Sonnet uses 2x tokens of Sol/Kimi K3 for no performance gain
zainhas · x · 2026-08-16
A chart comparing long horizon SWE scores with average token usage reveals that Sonnet consumes twice as many tokens as Sol and Kimi K3 without any performance improvement. The author speculates that GLM5.3 would significantly outperform others in this metric.
More from Models
- Gemini 3.7 Flash sees no performance gains in Tibetan/Sanskrit/Chinese — SebastianNehrd2 · 2026-08-16
- Using /goal command spikes token usage 13x for minimal score gain — zainhas · 2026-08-16
- Grok's 'Auto' mode praised for balancing speed and reasoning — mark_k · 2026-08-16
- Community debates if Qwen 3.8 27B outperforms the larger 3.5 122B — MackThax · 2026-08-16
- Claude Sonnet 5 numbers crush Sonnet 4.5 in comparison — zainhas · 2026-08-16
- Ternary LLMs making a comeback? Multiple 1.58-bit models released, some beat full-precision baselines — Individual-Dot5488 · 2026-08-16