Saturday model drop beats Opus 5 on agent benchmarks at $0.20 vs $12 per task
sophiamyang · x · 2026-10-17
Sophia Yang highlights an unusual Saturday model drop:
- Stronger and cheaper on agentic tasks: DeepSWE score of 77.0 (up from 74.2), beating Opus 5's 74.0, at roughly $0.20 per task vs $12 for Opus 5.
- Wins on four benchmarks vs Opus 5: DeepSWE, AutomationBench, Agents' Last Exam, and Apex; trails on HLE, Terminal-Bench, ProgramBench, NL2Repo and visual benchmarks.
- Architecture: a 552B-param MoE plus 196B of Engram memory (763B total), with only 8B active during prefill and 16B during decode; 1M context.
- KV cache stored at 4-bit, 890 bytes per token, about a quarter of V4-Flash; pricing from $0.30 per million (listing truncated).
More from Models
- Terence Tao: OpenAI's quasi-Riemann result is the field's biggest advance in 50 years — basedjensen · 2026-10-17
- Chartography leaderboard: Claude Sonnet 5.5 tops chart understanding at 0.902 pass@1 — NielsRogge · 2026-10-17
- Pro $200 plan chewed through by daily research automations, user seeks tips — MidnightIndex35 · 2026-10-17
- Delta appears to block Grok bot's IP range, thwarting flight-booking attempt — lucasmeijer · 2026-10-17
- Qwen-Image-Edit shows East Asian facial prior causing identity drift in edits — Gen-Talks · 2026-10-17
- AI turns a nuclear engineer's PWR rod-bow challenge from X into a structured analysis — burhop · 2026-10-17