GPT 6.1 Terra/Sol benchmarked on fresh Codeforces problems: Sol 6.1 fastest on 3 of 4
tuhdo · reddit · 2026-10-27
A Reddit user benchmarked models on four fresh (Sep–Oct) high-difficulty Codeforces problems — 2271G (3500), 2271H2 (2000), 2273F (3500), 2261F2 (5000) — unlikely to be in training data.
- Sol 6.1 posted the fastest solve time on 3 of 4 problems (13m24s, 1m21s, 25m06s), often at Low/Medium reasoning effort
- Terra 6.1 Max was fastest on 2271H2 (15m33s); on 2273F, High effort beat Max by 4m10s, but Max's solution ran 2× faster with 2× less memory
- Terra 6.1 at low effort refuses to solve quickly — framed as honest token/time saving rather than a flaw
- Smaller models (Luna, Haiku) struggle even with 200-point problems like 2271H2
Takeaway: Sol 6.1 remains the smartest but tends to over-engineer like "a mathematician writing code"; Terra 6.1 under-engineers. The author suggests pairing Sol's slightly over-engineered plans with Terra's lazier implementations to balance the weaknesses.
More from Models
- AI text detector posts 99.8% Arabic accuracy with under 1% false positives across languages — markjeffrey · 2026-10-28
- Day 1 of AI model competition livestream: Gemini takes clear lead — my_cat_can_code · 2026-10-28
- RWE-Bench: GPT-6 Astra leads real-world engineering eval at 37%, most models score 0% — wavefnx · 2026-10-28
- X launches referral program that rewards Premium recruiters with unlimited usage resets — XFreeze · 2026-10-28
- Lafa turns pretrained LLMs into flow models: 2.54x faster decoding at 7B scale — KyeGomezB · 2026-10-28
- Swiss lab Giotto launches sovereign models Alto and Solo, touting top benchmark scores — MannyKayy · 2026-10-28