Thinking Machines Lab fine-tuned a version of Moonshot AI’s Kimi-K2.6 to turn plain-language business questions into SQL queries, and reported that the resulting model beat the human accuracy baseline on an expert-verified benchmark without using any of the multi-stage prompting pipelines that dominate the text-to-SQL leaderboard today. The model, called ReViSQL-K2.6, reached 92.97 percent accuracy on the Arcwise-Plat-SQL benchmark when the lab sampled 16 candidate queries and kept the majority answer, just ahead of the 92.96 percent mark set by human query writers on BIRD, the dataset Arcwise-Plat-SQL derives from. Under a single greedy pass, the model scored 91.37 percent at roughly $0.035 per task.
The comparison points are the lab’s own, drawn from its blog post and technical report rather than an independent leaderboard. Thinking Machines says its model outperforms GPT-5.6 Sol Ultra and Claude Fable 5 while running at 12 to 15 percent of their per-task cost, and it says the 16-sample version beats the strongest scaffolded open-source pipeline, OpenSearch, by 8.4 points at a 37 percent lower price. Those are the lab’s stated baselines, not ones this article introduces, and the human accuracy figure the model is measured against comes from the same benchmark family Thinking Machines curated.
Most attempts to close this kind of gap add machinery around a fixed model: a stage that narrows the relevant database columns, a stage that generates candidate queries, a stage that repairs broken ones, a stage that votes among survivors. Thinking Machines argues that approach caps out at whatever the underlying model can reason its way to, and that the fix belongs inside training rather than around it. That is a direct challenge to the scaffolding and prompt-engineering layer that a large share of applied AI teams have spent the past two years building, and it is worth treating as a live disagreement rather than a settled one. Plenty of production systems still rely on exactly the kind of orchestration this result argues against.
Two changes account for most of the reported gain. The first is data quality. An audit of 2,500 sampled BIRD training instances found errors in 61.1 percent of them, with more than half of the “gold” SQL answers themselves wrong. Thinking Machines paired a large language model auditor with human experts to correct the set, releasing the fixed version as BIRD-Platinum, and says training on it alone lifted accuracy to 88.55 percent before any reward changes.
The second is reward design. Standard reinforcement learning with verifiable rewards for SQL grades a generated query only by whether it returns the same result as the correct query on one test database, which can reward wrong logic that happens to match by coincidence. Thinking Machines found this happening in 32.8 percent of positive rewards during a pilot run and added a semantic-equivalence check, a solver called VeriEQL, to catch it. A separate failure mode showed up in 24.2 percent of errors: the model ignoring extra context supplied in the prompt and guessing from its pretraining priors instead. The lab addressed that with a rule-based reward requiring the model to show it translated that context into an explicit constraint before answering.
Thinking Machines frames the result as consistent with earlier work it published on training models for financial judgment, arguing that expert-level task knowledge scales better as a training signal than as prompt scaffolding. Teams currently maintaining multi-stage agentic pipelines for narrow, verifiable tasks should treat this as a prompt to test whether a cleaned dataset and a shaped reward on a smaller model can replace the pipeline rather than sit on top of a frontier one.
Reported by Thinking Machines Lab in its blog post “Putting Task Expertise into RL Achieves State-of-the-Art Performance on Text-to-SQL.”
