# Lovable's Opus 5 upgrade cuts build variance 22%

By Steven Van · 2026-07-24

Lovable says the model scores highest yet on shipping complete, working apps, with less run-to-run variance than Opus 4.7.

Lovable added Claude Opus 5 the day it launched, and unlike a generic "new model available" note, it came with numbers tied to the one metric that matters for a vibe-coding platform: does the app actually work? Lovable says it pushed Opus 5 hardest on shipping complete, working apps, and it's **up 22% over Claude Opus 4.7 on its most demanding coding tasks, with the highest finished-app quality Lovable has measured and noticeably less variance run to run**.

[Source post](https://x.com/i/status/2080703545389809780)

## Finished-app quality, not demo quality

Most model announcements lead with abstract benchmarks. Lovable's framing is refreshingly specific: the top performer in its family for producing apps that are done, not apps that look done in a screenshot and fall apart on the second click. That's the honest bar for vibe coding, where the gap between "it generated something" and "it generated something I can ship" is the whole product.

## Less variance is the underrated win

The detail worth dwelling on is "noticeably less variance run to run." Anyone who builds with AI knows the frustration of re-prompting the same thing and getting a great result once and a broken one the next time. Lower variance means the model is more reliably good, which for a tool people use to ship real products matters as much as a higher peak score. Predictable beats occasionally-brilliant when there's a deadline.

## The economics behind it

Opus 5 is Anthropic's new flagship, pitched as roughly Claude Fable 5-level intelligence at half the price ($5/$25 per million tokens), and it roughly doubled Opus 4.8's score on the hardest agentic coding benchmark (Frontier-Bench: 43.3% vs 18.7%). It also adds a low/medium/high effort toggle. For a platform like Lovable that runs the model on every build, better output at a lower price compounds directly into either cheaper builds or more attempts per dollar, both of which help users ship.

## Try it

Opus 5 is live in Lovable now. The honest test is a real project, not a toy: describe a complete app with a few interacting features and see whether the first result is closer to shippable than it used to be, and whether re-running it stays consistent.

![Lovable](<https://www.google.com/s2/favicons?domain=lovable.dev&sz=128>)

Lovable

Build software with AI.

[View Lovable →](<https://creatorstoolbox.com/tools/lovable>)

Sources

[Lovable on X](<https://x.com/Lovable/status/2080703545389809780>), [Anthropic](<https://www.anthropic.com/>).

[Read Lovable's Opus 5 upgrade cuts build variance 22% on Creators Toolbox](<https://creatorstoolbox.com/blog/claude-opus-5-in-lovable>)

---
Canonical source: https://creatorstoolbox.com/blog/claude-opus-5-in-lovable