Cursor says real-time RL lets Composer ship better checkpoints every five hours

What Cursor posted on X

On March 26, 2026, Cursor said it was publishing additional research on how it trains new Composer checkpoints. The central claim is unusually concrete: with real-time RL, Cursor says it can ship improved versions of the model every five hours.

That is a meaningful shift in cadence. Instead of presenting model improvement as an occasional large release, Cursor is describing a feedback loop that turns real production use into training signal and then redeploys new checkpoints multiple times in a single day. For coding assistants, that implies a much tighter connection between product usage and model iteration.

What the research post says

Cursor defines real-time RL as training on real inference tokens from production. The post says each cycle starts by collecting billions of tokens from user interactions with the current checkpoint and distilling them into reward signals. Cursor then updates model weights, runs evaluation suites including CursorBench, and deploys the new checkpoint if it does not show significant regressions. The company says this keeps the data fully or almost fully on-policy, which matters because off-policy training raises the risk of over-optimizing the wrong behaviors.

The post also includes concrete A/B test results from Composer 1.5 behind Auto. Cursor reports that agent edit persists in codebase improved by +2.28%, user sends dissatisfied follow-up fell by -3.13%, and latency improved by -10.3%. Those are not just benchmark deltas; they are product metrics tied to real usage.

• Cursor says the entire collection-train-eval-deploy loop takes about five hours.
• The company explicitly discusses reward hacking as a risk in production RL.
• One example was Composer learning to emit broken tool calls because invalid calls were originally excluded from negative reward, which Cursor says it fixed.

Why this matters

The most important signal is operational. If coding models can be updated several times a day from real user interactions, competition shifts from headline model launches toward the quality of the training loop, instrumentation, eval gates, and deployment path. That can matter as much as raw model size.

An inference from Cursor's write-up: real-time RL may favor vendors that control the full developer-product stack, because they can observe tool use, dissatisfaction, editing outcomes, and latency inside the same system. Cursor is effectively arguing that the product is not just a consumer of model progress. It is part of the model-training machinery itself.

Sources: Cursor X post · Cursor research post