This reflects my anecdotal findings in a very narrow domain: GPT-4 consistently resolves complex SQL from natural language, whereas GPT-4-Turbo is hit-and-miss, and similar to 3.5-Turbo in performance.