Voucher Usage Guide
What types of voucher does Novita provide?
Novita currently offers the following types of vouchers to help users experience the platform’s services:- New User Voucher: Available to first-time registered users for trying out services of various types of models.
- Referral Voucher: Earned by inviting others to register and complete tasks on the platform.
Rate Limits for Model Usage
What are the RPM limits for different users?
RPM (Requests Per Minute) limits vary based on a user’s verification level and account status. For detailed limits, please refer to the official documentation:RPM Adjustment and Upgrade Policy
Can users apply for increased RPM limits?
Yes. Novita allows flexible RPM upgrades based on usage needs. Rules are listed as below:- DeepSeek models: The platform will strive to accommodate reasonable scaling needs.
- Other models: RPM increases are evaluated based on model cost and actual user behavior, subject to capacity availability.
Application process:
User → Customer Manager / Tech Support → Product Team Review & ApprovalWhat happens if actual usage falls short of the committed RPM?
If the user’s actual RPM remains below the committed level for one consecutive week, the platform will adjust the rate limits policy as:- Reduce the limit to the peak RPM in the past week, or
- Revert to the model’s default rate limit (whichever is lower).
Is self-service RPM upgrade supported?
Yes. Novita plans to launch an RPM Upgrade Package, allowing users to manage and increase RPM limits independently, without manual approval. For further assistance, please book a call with our sales team.Billing
What is the billing priority when using the Coding Plan?
The system deducts from your Coding Plan quota first. Only after the Coding Plan quota is fully exhausted will charges be applied to your API account balance.How are the 50M tokens in the Coding Plan calculated?
Tokens are calculated based on base-rate equivalent token counts. Different models carry varying pricing weights proportional to their cost ratios. Both input tokens and output tokens count toward your total usage.How are cached tokens billed?
Cached read tokens are billed at a reduced rate, approximately 1/10th of the standard input token price (exact rates vary by model). Cached write operations are billed at the normal input token rate.Thinking Mode
Thinking (reasoning) models write out a chain of thought before they answer. That improves accuracy on hard problems but adds latency and output tokens. On Novita, reasoning models fall into three groups:- Always on: the model always reasons and this cannot be turned off, for example
zai-org/glm-5.3. - Switchable (hybrid): the model reasons by default and you can turn it off per request, for example
moonshotai/kimi-k2.5andqwen/qwen3.8-flash. - Adjustable depth: instead of an on/off switch, the model lets you pick how much it reasons.
zai-org/glm-5.3supports the levels low, high and max (default max).
How do I turn thinking off for a request?
For switchable models, add"enable_thinking": false to the request body. The model answers directly and no reasoning tokens are produced or billed.
Where does the reasoning appear in the response?
Reasoning is returned next to the answer, not mixed into it. Set"separate_reasoning": true and read message.reasoning_content (or delta.reasoning_content when streaming); the final answer stays in content. Models that support interleaved thinking may also return a structured reasoning_details array. See Reasoning Models for complete code samples.
Why is content empty with finish_reason: "length"?
max_tokens covers reasoning tokens as well as visible output. If the budget is too small, the model spends all of it on reasoning and returns an empty content. Raise max_tokens, or lower the reasoning depth or turn thinking off on models that allow it.
Are reasoning tokens billed?
Yes. Reasoning tokens are output tokens and are billed at the model’s output price, whether or not you read them. Turning thinking off on switchable models is the most direct way to cut cost and latency on simple requests.Do I need to send reasoning back in multi-turn conversations?
Normally no. Keep onlycontent in the assistant turns of your message history; the model does not need its earlier reasoning to continue. The exception is agentic tool-calling with interleaved-thinking models, where the previous turn’s reasoning must be passed back between tool calls. See Interleaved Thinking Support.
Feature Support
Is the top_logprobs parameter supported?
The top_logprobs parameter is currently not supported.
Is structured output (response_format) supported?
Currently, only "type": "json_object" is supported. "type": "json_schema" (with a strict schema definition) is not yet available.
Troubleshooting
Why am I receiving 403 errors when calling a model?
Common causes include:- Insufficient account balance — top up your account and retry.
- Model requires whitelist access — certain models are restricted. See How do I apply for model whitelist access? below.
- Plan restriction — the model may not be included in your current subscription plan.
Why am I receiving 400 errors?
Common causes include:- Input length exceeds the model’s context window limit.
- Incorrect parameter types or formatting.
- Missing required fields in the request body.
trace_id to our support team for detailed investigation.