Lovable 早期测试 GPT-5.5：最难任务通过率 41.6%，比 GPT-5.4 提升 12.5%
Original title: Testing GPT-5.5 in early access: what we are seeing so far
The title and summary in the selected language are awaiting translation.
Lovable 在早期访问中测试 GPT-5.5，其内部基准显示最难任务通过率从 GPT-5.4 的 36.9% 升至 41.6%，每次请求工具调用减少 23.1%，用户卡住的消息占比下降 9.9%。GPT-5.5 每请求输出 token 减少 33%，日常任务成本效率提升约 15%，将很快向 Lovable 构建者开放。
Lovable has been testing GPT-5.5 in early access and our evals show it's the most capable model we've tested for getting builders unblocked and is meaningfully stronger than GPT-5.4 on the more complex tasks that can stall a build session.
This post breaks down how we benchmarked it, where it wins, and what we're seeing in production.
How we evaluate new models
We run every candidate model against Lovable's internal benchmark suite, which measures end-to-end app-building performance on:
- Production-readiness: security, reliability, and the kind of edge-case handling a senior engineer would expect in a code review
- Agentic task completion: multi-step requests that involve tool use, file edits, and iterative debugging
- Common build scenarios: authentication flows, real-time syncing, database integration, multi-file edits, API wiring
- Unblocking users on the hardest tasks: specific moments where builders stall, for example extensive UI polishing, long debugging loops, backend config issues, and complex requests that require the model to propose a path forward rather than just execute
GPT-5.5 vs. GPT-5.4
| BENCHMARK | GPT-5.4 | GPT-5.5 | DELTA |
|---|---|---|---|
| Hardest-tasks benchmark (Senior engineer bar) |
36.9% | 41.6% | +12.5% |
| Tool calls per request | 11.74 | 9.03 | -23.1% |
| High user success rate | 27.36% | 30.62% | +11.9% |
| % of stuck user messages | 3.086% | 2.780% | −9.9% |
What changed under the hood
Three things stand out in how GPT-5.5 handles Lovable workloads:
- Deeper reasoning on high-stakes requests. On our hardest benchmark GPT-5.5 is 12.5% ahead of GPT-5.4 at the same cost. It weighs consequences more carefully and produces output closer to what a senior engineer would accept on first review.
- Tighter tool use. On average it makes 23.1% fewer tool calls per request and 33% fewer output tokens per message. In practice: fewer intermediate steps, less course-correction, more intentional edits. This makes the model around 15% more cost-efficient than GPT-5.4 on everyday tasks.
- Better unstuck behavior. GPT-5.5 is 9.9% more likely to resolve complex tasks blocking users across the board. We've seen it produce novel solutions to problems with strong reasoning when there's no obvious path forward.
What this means for builders
GPT-5.5 is the model to pull in when builders are stuck. The UI fix that won't land, the bug that won't go away, the backend config that's silently breaking your app — GPT-5.5 resolves those in far fewer turns, with a higher success rate, at a lower cost per session.
Builders want continuous progress, not endless iteration. GPT-5.5 breaks through the walls people usually hit on more complex tasks, like authentication flows and real-time syncing, in far fewer turns. It really shines when the work gets hard, handling tough tasks with far less back-and-forth.
— Fabian Hedin, CTO & Co-founder, Lovable.
This will be rolled out to Lovable builders soon.
Source: Lovable · Blog · lovable.dev