Navicat Gone, Spring Shrunk: GPT-6 AI Takes Over Backend Workflows

Last week the author uninstalled Navicat after seven years, not to switch tools but because the need for a traditional database client disappeared.

On September 22 OpenAI completed the GPT-6 matrix: Sol at $2 per million input tokens and $10 per million output tokens, Luna at $0.10 and $0.50 respectively, both at half the previous promotional price. Around the same time the flagship Astra received a Computer Use upgrade. Together these changes affect backend work more concretely than any coding plugin.

Astra scores 72.6% on OSWorld 2.0 for computer operation, about 47% faster than GPT-5.6 Sol, and can run a single task for up to forty minutes. It can open a browser, modify a CRM, fill forms, and — in a backend context — connect to databases, write SQL, inspect slow-query logs, and adjust configuration from screenshots without any human clicking in a GUI.

The author now supplies a read-only database account and lets the model query autonomously; the Navicat wrapper has lost most of its purpose.

Pricing details: Sol $2/$10, Luna $0.10/$0.50, cached input at a further 10% discount, while the Terra model is retired. OpenAI emphasizes cost-per-task rather than benchmark scores.

On AutomationBench Sol averages roughly $0.27 per task, outperforming Astra's low-intensity tier at about one-quarter the cost. On DeepSWE Luna achieves 66.6% at a per-task cost 93% lower than Opus 5 — translating to work that previously took half a day now costing a few cents.

Spring is not eliminated but compressed into a template. Its value lies in conventions and glue that stitch business logic to infrastructure. The boilerplate layers — Controller, Service, Mapper — are generated in one shot by GPT-6, complete with tests and exception handling. The model consumes repetitive labor; what remains for humans are transaction boundaries, idempotency guarantees, and production incident judgment.

The framework hasn't died; it has shifted from "three months of learning" to "the model writes, you review."

Backend development persists, yet the "code-only" role is shrinking. Professionals who can define problems and audit model output become more valuable.

Actionable advice: this week pick your most hated repetitive task, run it end-to-end with Sol or Luna, and observe where the model fails. Those failure points are exactly the moat you must defend.