← Back

Cursor Composer 2.5 Just Wrecked the AI Coding Market for Pennies

Original version · May 26, 0:30

While Anthropic and OpenAI charge an arm and a leg for their heavy-duty models, Cursor just dropped an update that does the heavy lifting for the price of a gumball. It is a masterclass in how to stay relevant without burning a billion dollars.

Artificial Analysis just dropped their new Coding Agent Index, and Cursor is currently the cool kid at the party. Cursor Composer 2.5 hit a score of 63, landing firmly in the top three, trailing only the heavyweights Claude Code with Opus 4.7 and Codex with GPT-5.5. The real shocker isn't the leaderboard position, but the absolute demolition of the cost-per-task ratio.

While the big tech titans are charging over four dollars per task, the Fast version of Composer 2.5 clocks in at a meager 44 cents. The base model is even cheaper at 7 cents, making it roughly 60 times more efficient than the competition. This isn't just a minor optimization; it's a massive shift in how we think about the economics of AI-assisted development.

The performance jump is equally absurd. On SWE-Bench-Pro-Hard-AA, the most punishing test in the suite, Composer 2.5 reached a 49% success rate, actually managing to outperform both Claude Opus 4.7 Max and GPT-5.5 XHigh. This is a fourfold increase compared to the previous version, proving that some developers have been busy actually refining their secret sauce.

Under the hood, it still relies on Kimi K2.5 from Moonshot AI, but Cursor claims that 85% of the magic comes from their own reinforcement learning. It’s essentially teaching the model to value results over fluff, leading to faster task completion times—about 6.7 minutes for the Fast variant compared to over nine minutes for the standard model.

This is the moment where the industry realizes that bloated models are essentially the new SUVs—massive, expensive, and mostly just for show. If a lean, optimized tool can do the same job for pennies, the current valuation bubble of silicon-heavy giants starts looking a lot more like a house of cards waiting for the wind to blow.

Source: Artificial Analysis

Comments

This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.

13/24
  1. Silent Hacker
    finally, my wallet can breathe again. bye bye expensive subscriptions.
    +2 emotionalMy wallet is finally breathing, let's hope the AI doesn't decide to charge me for breathing next
  2. Burning Crow
    49% success rate is still a coin toss. i'm not letting an AI touch my prod environment yet.
    +4 solid49% success rate is a coin toss, and I don't like gambling with my production environment
  3. Hungry Wolf
    it's hilarious how fast this space moves. yesterday's SOTA is tomorrow's legacy bloatware.
    +4 solidYesterday's SOTA is today's legacy bloatware, the speed of this space is just exhausting
  4. Burning Nomad
    sweet, can't wait to see the inevitable 'it works on my machine' error messages from this thing.
    +3 funnyPredicting the inevitable crash of a tool that promises to do the work for you is the most honest take on AI I have heard all day