What LLM does Replit use? Current Agent models explained

In short

Replit Agent is a multi-model system. Auto routing is available in Economy, while manual choices depend on mode and account access. The model inside an app is a separate decision.

• Lite currently offers fast models for scoped changes, including Kimi, GPT, Gemini, and DeepSeek choices.
• Economy currently includes Auto and selected Claude, GPT, Gemini, and GLM models for cost-conscious work.
• Power currently includes selected Claude, Kimi, and GPT models for more capable work.
• A primary-model choice guides most of the task. Replit can still route selectively when higher effort is needed.
• Replit AI Integrations is a different feature. It adds model APIs from OpenAI, Anthropic, Google, or OpenRouter to an app you build.
• Compare models on the same task, code state, acceptance criteria, tests, and budget before standardizing a team workflow.

Why is there no single Replit LLM?

Replit's default behavior can balance speed, cost, and accuracy for each request. The model selector also lets eligible users pick the primary model used for most of a task. This makes the answer dependent on the selected mode, model choice, effort setting, rollout, organization rules, and the date checked.

Open Replit's current model selector documentation and check the list shown in your account.

What LLM does Replit use now?

The official selector page groups the current catalog by Agent mode. The names below reproduce that documented grouping on the publication date. They do not promise access for every account.

Lite model choices

• Kimi K2.7
• GPT-5.6 Luna
• Gemini 3.5 Flash
• DeepSeek V4 Flash

Replit describes Lite as a mode for quick edits, visual changes, bug fixes, and other small tasks. Use it when the requested change is narrow and already understood.

Economy model choices

• Auto
• Claude Sonnet 4.6
• Claude Sonnet 5
• GPT-5.6 Luna Fast
• GPT-5.6 Terra
• Gemini 3.1 Pro
• GLM 5.2

Replit positions Economy for everyday builds, learning, and work that needs fewer credits per task. Auto leaves the primary model choice to Replit's routing.

Power model choices

• Claude Fable 5
• Claude Opus 4.8
• Claude Opus 4.8 Fast
• Claude Opus 5
• Claude Opus 5 Fast
• Kimi K3
• GPT-5.6 Sol

Replit positions Power for complex tasks, larger codebases, and harder problems. Some fast variants require Turbo access on an eligible plan. Higher capability or speed can increase credit use.

The selector shown in an account is authoritative. Model selection is rolling out, and the list can vary with organization settings and authorization. If the Primary model control is missing, the feature may not be available to that account yet.

Why do older answers name only Claude?

Older answers may accurately describe an earlier Replit release. For example, Replit's 17 April 2026 changelog said Claude Opus 4.7 powered Power mode at that time. The current selector now documents a broader and newer catalog. Treat a dated launch post as release history, not as a permanent platform specification.

Compare the April 2026 announcement with the current selector before repeating an older model claim.

How do mode, model, effort, and Turbo differ?

• Mode sets the broad cost, speed, and capability tier. Replit currently exposes Lite, Economy, and Power.
• Primary model selects an available model inside the chosen mode for most of the next task.
• Effort controls how much reasoning the selected model uses, from Low through Max. Higher settings can add cost on difficult work.
• Turbo is a paid speed control inside Power. Official guidance says it can be faster and can cost substantially more than standard Power.
• Compare models runs the same chat prompt through up to four selected models. Each run uses credits separately.

Review Replit's current mode guidance and Turbo availability.

How does the agent use a model?

An LLM generates and reasons over text, but Replit Agent also receives project context and can use tools. Those tools can inspect files, edit code, run commands, test behavior, and guide publishing. Checkpoints, prompts, repository state, integrations, tests, permissions, and Replit's orchestration can change the result even when the primary model stays the same.

Read Replit's explanation of models, tools, and Agent responsibilities.

Do not confuse Agent models with models inside your app

Replit Agent helps build the project. Replit AI Integrations lets the finished project call model APIs. The integration catalog currently includes OpenAI, Anthropic, Google, and OpenRouter. A team can use one model to build an app and a different model inside that app.

Agent choice

which primary model helps plan, edit, test, and explain the project inside Replit.

Application choice

which provider and model the deployed software calls for its own user-facing feature.

Credential choice

Replit-managed credentials on an eligible plan or the builder's own provider key.

Billing choice

Agent effort-based usage, managed integration API usage through Replit credits, or direct provider billing with a personal key.

Data choice

which project content, prompt, user data, and output can reach Replit, a provider, or another service.

Check the current Replit AI Integrations providers, model list, billing, and privacy settings.

Verify the data path for the selected feature

Do not infer data handling from a model name. Identify whether the task uses Replit Agent, a managed AI integration, OpenRouter, or a direct API key. Then review the applicable Replit terms, provider terms, retention, training, logging, region, and organization controls. Avoid placing secrets or real personal data in a test unless the approved setup requires it.

Replit's current integration documentation describes different OpenRouter defaults for paid and free endpoints. It also describes stricter zero-data-retention routing for Enterprise. These settings concern models called by an app through that integration. They do not automatically describe every Replit Agent data flow.

Measure cost at the task level

Replit uses effort-based billing for Agent work. A model label alone does not predict the final charge because task complexity, mode, effort, tools, retries, testing, and comparison runs can matter. Set a budget, inspect usage by checkpoint, and compare accepted outcomes rather than raw prompt counts.

Use Replit's current AI billing guide to monitor Agent and integration spending.

Run a controlled model comparison

Choose a representative task that is important but recoverable. Start every candidate from the same checkpoint. Keep the prompt, files, constraints, and acceptance criteria fixed. A useful comparison measures the completed change, not the confidence of the explanation.

1. 1Define the task, excluded changes, required files, user outcome, and maximum budget.
2. 2Create a clean checkpoint and record the current tests, build result, known defects, and data state.
3. 3Run one model at a time, or use Replit's comparison feature for a chat-only evaluation.
4. 4Review the diff for scope, correctness, security, accessibility, maintainability, and unnecessary dependencies.
5. 5Run the same automated checks and the same manual user flow against each result.
6. 6Record elapsed time, credits, retries, human correction time, passed criteria, and remaining risk.
7. 7Repeat with a second task before setting a team default. Keep a fallback for model or catalog changes.

Use Better Design's AI guardrails to review generated interface and application work.

Which Replit model should you choose?

Start with Auto when you want Replit to balance the request inside Economy. Pick a specific model when reproducibility, provider preference, or comparison matters. Use Lite for a narrow edit and Power for a complex task. Raise effort only when the harder reasoning justifies added cost. Confirm every result with code review and tests.

Frequently asked questions