CL-SN55 941 —
GPT-61S 885 -0.6%
GPT-56T 861 —
GPT-6S 825 -1.4%
CL-OP55X 819 -1.4%
MUSE-SPK 793 -1.4%
QWEN-38X 779 -2%
GPT-6A 778 -0.5%
CL-OP5H 712 -1.7%
CL-OP5X 710 -1.7%
KIMI-K3X 700 -2.5%
CL-OP46H 695 -1.7%
CL-FAB5H 693 -1.7%
CL-OP55H 693 -1.7%
CL-OP47H 684 -1.7%
GEM-37FH 649 -1.8%
GEM-38FH 634 -3.4%
GPT-56S 574 -0.5%
CL-OP47 540 -0.6%
INKL 531 —
GEM-3P 494 —
CL-OP46 490 —
CL-OP48 483 —
GEM-31P 459 -5.4%
GPT-6L 436 —
CL-SN55 941 —
GPT-61S 885 -0.6%
GPT-56T 861 —
GPT-6S 825 -1.4%
CL-OP55X 819 -1.4%
MUSE-SPK 793 -1.4%
QWEN-38X 779 -2%
GPT-6A 778 -0.5%
CL-OP5H 712 -1.7%
CL-OP5X 710 -1.7%
KIMI-K3X 700 -2.5%
CL-OP46H 695 -1.7%
CL-FAB5H 693 -1.7%
CL-OP55H 693 -1.7%
CL-OP47H 684 -1.7%
GEM-37FH 649 -1.8%
GEM-38FH 634 -3.4%
GPT-56S 574 -0.5%
CL-OP47 540 -0.6%
INKL 531 —
GEM-3P 494 —
CL-OP46 490 —
CL-OP48 483 —
GEM-31P 459 -5.4%
GPT-6L 436 —
← Back to feed

Cursor Composer 2.5 Takes #3 on AA Coding Agent Index at $0.07 Per Task — 60x Cheaper Than Codex

Artificial Analysis has published its independent evaluation of Cursor’s Composer 2.5 on the AA Coding Agent Index, placing it third globally with a score of 62 — a 14-point gain over Composer 2 (48). The result puts Cursor ahead of every single-model agent that costs less than a dollar per task, and within striking distance of the frontier at a fraction of the price.

The Leaderboard Position

AgentModelIndex ScoreCost Per Task
Claude CodeClaude Opus 4.7 (max)66$4.10
CodexGPT-5.5 (xhigh)65$4.82
CursorComposer 2.562$0.07

The agents above Composer 2.5 cost 10x to 60x more per task depending on which Composer variant you use. At $0.07 (standard) and $0.44 (Fast), Composer 2.5 is the only coding agent scoring above 60 that sits on the cost-quality Pareto frontier.

Benchmark Breakdown

The 14-point index gain over Composer 2 is concentrated in the hardest problems:

  • SWE-Bench-Pro-Hard-AA: +35 points (12% → 47%). At 47%, Composer 2.5 matches Claude Opus 4.7 (max) in Claude Code on this subset.
  • Terminal-Bench v2: +2 points (64% → 66%)
  • SWE-Atlas-QnA: +3 points (69% → 72%)

Wall time averages 6.7 minutes per task in Fast mode, third-fastest among tested agents — behind Claude Opus 4.7 (medium) in Claude Code at 5.8 minutes and GPT-5.5 (medium) in Cursor CLI at 6.2 minutes.

What’s Under the Hood

Composer 2.5 continues training on Moonshot AI’s open-weights Kimi K2.5 base, with Cursor reporting approximately 85% of total compute coming from its own additional training and reinforcement learning on top of that base. The model is not exposed via any public API — it runs exclusively inside Cursor IDE and Cursor CLI.

Pricing is structured in two tiers: standard at $0.50/$2.50 per million input/output tokens, and Fast at $3.00/$15.00. The Fast variant runs roughly 30% faster (6.7 vs 9.3 minutes per task) but costs approximately 6x more per task ($0.44 vs $0.07).

The Practical Implication

For engineering teams running coding agents at volume, Composer 2.5 makes the cost-performance trade-off much sharper. At the frontier, you are paying $4-5 per task for a 3-4 point score advantage over a $0.07 agent. Whether that delta is worth 60x the cost depends on what you are automating.

The result also validates the strategy of building on strong open-weights foundations rather than training from scratch. Kimi K2.5 + Cursor’s RL investment has produced a model that is independently competitive — not just a cheap wrapper around an Anthropic or OpenAI API.