# What AI Model Does Lovable Use? (2026 Answer) - Tool: Lovable - Difficulty: Beginner - Fix time: ~5 min read - Compatibility: Applies to Lovable's builder agent on all plans; the Cloud AI gateway (Gemini/GPT models) is a separate, optional feature - Last updated: September 2026 ## TL;DR Lovable's builder agent runs on Anthropic's Claude models — Claude Opus 4.5 is credited with a step-change in reliability on long, multi-file tasks, and Lovable has tested the newer Claude Sonnet 5 against its hardest real pull requests. There is no model picker; Lovable routes tasks to Claude internally without publishing the exact production version. This is separate from the Cloud AI gateway, which powers in-app AI features with Gemini and GPT models instead. ## Why the answer is 'Claude,' not a single fixed model name Lovable's own public comparison guide states plainly that Lovable runs on Claude, and positions its differentiation as the platform layer built around the model rather than the model itself. Anthropic's official case study on Lovable describes the architecture as 'a harness around a main agent that can use subagents to orchestrate tasks effectively, with the right models at each step,' and every new Claude release is eval-tested before it gets rolled into Lovable's builder. Anthropic specifically credits Claude Opus 4.5 with a 'step change in reliability on long-horizon tasks' — the kind of multi-file, multi-step work that a real Lovable prompt often requires. More recently, Lovable ran Claude Sonnet 5, released by Anthropic on June 30, 2026, against dozens of its most challenging real pull requests, which signals that newer Claude generations are actively being tested and adopted as they ship, not locked to one version indefinitely. What is not published anywhere is a model picker or the exact production model version in use at any given moment. Lovable does not expose a way to choose the builder's underlying LLM, and no docs page or changelog entry describes one — the vendor relationship (Anthropic, Claude) is confirmed, but the specific deployed version is intentionally an internal detail that can change without a public announcement. - Lovable's own comparison guide and Anthropic's official case study both confirm the builder runs on Claude, not GPT or Gemini - Claude Opus 4.5 is credited with a step change in reliability on long-horizon, multi-file tasks inside the builder - Lovable has tested Claude Sonnet 5 (released June 30, 2026) against real challenging pull requests, showing newer generations get adopted over time - There is no model picker for the builder — Lovable routes tasks to Claude internally and does not publish the exact production version ## Before you start - None — this is an informational page, not a setup task ## How to fix it ### 1. Know that the builder itself always uses Claude *This is the confirmed, sourced answer to the underlying question* Whenever you send a prompt in Lovable's editor, the response is generated by an agent harness running on Anthropic Claude models, with subagents used to orchestrate multi-step work. This is confirmed by both Lovable's own documentation language and Anthropic's official customer case study — it is not a rumor or a community guess. **Expected result:** You can confidently answer 'does Lovable use Claude?' with yes, citing Lovable's own comparison guide and Anthropic's case study. ### 2. Don't confuse the builder model with the Cloud AI gateway models *These are two entirely different things that both involve AI inside a Lovable project* The builder (what writes your code from a prompt) runs on Claude. The Cloud AI gateway (what powers AI features you build into your own published app, like a chatbot or image generator) defaults to Gemini 3.7 Flash for chat, GPT Image 2 for images, Veo 3.1 Lite for video, and other models per category — see the dedicated guide on building AI agents in Lovable for that side of things. If someone asks 'what model does Lovable use for AI features,' clarify which of the two they mean before answering. **Expected result:** You can correctly distinguish between 'the AI that builds my app' and 'the AI my app uses at runtime' — they are not the same model family. ### 3. Understand there is no model picker *Some users look for a settings toggle that does not exist* Lovable does not expose a model selector for the builder anywhere in the editor or account settings. It routes tasks to Claude models internally, and the exact production model version at any given time is not publicly disclosed. Do not expect to prompt Lovable into using a different LLM for code generation — that layer is fixed by design. **Expected result:** You stop looking for a nonexistent setting and focus your prompting effort on clarity instead of model selection. ### 4. Understand why the underlying model matters for your output quality *Model generation directly affects how well Lovable handles complex, multi-file prompts* Claude Opus 4.5's credited improvement in long-horizon task reliability is a big part of why large, multi-step prompts hold together better in Lovable today than they did on earlier Claude generations. The ongoing testing of Claude Sonnet 5 against real hard pull requests suggests further quality improvements are coming as Lovable adopts newer models. Plan Mode can also deploy research subagents on top of the base builder model for complex features — billed on top of the flat 1-credit message cost, but often worth it for planning quality on anything nontrivial. Even with a strong underlying model, some tasks — deep refactors, unusual integrations, production hardening — sit past what any chat-to-app builder handles reliably on its own, which is where teams typically bring in a RapidDev engineer to finish the last stretch. **Expected result:** You understand that output quality shifts over time as Lovable adopts newer Claude generations, even without you changing anything about how you prompt. ### 5. Check for updates if you need the latest picture *The model relationship is confirmed, but the exact deployed version is a moving target* For the most current public information, Anthropic's customer case study page on Lovable and Lovable's own changelog are the primary sources — treat community claims about specific version numbers or dates with more skepticism than either of those two sources. **Expected result:** You know where to check next time this question comes up, instead of relying on outdated forum posts. ## Best practices - Do not try to prompt Lovable into 'using GPT-4' or any other specific model — there is no picker, and it always routes through Claude - Expect long-horizon, multi-file prompts to hold together better on Opus 4.5-era Lovable than on earlier Claude generations - If you need a specific model for an in-app AI feature, that is the Cloud AI gateway (Gemini or GPT variants), not the builder — configure it separately - Use Plan Mode for big features — it can call research subagents on top of the base model, which is billed but improves plan quality - Check the changelog periodically, since Lovable adopts new Claude releases without announcing every internal swap - Treat 'what model does Lovable use' as an evolving answer — the vendor relationship (Anthropic, Claude) is confirmed, but the exact deployed version is not published ## Frequently asked questions ### Does Lovable use Claude? Yes. Lovable's builder agent runs on Anthropic Claude models, confirmed by both Lovable's own comparison guide and Anthropic's official customer case study. Claude Opus 4.5 is specifically credited with improving reliability on long, multi-file tasks. ### Does Lovable use GPT-4 or ChatGPT to build apps? No. The builder that generates your code runs on Anthropic Claude, not OpenAI's models. GPT models do show up elsewhere in Lovable — as options inside the Cloud AI gateway for in-app chat and image features — but not for the builder itself. ### Can I choose which AI model Lovable uses to build my app? No. There is no model picker for the builder. Lovable routes every prompt to Claude internally and does not expose a setting to change that or to see the exact production model version. ### What model powers AI features inside my published Lovable app, versus the builder? That's the Cloud AI gateway, a separate system from the builder. It defaults to Gemini 3.7 Flash for chat, GPT Image 2 for images, Veo 3.1 Lite for video, and other models per feature category — configurable per feature, unlike the fixed Claude-powered builder. ### Is Lovable using Claude Opus or Claude Sonnet right now? Anthropic credits Claude Opus 4.5 with a major reliability improvement in the builder, and Lovable has separately tested the newer Claude Sonnet 5 against its hardest real pull requests. The exact production model at any given moment isn't publicly disclosed, so treat this as an evolving relationship rather than one fixed version. ### Why did my Lovable output quality seem to change recently without me changing anything? Lovable periodically adopts newer Claude model generations into the builder without announcing every swap, and Plan Mode's research subagents can also affect output quality on complex prompts. Both can shift results even when your own prompting habits stay the same. ### What if the builder model still can't solve my problem? Even the best model in the builder harness has limits on very complex or unusual requirements. RapidDev's engineers step in for exactly that gap — taking over the logic or refactors that need more than a chat prompt can reliably deliver. --- Source: https://www.rapidevelopers.com/lovable-issues/what-model-does-lovable-use © RapidDev — https://www.rapidevelopers.com/lovable-issues/what-model-does-lovable-use