Yesterday in AI: 20 May 2026 — Cursor Matches Opus 4.7 at One-Tenth the Cost
Cursor Composer 2.5 ties Claude Opus 4.7 on SWE-bench at 1/10th the cost; xAI launches Grok Build coding CLI with 2M context; OpenHuman hits 11,600 GitHub stars in 15 days.
By OMC Editorial on 2026-05-21
TL;DR — Cursor's Composer 2.5 matches Claude Opus 4.7 on SWE-bench Multilingual at $0.50/M tokens—about 1/10th the price; xAI launched Grok Build, a 2M-context coding CLI with 8 concurrent agents aimed at Claude Code; OpenHuman hit 11,600 GitHub stars in 15 days with a local-first AI agent that front-loads your full digital context.
---
1️⃣ Cursor Composer 2.5: Frontier Benchmark at a Fraction of the Cost
- What: Cursor released Composer 2.5 on May 18—its in-house agentic coding model scoring 79.8% on SWE-bench Multilingual, nearly matching Claude Opus 4.7 80.5%, at $0.50/$2.50 per million input/output tokens.
- Why it matters: Cursor previously depended on Anthropic and OpenAI models it also competed against; Composer 2.5 breaks that dependency while cutting per-token costs roughly 10–30x versus comparable frontier models.
- Key number: 79.8% SWE-bench Multilingual—0.7 points behind Opus 4.7—at $0.50/M input tokens.
Composer 2.5 is built on Moonshot's Kimi K2.5 open-source checkpoint and trained with 25× more synthetic tasks than its predecessor, with targeted textual feedback injected at the exact trajectory points where the model underperformed. Cursor credits those targeted RL improvements for closing most of the gap to frontier-class models without a clean-sheet pretraining run.
More consequentially, Cursor disclosed that its next model will be trained from scratch on SpaceXAI's Colossus 2 supercluster—equivalent to roughly one million H100 GPUs—using 10× more total compute than Composer 2.5. That infrastructure access comes from the April 21 agreement that gives SpaceX the right to acquire Anysphere Cursor's parent for $60 billion later this year.
📎 Cursor Bloghttps://cursor.com/blog/composer-2-5 · TechTimeshttps://www.techtimes.com/articles/316917/20260520/cursor-composer-25-matches-claude-opus-47-coding-benchmarks-one-tenth-cost.htm · DevOps.comhttps://devops.com/cursors-composer-2-5-brings-smarter-more-reliable-ai-coding-agents/
---
2️⃣ xAI Enter