Core AI Settings

Learn how to configure moderation, foundation model selection, and temperature for your AI character

The Core AI Settings section defines the foundational behavior of your AI character by controlling safety filters, the underlying language model, and the creativity level of its responses. These settings have a significant impact on how your character interacts with users, balancing safety, accuracy, and creativity.

Main Features

1. Enable Moderation Filter

This setting allows you to filter out potentially harmful content, including hate speech, profanity, or inappropriate language. You can turn the moderation filter on or off using the toggle located at the top of the page. By default, this setting is enabled.

Disabling the Moderation Filter makes some foundation models unavailable.

Features like Narrative Design and Multilingual support will not work when moderation is disabled.

2. Select Foundation Model

Choose from a variety of Large Language Models (LLMs) from leading providers:

OpenAI
Anthropic
Google
Llama

Model availability depends on whether the Moderation Filter is enabled.

Supported LLMs

Below is a list of Large Language Models (LLMs) available in the Convai Playground under Core AI Settings. Models marked as ✅ Flagship are the providers’ top-tier, most capable models — but usage of these is subject to the Flagship Interaction Cap based on your plan.

Flagship LLMs This is the limit on the number of interactions you can perform using Flagship LLMs.
Example: In the Indie Dev plan, you have a total monthly quota of 3000 Interactions. However, the Flagship LLM Interaction Cap is 1500. If you use GPT-4.1 after 1500 interactions, your Flagship LLM quota will be exhausted. You will then need to switch to a non-Flagship LLM for the remaining 1500 interactions.

Realtime / Live Models

OpenAI

Model

Model Code

Flagship

GPT Realtime 1.5 (beta)

gpt-realtime-1.5

GPT Realtime Mini (beta)

gpt-realtime-mini

Google

Model

Model Code

Flagship

Gemini 2.5 Flash Live (beta)

gemini-2.5-flash-live

Gemma 4 31B Fast (beta)

realtime-gemma-4-31b-it

Gemma 4 26B A4B Fast (beta)

realtime-gemma-4-26b-a4b-it

Standard Models

OpenAI

Anthropic

Model

Model Code

Flagship

Claude 4.5 Sonnet (beta)

claude-4-5-sonnet

Claude 4.5 Haiku (beta)

claude-4-5-haiku

Claude Sonnet (latest)

claude-sonnet

Claude Haiku (latest)

claude-haiku

Google

Model

Model Code

Flagship

Gemini 3.5 Flash

gemini-3.5-flash

Gemini Flash (latest)

gemini-flash

Gemini 3.1 Flash Lite

gemini-3.1-flash-lite

Gemini Flash Lite (latest)

gemini-flash-lite

Gemini 2.5 Flash

gemini-2.5-flash

Gemini 2.5 Flash Lite

gemini-2.5-flash-lite

Gemini 3.8 Flash offers these reasoning profiles:

Model

Model code

Gemini 3.8 Flash (Low)

gemini-3.8-flash-low

Gemini 3.8 Flash (Medium)

gemini-3.8-flash-medium

Gemini 3.8 Flash (High)

gemini-3.8-flash-high

The Low, Medium, and High profiles select how much reasoning Gemini 3.8 Flash performs. Their provider token prices are the same; additional reasoning increases output-token usage.

All three profiles use 0.75 Convai credits per 1,000 input tokens and 3.75 credits per 1,000 output tokens, including reasoning. Each turn also includes the Platform Fee of 6 credits. Speech, session duration, memory, knowledge retrieval, and other enabled services add their own usage, and Convai rounds each turn's total up to the next whole credit. See How Convai credits work.

Gemini 3.8 Flash latency and credit measurements

These staging measurements use 30 synthetic requests per profile across conversation, questions using supplied reference notes, and planning, with approximately 3,900 input tokens per request, on September 5, 2026. Requests run one at a time. Latency measures the language-model request to the first text response, including reasoning. It excludes connection setup, speech recognition, speech synthesis, and delivery to your device:

Profile

First text, median

First text, 95th percentile

Mean LLM credits

Mean generation credits

Low

1.06 s

3.65 s

4.03

10.03

Medium

4.26 s

7.49 s

8.63

14.63

High

5.57 s

9.28 s

11.03

17.03

The LLM column quotes the language model in isolation, using measured input, visible output, and reasoning tokens. The generation column shows observed staging charges including the six-credit Platform Fee. Both columns exclude session duration charges and other services. These workload averages do not guarantee production costs or response times.

Google's introductory provider prices are $0.75 per million uncached input tokens and $3.75 per million output tokens, including reasoning, through December 31, 2026. Provider prices describe Google's charges and differ from Convai credits. See Google Gemini API pricing.

Qwen

Model

Model Code

Flagship

Qwen3.6 27B (beta)

qwen3.6-27b

Qwen3.6 35B A3B (beta)

qwen3.6-35b-a3b

Llama

Model

Model Code

Flagship

Llama 4 Maverick (beta)

llama-4-maverick

Llama 4 Scout (beta)

llama-4-scout

Llama3 70B

llama3-70b

xAI

Model

Model Code

Flagship

Grok 4.3

grok-4.3

3. Temperature Control

Function: Adjusts the randomness and creativity in the AI’s responses.
Slider Range: 0.0 (most deterministic) to 1.0 (most creative).

Temperature Range

Behavior

Use Case

Low (0.0–0.3)

Deterministic, consistent

Factual Q&A, compliance-critical interactions

Medium (0.4–0.7)

Balanced accuracy and creativity

Conversational agents, customer support

High (0.8–1.0)

Diverse, creative, sometimes unpredictable

Storytelling, brainstorming, roleplay

Lower temperature sharpens the probability distribution for more predictable word choices.

Higher temperature flattens the distribution, allowing less likely words to appear more frequently.

4. Reasoning Level

Found under Advanced Settings, next to Temperature.

Function: Controls how much internal reasoning the model does before answering.
Availability: Only shown for models that support it. Models without reasoning control — such as the Gemma, Llama, Qwen and GLM families — do not display this setting.

Reasoning trades latency for answer quality. More reasoning generally produces better handling of multi-step questions and instructions, at the cost of a slower first response.

Available options

The exact list depends on the selected model, because each provider exposes a different scale.

Option

Behavior

Auto

No level is sent. The model applies its own adaptive default, reasoning more on hard requests and less on easy ones.

Off / Minimal

The lowest setting the model offers. Fastest first response.

Low

A small amount of reasoning.

Medium

Balanced. Most providers' own default.

High and above

Maximum reasoning. Slowest, best on complex multi-step requests.

Auto is usually the right starting point. Current models already adapt their own reasoning to the difficulty of each request, so Auto typically keeps easy turns fast while still allowing the model to think when a request genuinely needs it. Pin an explicit level when you need predictable latency, or when you have measured that a specific level performs better for your use case.

Higher reasoning levels increase both response latency and token consumption. If your experience is latency-sensitive — a live voice agent, for example — measure the effect before raising the level.

Switching models

If you change the foundation model, your reasoning level is kept when the new model also supports it. When it does not, the setting falls back to Auto, so the character never sends a value its model would reject.

Characters created before this setting existed

Characters that have never had a reasoning level set show Model default and keep the behavior they have always had. Editing and saving other settings will not change this. Selecting any other option opts the character in, and there is no way back to Model default afterwards — choose Auto if you want the model to decide.

Conclusion

The Core AI Settings give you precise control over your character’s foundation model, safety filters, and response style. By adjusting these parameters, you can create an AI that balances safety, reliability, and creativity to suit your specific application.

PreviousPersonality Traits

NextGuardrails

Last updated 1 day ago

Was this helpful?