LiveSQLBench’s new runner-up gains 4.83 points over a verified model—but costs about 13 times more per task

LiveSQLBench’s Base-Full v1 leaderboard has two late-September additions that sharpen an increasingly important trade-off: higher text-to-SQL success rates are arriving with very different evidence and cost profiles.

Jitto Build, configured with Claude Opus 5.5 at high effort, is listed at 41.33% success rate on the LiveSQLBench Base-Full v1 split, with a submission date of September 26, 2026 and a reported cost of $0.6731 per task. That places it second on the currently displayed overall table, behind C3 AI’s DIA at 48.00%; the leaderboard does not show a per-task cost for DIA.

Three days later, OpenCode-Grok4.7 entered at 37.50%. Its displayed cost is approximately $0.0976 per task, but the leaderboard explicitly says that figure is a submitter estimate using Grok 4.3 pricing as a proxy rather than actual billing. That caveat matters: it makes the number useful as a directional estimate, not a clean measured-cost comparison.

A 4.83-point gain with a large cost multiple

The nearest entry carrying LiveSQLBench’s verified-submission badge is Gemini 3.1 Pro at 36.50% and $0.0507 per task, dated April 26, 2026. Jitto Build’s score is 4.83 percentage points higher, while its displayed per-task cost is about 13.3 times Gemini’s. OpenCode-Grok4.7 is 1.00 point higher than Gemini, with a displayed proxy cost about 1.9 times as large.

Those comparisons do not establish which system is economically preferable. The leaderboard defines success rate as the share of tasks passing test cases and says costs are shown only when the selected release has cost data. It also distinguishes verified submissions—codebases evaluated by the BIRD team—from other listed submissions. Jitto Build does not carry the verified badge on the current table, while OpenCode-Grok4.7 has neither a result link nor measured billing behind its cost estimate.

Read the labels, not just the rank

Base-Full v1 contains 600 tasks across 22 databases, according to LiveSQLBench, and the table compares systems on that specific release. The new entries are therefore meaningful movement on a named split, not a cross-benchmark claim.

But the ranking now bundles at least three different evidence states: team-verified submissions, externally reported results and ordinary submissions. It also mixes measured costs, estimates and blanks. The right conclusion is narrow: Jitto Build and OpenCode-Grok4.7 raise the visible success-rate frontier below the leader, but neither row closes the reproducibility and cost-accounting gap that buyers would need for a deployment decision.