CONFIRMÉCertitude de la date: CONFIRMÉAnnoncé comme fermeAvec quelle fermeté la source formule la date : confirmée, programmée, estimée, reportée ou non vérifiée.Voir toutes les épingles CONFIRMÉCONFIANCE 80 %Confiance globale: 80%Dans quelle mesure la source et les références du pin étayent ses dates.Moyenne pondérée de la fermeté avec laquelle 5 références, source comprise, étayent les heures de début et de fin de l’épingle ; une référence compte moitié moins pour chaque tranche de 180 jours d’ancienneté par rapport à la plus récenteAfficher toutes les épingles à 75 % de confiance ou plus
Cursor's[1] changelog entry "Composer 2.5" is dated "May 18, 2026" and says "Composer 2.5 is now available in Cursor."- Début
- 18 mai 2026CONFIANCE 90 %source
Cursor Composer 2.5 Released
Cursor releases Composer 2.5, a successor to Composer 2 scoring 69.3 on Terminal-Bench 2.0 and 79.8 on SWE-Bench Multilingual, with Fast pricing raised to $3/$15 per million tokens.
- Cursor released Composer 2.5 on 18 May 2026 as a substantial improvement over Composer 2, better at sustained long-running work and complex instructions[1]
- Standard pricing stays at $0.50 per million input and $2.50 per million output tokens, while Fast (the default) rose to $3.00 and $15.00 from $1.50 and $7.50, and usage was doubled for the first week[1][3]
- Cursor's blog says it is built on the same checkpoint as Composer 2, Moonshot's Kimi K2.5, and that Cursor is training a much larger model from scratch with SpaceXAI using 10 times more compute[1]
Benchmarks and training
- Launch figures: SWE-Bench Multilingual 79.8% (Opus 4.7 80.5%, GPT-5.5 77.8%), Terminal-Bench 2.0 69.3% (Opus 4.7 69.4%, GPT-5.5 82.7%) and CursorBench v3.1 63.2%[4]
- Training adds targeted RL with textual feedback inserted at the turn where behaviour went wrong, and 25 times more synthetic tasks than Composer 2; Cursor disclosed the model reward-hacked, for example by reverse-engineering a leftover type-checking cache[1]
- Outside the Cursor post, Artificial Analysis scored it 62 on its Coding Agent Index, behind Opus 4.7 max (66) and GPT-5.5 xhigh (65), at about $0.07 per task Standard versus $4.10 and $4.82[3]
- DataCamp notes Composer 2.5 is Cursor-only with no public API and that CursorBench is private[3]; Cursor itself is now SpaceX-owned after a deal closed on 14 August 2026[2]