River open recipe: post-trained open models beat GPT-6 Astra Pro and Opus 5.5 on text-to-SQL for under 1% of cost

kylekosic · x · 2026-10-08

River AI launched River Recipes, an open-source initiative turning research into vetted, reproducible training recipes, starting with text-to-SQL: post-trained open models beat GPT-6 Astra Pro and Claude Opus 5.5 on Arcwise-Plat for less than 1% of the cost.

Key points:

Built on ReViSQL (RL + LoRA on Kimi K2.6), rebuilt on River's platform with several implementation fixes for performance and stability.
Four open models reached human-level 92.96%; five outperformed GPT-6 Astra Pro at max effort by 4.0 points on average at <2% cost per question, with GLM-5.3 Flash leading (+4.6 points, 0.5% cost).
Post-training behavior: large MoE models became more concise while smaller models reasoned and explored databases more, driving up token costs differently per model.

Original post →

River open recipe: post-trained open models beat GPT-6 Astra Pro and Opus 5.5 on text-to-SQL for under 1% of cost

More from coding & agent