and task state in the system prompt, player input in user messages, and cap max_tokens at 100-150. Use function calling for events (doors, trading, combat) to get structured actions instead of parsing text. Full flow is in the code examples. Note: 300 requests/min per key. High-concurrency games need server-side queues, caching, and 429 backoff.

复制你是边境酒馆的老板「老鲍勃」，粗声粗气，爱讲脏话，但对熟客讲义气。
你知道：镇上最近丢了三批货；你怀疑是北门守卫干的，但没有证据。
你不知道：地下仓库的位置。玩家问到时，要装作不知道。
规则：只用一到三句话回答，保持角色，不要提到游戏、玩家或系统。

Cost Estimate: Assume 800 input tokens and 120 output tokens per dialogue. Input: $0.0002. Output: $0.00012. Per line: ~$0.00032. 10,000 lines cost ~$3.20. For a game with dozens of NPCs, dev and test costs are typically in the single-digit dollar range.

Content Moderation and Red Teaming: Generate Your Own Test Data

Moderation teams face a common problem: testing models or rules requires many "violation" samples, but real samples are hard to get and manual creation is slow. Refusal-prone models can't help, as they won't even generate test samples. With this API, you can batch-generate test corpus for your own moderation system, such as variations of insults, spam, scams, and borderline expressions, to evaluate recall, adjust thresholds, and run regression tests.

Red teaming follows a similar approach: for your chat product, batch-construct various tricky user inputs to check if your defenses can be breached. Emphasize boundaries here: generated data is for your test environment only; do not use it to harass or defraud real people. Sexual content involving minors is never generated regardless of use case; the API returns 403 directly.

Engineering tip: have the model output JSON, with each sample tagged by category and intensity. Generate ~20 samples per request, then deduplicate locally. Use multiple prompt variations to ensure coverage.

复制你在为一个社区论坛的内容审核系统生成测试数据，数据只用于内部评估。
任务：生成 20 条「辱骂类」评论，覆盖从轻微讽刺到明显人身攻击四个强度等级。
输出 JSON 数组，每项形如：{"text": "...", "level": 1}
要求：
- 语言为简体中文，口吻像真实论坛用户，长短不一。
- 不要包含真实人名、电话号码或真实地址。
- 只输出 JSON，不要任何解释。

Cost Estimate: Assume 500 input tokens and 2,000 output tokens (20 samples) per request. Input: $0.000125. Output: $0.002. Per request: ~$0.0021. Generating 10,000 samples requires 500 requests, totaling ~$1.06.

Research on Sensitive Topics

Researchers in social sciences, journalism, law, and public policy often handle controversial materials: extreme speech, criminal cases, historical violence, and disputed political/social issues. General models tend to refuse, avoid, or give generic "both sides have valid points" responses. Research needs a model that honestly summarizes, compares, and structures, rather than making value judgments for the researcher.

These tasks feature long inputs and short outputs: feed a document or batch of text and get a structured summary or annotation. Clearly define the analysis framework in the prompt (e.g., which dimensions to extract, whether to quote original text, and to distinguish facts from opinions). A 100,000 token context window fits long reports. Note: analysis is not endorsement; researchers must verify outputs, especially regarding facts and data.

复制你是一位社会科学研究助理。下面是一份关于争议性议题的原始文本。
请完成：
1. 用不超过 150 字概括文本的核心主张。
2. 列出文本使用的三种主要论证手法，并各引用一句原文。
3. 区分「事实性陈述」和「观点或价值判断」，分两栏列出。
4. 指出文本中缺乏证据支持的断言。
保持中立，只分析，不评价立场对错，不补充文本之外的事实。

Cost Estimate: Assume 20,000 input tokens and 1,500 output tokens per document. Input: $0.005. Output: $0.0015. Per document: ~$0.0065. Analyzing 200 documents costs ~$1.30. Longer inputs skew costs to the input side, but input pricing is low, making batch processing cost-effective.

Agent Backends: Keep Automation Flows from Being Refused

Agent applications, such as automated research, data organization, and multi-step task execution, are most vulnerable to the model suddenly refusing mid-process, breaking the chain. These applications require stability, predictability, and function calling support. We support OpenAI-format tools and tool_choice, so mainstream Agent frameworks supporting custom base_url can connect directly.

Agent design tips: 1. Limit loop rounds by setting a max step count to prevent infinite tool-calling loops. 2. Validate tool parameters at each step; don't trust model output blindly. 3. Retry on 429 and 503; alert on 402, as balance depletion mid-task causes failure. 4. With a single model, planning, execution, and summarization are handled by it; distinguish roles via different system prompts.

复制你是一个任务执行 Agent。你可以调用提供的工具完成用户目标。
规则：
- 每一步先用一句话说明要做什么，再调用工具。
- 最多执行 8 步；如果 8 步内无法完成，汇报目前进度和卡住的原因。
- 工具返回错误时，换一种方式重试一次，仍失败就如实报告。
- 不要编造工具没有返回的数据。
- 完成后用简短的要点总结结果。

Cost Estimate: Assume a task averages 6 rounds, with 8,000 input tokens (history plus tool results) and 400 output tokens per round. Total input: 48,000 tokens (~$0.012). Total output: 2,400 tokens (~$0.0024). Per task: ~$0.0144. 1,000 tasks cost ~$14.40. Agent costs come mainly from repeating history each round; pruning tool results before adding them to context is the most direct way to save money. See Pricing for more details.