Sapience System Reads Only ~15K Tokens Yet Beats GPT-6 Astra on MRCR, 97.7 vs 96.3
gordic_aleksa · x · 2026-09-15
A surprising long-context result: Sapience, a custom system that reads only 15K tokens and delegates answering to open model DeepSeek V4 Pro, scores 97.7 on OpenAI's MRCR benchmark — beating GPT-6 Astra's 96.3, which must ingest the full 500K–1M token context.
The author notes similar gains on RULER and other long-context benchmarks, suggesting it's not overfit to a single benchmark — hinting that smart retrieval may beat brute-force full-context reading.
Related event: Sapience tops GPT-6 Astra on MRCR reading just 15K tokens(2 posts)→
More from Models
- Perplexity early evals: Gemini 4 Carbon tops WANDR benchmark for long-form research — AravSrinivas · 2026-10-31
- DeepSeek's silent downtime upgrade caught by the pelican bicycle test — teortaxesTex · 2026-10-31
- ARC Prize 2026 closes in 3 days as Yi-Chia Chen tops ARC-AGI-3 at 75.56% — GregKamradt · 2026-10-31
- Fable 5.5 Max stuns with gorgeous SVG of a pelican riding a bicycle, headed to Claude Code — ChrisGPT · 2026-10-31
- OpenAI reportedly to train on users' Daybreak sessions with no opt-out — dyn___ · 2026-10-31
- NVIDIA Nemotron 3.5 scales LLM self-verification, +9-10 points on agentic benchmarks — sarahcat21 · 2026-10-31