TL;DR
- Strong on agentic tasks, better and cheaper per task: higher scores with shorter outputs. On DeepSWE it gets 77.0 versus 74.2 before, and beats Opus 5's 74.0, at roughly $0.20 per task against ~$12 for Opus 5.
- Beats Opus 5 on four benchmarks: DeepSWE, AutomationBench, Agents' Last Exam and Apex. It trails Opus 5 on HLE, Terminal-Bench, ProgramBench, NL2Repo and the visual benchmarks.
- The model: a mixture of experts with 552B parameters plus 196B of Engram memory (~763B total), only 8B active during prefill and 16B during decode. 1M context.
- Tiny memory per token: KV cache stored at 4-bit, about 890 bytes per token, roughly a quarter of V4-Flash.
- Price: $0.30 per million tokens in and $1.20 out at peak, half off-peak.