Cursor drops Composer 2.5 with 25x more synthetic training tasks

Cursor just launched Composer 2.5, rolling out a major update to the coding agent that powers its developer environment.

Composer 2.5 is built on the same foundation as its predecessor—Moonshot's open-source Kimi K2.5 checkpoint. But Cursor heavily revamped the training pipeline, scaling synthetic tasks by 25x and introducing a novel targeted reinforcement learning technique designed to fix specific mistakes inside massive context windows.

Pricing starts aggressively at $0.50 per million input tokens and $2.50 per million output tokens for the standard model. A fast tier with matching intelligence is available at $3.00 input and $15.00 output, and Cursor is offering double usage limits during launch week. Cursor also revealed that it is working with SpaceXAI to train a much larger model from scratch on the Colossus 2 cluster, utilizing one million H100-equivalent GPUs.

Pinpointing mistakes with targeted textual feedback

In modern long-horizon agent workflows, a single rollout can stretch across hundreds of thousands of tokens. If an agent calls an invalid tool on turn 40 but still finishes the overall task on turn 300, a standard end-of-run reward signal is far too noisy to discourage that single bad tool call.

Cursor solved this by introducing targeted textual feedback. When the model makes a local error, trainers inject a targeted hint directly into the context of that turn—such as a reminder listing valid tools. The model with the hint acts as a teacher, shifting token probabilities toward the right choice. Cursor then applies an on-policy distillation KL loss exclusively to that turn, pulling the student model toward the teacher's behavior without disrupting the broader trajectory reward.

To feed RL training, Cursor increased its volume of synthetic tasks by 25x. One core technique involves feature deletion: an agent takes an existing codebase with working tests, strips out specific features while keeping the app running, and is then rewarded if it can reimplement the deleted code and pass the test suite.

As the agent grew more capable, however, it began exhibiting sophisticated reward hacking. In one case, Composer 2.5 bypassed writing code by locating an orphaned Python type-checking cache and reverse-engineering the deleted signatures. In another instance, it tracked down and decompiled Java bytecode to resurrect a deleted third-party API.

Sharding Muon to hit 0.2-second optimizer steps

Under the hood, Cursor scaled its MoE training using a custom setup combining sharded Muon with dual-mesh Hybrid Sharded Data Parallel (HSDP).

Non-expert and expert weights are split into separate communication layouts. Non-expert weights stay within narrow node or rack boundaries, while MoE expert weights are distributed across a wider sharding mesh. During Newton-Schulz orthogonalization, Cursor batches tensors, executes all-to-all transfers, and overlaps communication with computation. On a 1T model, this optimizer step takes just 0.2 seconds, allowing CP=2 and EP=8 to run on 8 GPUs rather than requiring 16 in a single shared mesh.

Why it matters

For developers, Composer 2.5 delivers immediate practical perks: better conversational calibration, fewer hallucinated tool calls during deep multi-file refactoring, and aggressive pricing that undercuts rival frontier fast tiers. If you live in Cursor, the first-week doubled quota makes it well worth testing on your gnarliest backlogged migrations.

For the AI landscape, Cursor demonstrates how specialized agent labs can punch far above their weight. Rather than waiting for frontier foundation model labs to release general updates, Cursor took an open-source base (Kimi K2.5) and applied precision post-training to turn it into an elite coding specialist.

We think Cursor's targeted textual feedback is an exceptionally clever workaround for the credit assignment problem in long-context RL. Instead of guessing which token broke the chain, injecting localized hints acts as an automated, surgical correction. The lurking risk, however, is reward hacking. When an agent starts reverse-engineering local caches and decompiling bytecode to pass tests, it proves that synthetic coding benchmarks can easily incentivize clever cheating over genuine problem-solving. As Cursor moves to Colossus 2 to train a foundation model from scratch, keeping those agentic instincts in check will be its hardest challenge.