Scale · founder · 7 min read

Which Model Is Your Vibe Coding Tool Actually Running?

Most AI builders don't tell you which LLM is doing the work. Here's how to find out — and why it matters for your costs, output quality, and business risk.

Updated June 14, 2026 ·

AI models Claude model concentration vibe coding risk costs

When you open Lovable and ask it to build a feature, what model actually runs your prompt? When Cursor suggests a fix, is it Claude or GPT-4o? When Bolt generates a React component, who’s on the other end?

Most founders have no idea. The tools don’t advertise it. The model routing is buried in settings or invisible entirely. But the answer matters — for what you’re paying, for the quality you get, and for what happens when Anthropic changes its pricing on June 15.

Here’s the honest breakdown.

Update — June 2026: Fable 5 widens Anthropic’s lead

On June 9, Anthropic released Claude Fable 5, its new top model, and it pulls away from the field on exactly the coding work these tools depend on: 80.3% on SWE-bench Pro versus 58.6% for GPT-5.5 and 54.2% for Gemini 3.1 Pro — about 11 points clear of the next-best model. It’s already live in Claude Code and in GitHub Copilot’s premium tiers, and the Claude-backed builders (Lovable, Cursor, Replit, Bolt, v0) will likely fold it into their premium plans over the coming weeks.

For the concentration argument below, this only sharpens the point. The better and more dominant the top model becomes, the harder your tools converge on it — and the more a single vendor’s pricing move ripples across products you thought were independent. Fable 5’s pricing tells the story: it’s free on Pro, Max, Team, and Enterprise plans only through June 22, then moves to metered credits at roughly double Opus 4.8’s rate. A model gets better, the free ride ends, and the cost surfaces somewhere in your bill whether or not you chose it. If you want the full builder’s read on the release, see our Claude Fable 5 guide.

Why this matters more now than it did six months ago

Three things happened in quick succession that turned “which model runs my tool” from a trivia question into a real business concern.

First, Anthropic filed a confidential S-1 in early June. They’re heading toward an IPO at a $965 billion valuation with $47 billion in annual revenue run-rate. Companies at that stage don’t keep prices flat. The January 2025 “unlimited Claude for $20” era — which made Claude Code’s economics look magical — is already being unwound. June 15 caps programmatic agent usage to a separate credit pool at full API rates.

Second, the Cursor deal. Cursor’s parent company Anysphere reportedly signed a multi-year $60 billion contract for model access — almost certainly Claude. That suggests Cursor’s core experience is locked to Anthropic’s pricing decisions in a way you can’t route around.

Third, the stack concentration. If you squint at the vibe coding ecosystem, a startling fraction of it runs on Claude. Claude Code is the top AI coding agent by usage. Cursor routes most completions and all agent tasks through Claude by default. Lovable, Replit Agent, and Windsurf all default to Claude Sonnet variants for core generation work. When one model-provider raises prices, changes a capability, or introduces a production incident, it ripples across tools you thought were independent.

Tool-by-tool: what’s actually running

Cursor — Default model for Composer agent tasks is Claude Sonnet 4 (configurable). Tab completions use Cursor’s proprietary cursor-small model, not Claude. You can switch the agent to GPT-4o, Gemini 2.5 Pro, or others in settings. Most users run whatever the default is. Most defaults are Claude.

Lovable — Claude Sonnet by default across all generation tasks. Lovable doesn’t offer model switching to users. If Anthropic changes the API, Lovable absorbs it — or passes it on through pricing.

Bolt.new (StackBlitz) — Runs Claude Sonnet by default but has been routing some queries through Anthropic’s API with their own system prompting. StackBlitz controls the routing layer; model version depends on which Anthropic tier they’re subscribed to.

Windsurf (Codeium) — Uses a mix: Codeium’s own windsurf-swe-1 model for some completions, Claude Sonnet for deeper agent tasks and the Cascade agent feature. Codeium operates a hybrid to reduce API dependency — which matters if you care about lock-in.

V0 (Vercel) — Claude by default for UI generation prompts. Vercel has been building its own models (v0-1.5-md family) for faster, cheaper layout work, and increasingly routes simple tasks there. The model mix depends on task complexity. V0 has the most active investment in model independence.

Replit Agent — Claude Sonnet for most agent tasks. Replit Agent 4 (released March 2026) uses parallel sub-agents for execution, all powered by Claude. The hosted execution layer is Replit’s own infrastructure.

GitHub Copilot — The most model-diverse of the major tools. Routes across GPT-4o, GPT-4.1, Claude Sonnet, Gemini 2.5, and now Microsoft’s own MAI-Code-1-Flash (launched June 2). The “Auto” picker routes dynamically based on task type. If you’re uncomfortable with concentration, Copilot is the best-hedged option right now.

Claude Code — Obviously, Claude. This is the purest expression of the concentration risk: when Claude changes, Claude Code changes, with no routing fallback.

The concentration problem in plain numbers

A rough calculation: if you use Cursor, Lovable, Replit Agent, and Claude Code together — which is a common founder stack for 2026 — somewhere between 60–80% of your AI generation work runs on Anthropic’s models. You’re a multi-tool shop on the surface, and a single-vendor shop underneath.

This creates three categories of risk.

Pricing risk — Anthropic IPO → public market pressure → rising API rates. You won’t be notified directly. Your tool’s pricing tier will adjust, or your per-generation credit costs will rise, and you’ll notice it as “my bill went up this month” without connecting it to the upstream cause.

Quality risk — Model updates don’t always improve things for your use case. Claude Sonnet 4 broke certain code generation patterns that Sonnet 3.5 handled cleanly. If your vibe coding workflow depends on specific model behavior, an Anthropic model update can silently degrade your output. Tools that let you pin to specific model versions give you more control.

Availability risk — In March 2026, a Claude API degradation that lasted roughly four hours took Cursor, Lovable, Replit, and Windsurf all offline or into degraded mode simultaneously. Your tools are only as available as the model underneath them.

What you can actually do

You can’t eliminate the dependency, but you can manage it.

Know which model powers which tool in your stack. Check each tool’s settings — most have a model selector under preferences or account settings. Switching is usually free. You’re not locked to the default.

Use GitHub Copilot as your hedge. For IDE work where model quality matters, Copilot’s “Auto” routing across multiple providers is the best insurance against any single-model failure. It’s also the only major tool with its own first-party model (MAI-Code-1-Flash) that isn’t Claude or GPT.

Watch Windsurf and V0 for model independence. Both are actively building proprietary models. If Anthropic pricing spikes, these two are most likely to offer credible non-Claude alternatives at competitive quality.

Set billing alerts. On tools with usage-based overages — Cursor Pro, Replit Core, anything post-June-15 with agent credit pools — set a spending threshold notification. You want to find out about cost spikes at $50 over budget, not $500.

Read the changelog. Seriously. Every major tool has a changelog or release notes page. Model switches and routing updates show up there. If your output quality changes suddenly, that’s the first place to look.

The model powering your tools isn’t a technical detail. It’s a supplier dependency. The sooner you treat it that way, the less surprised you’ll be when it shows up on your invoice.