Cursor Composer 2: Cursor's Third In-House Coding Model

Cursor has released Composer 2, its third in-house coding model and the most capable yet. Available immediately in Cursor and in early alpha through a new interface called Glass, Composer 2 delivers frontier-level coding performance at a fraction of the cost of competing models.

Benchmark Results

Composer 2 posts strong numbers across three major coding benchmarks:

• CursorBench: 61.3 (up from 44.2 for Composer 1.5)
• Terminal-Bench 2.0: 61.7 (up from 47.9)
• SWE-bench Multilingual: 73.7 (up from 65.9)

These scores place Composer 2 ahead of Claude Opus 4.6 on all three benchmarks. The model supports prompts with up to 200,000 tokens and can generate code, fix bugs, and interact with command-line interfaces.

Core Innovation: Compaction-in-the-Loop RL

The headline technical advance behind Composer 2 is what Cursor calls compaction-in-the-loop reinforcement learning. Rather than treating context summarization as a separate post-processing step, Cursor bakes it directly into the RL training loop.

The mechanism works in four steps: the model generates output until it hits a token-length trigger, receives a synthetic query to summarize its context, produces a condensed summary (averaging around 1,000 tokens compared to 5,000 for traditional approaches), and then loops back with the summary plus conversation state including plan status and remaining tasks.

Critically, the self-summaries are part of what gets rewarded. If the model summarizes poorly and loses vital information — a variable name, a past bug fix — it fails the downstream task and receives a negative reward. This feedback loop reduces compaction errors by 50% compared to previous methods.

170+ Turn Coding Sessions

The practical payoff of better self-summarization is dramatically longer reliable coding sessions. Cursor demonstrated this with a case study where Composer 2 completed a complex MIPS compilation task across 170 turns, compressing over 100,000 tokens of context into approximately 1,000 preserved tokens while still solving the problem successfully.

Pricing

Composer 2 ships in two tiers:

• Standard: $0.50 per million input tokens, $2.50 per million output tokens
• Fast: $1.50 per million input tokens, $7.50 per million output tokens (same intelligence, higher speed)

The Standard tier is roughly 10x cheaper than Claude Opus 4.6, reflecting Cursor's strategy of training exclusively on coding datasets rather than building a general-purpose model.

Glass Interface

Alongside the model launch, Cursor announced an early alpha of Glass, a new interface designed to take advantage of Composer 2's long-horizon capabilities. Details remain limited, but Glass appears to be a separate product surface beyond the existing Cursor editor.

Business Context

Cursor now has over 1 million daily active users. The company recently achieved a $29.3 billion valuation and is reportedly raising additional funding targeting approximately $50 billion.