AI Daily Brief — 19 March 2026

Cursor ships Composer 2 — its second-generation in-house coding model — claiming frontier coding performance: 61.7 on Terminal-Bench 2.0 (edging out Claude Opus 4.6), with pricing at $0.50 per million input tokens (roughly one-tenth of Anthropic’s). Within 24 hours, the base model is exposed as Kimi K2.5 from Moonshot AI.

Top stories

• Cursor Composer 2. Terminal-Bench 2.0: 61.7; input pricing $0.50 per M tokens. Pitched as in-house frontier-coding intelligence with continued pre-training and RL on top of an undisclosed base. via Medium
• The Kimi K2.5 disclosure. A developer probing Cursor’s OpenAI-compatible base URL finds accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast in API responses — exposing Composer 2’s base as Moonshot AI’s open-weight Kimi K2.5. Cursor’s VP of Developer Experience later concedes: “Not mentioning Kimi as the base in the blog post from the start was a mistake.” via Recording Law

By the numbers

• Composer 2 Terminal-Bench 2.0: 61.7
• Pricing: $0.50 per M input tokens (~10× cheaper than Anthropic Opus)
• Base model exposed: Kimi K2.5 (Moonshot AI, open-weight)

Compiled by AI Feed’s editor from publicly reported announcements on 19 March 2026.