The beloved Kimi K3 just got an upgrade. On SWE-Marathon v1.1, which tests multi-hour coding tasks, K3.1 Pro scores 51.7%, up from K3’s 48.1%. Claude Opus 5.5 scores 46.3% and GPT-6 Astra scores 42.5% on the same benchmark.
Moonshot also reports 64.4 on OfficeQA Pro v2 for financial document analysis and 92.7 on BenchCAD with tools for CAD generation.1
Run Kimi K3.1 Pro on Sail Flex for $0.70 input, $0.10 cached input, and $6 output per million tokens.
Sail offers three pricing tiers:
| Tier | Input | Cached input | Output |
|---|---|---|---|
| Flex | $0.70 | $0.10 | $6 |
| Balanced | $1.20 | $0.15 | $8 |
| ASAP | $1.50 | $0.20 | $10 |
For 1M uncached input tokens and 1M output tokens, Sail Flex costs $6.70, compared with $12 for a typical direct API, about 44% less.[2](#fn-2)
[Start building with Kimi K3.1 Pro on Sail](https://app.sailresearch.com/sign-up)
Footnotes #
- Sources: Vals (Terminal-Bench),SWE-Marathon (K3), andMoonshot (other results). BenchCAD: Voxel IoU, 1,000 files, with tools. Missing results shown as dashes.↩
- USD for 1M uncached input and 1M output tokens. Excludes tool charges. Direct API reference: $2 input and $10 output per million tokens. ↩
source & further reading
sailresearch.com — original article
How Channel3 reduced AI cost per product by 10× with Sail Research
Chasing Speed of Light on TPU v6e
Improving Decode Throughput on Intel Gaudi 3