<p>Cursor has released Composer 2.5, its fourth coding model in seven months. It’s built on Moonshot’s open-source Kimi K2.5 checkpoint — the same base as Composer 2 — and is already the most-chosen model in Cursor by usage.</p>
<h2>What’s new</h2>
<p>The benchmark numbers are a meaningful step up from Composer 2: Terminal-Bench 2.0 goes from 61.7% to 69.3%, and CursorBench v3.1 from 52.2% to 63.2%. Cursor says the model is better at long-horizon agentic tasks, complex instruction following, and maintaining context over extended sessions.</p>
<p>On the training side, the main changes are:</p>
<ul>
<li>25x more synthetic tasks than Composer 2, including feature deletion tasks grounded in real codebases</li>
<li>Targeted RL with textual feedback, which lets the model assign credit more precisely across rollouts spanning hundreds of thousands of tokens</li>
<li>Infrastructure improvements: Sharded Muon with distributed orthogonalization and dual mesh HSDP for MoE models</li>
</ul>
<h2>The cost angle</h2>
<p>Pricing is $0.50/M input and $2.50/M output tokens. For comparison, Anthropic’s Claude Opus 4.7 runs $25/M output and OpenAI’s GPT-5.5 is $30/M output. That’s a 10-60x cost difference, with Artificial Analysis ranking Composer 2.5 third on their Coding Agent Index.</p>
<p>GergelyOrosz put it plainly: charging roughly 1/20th the price of Opus 4.7 with similar coding characteristics is a real competitive move, and he expects Cursor to win back market share on that basis alone.</p>
<p>To mark the launch, Cursor doubled included usage for the next week and briefly offered 10x usage for a single day.</p>
<h2>Early caveats</h2>
<p>Not everyone is convinced. Some users in agent mode report the model losing context mid-task, and there’s the usual skepticism about whether benchmark gains translate to actual day-to-day productivity. Worth watching before drawing firm conclusions.</p>
<h2>What’s coming next</h2>
<p>Cursor also announced a partnership with SpaceXAI to train a significantly larger model from scratch using 10x more total compute on Colossus 2’s million H100-equivalents. No timeline given, but the framing suggests it’s meant to be a substantial capability jump rather than an incremental update.</p>


Collections

Cursor has launched Composer 2.5, its fourth coding model in seven months, built on Moonshot's open-source Kimi K2.5 checkpoint. Key improvements include Terminal-Bench 2.0 scores rising from 61.7% to 69.3% and CursorBench v3.1 from 52.2% to 63.2%, with better performance on long-horizon agentic tasks and complex instruction following. Training advances include 25x more synthetic tasks, targeted RL with textual feedback, and infrastructure improvements like Sharded Muon with distributed orthogonalization. Pricing is $0.50/M input and $2.50/M output tokens — roughly 10-60x cheaper than Claude Opus 4.7 or GPT-5.5 — making it a significant competitive move. Artificial Analysis ranks it third on their Coding Agent Index. Some early users report context loss in agent mode. Cursor also announced a partnership with SpaceXAI to train a much larger model using 10x more compute on Colossus 2's million H100-equivalents.

Cursor releases Composer 2.5, built on Kimi K2.5 with up to 10x better cost efficiency